Authors: Moran Sun, Tianlin Li, Yuwei Zheng, Zhenhong Zhou, Aishan Liu, Xianglong Liu, Yang Liu
Abstract: Emotion plays an important role in human cognition and performance. Motivated by this, we investigate whether analogous emotional signals can shape the behavior of large language models (LLMs) and agents. Existing emotion-aware studies mainly treat emotion as a surface-level style factor or a perception target, overlooking its mechanistic role in task processing. To address this limitation, we propose E-STEER, an interpretable emotion steering framework that enables direct representation-level intervention in LLMs and agents. It embeds emotion as a structured, controllable variable in hidden states, and with it, we examine the impact of emotion on objective reasoning, subjective generation, safety, and multi-step agent behaviors. The results reveal non-monotonic emotion-behavior relations consistent with established psychological theories, and show that specific emotions not only enhance LLM capability but also improve safety, and systematically shape multi-step agent behaviors.
Authors: Yuxing Lu, Yushuhong Lin, Jason Zhang
Abstract: Large language models applied to clinical prediction exhibit case-level heterogeneity: simple cases yield consistent outputs, while complex cases produce divergent predictions under minor prompt changes. Existing single-agent strategies sample from one role-conditioned distribution, and multi-agent frameworks use fixed roles with flat majority voting, discarding the diagnostic signal in disagreement. We propose CAMP (Case-Adaptive Multi-agent Panel), where an attending-physician agent dynamically assembles a specialist panel tailored to each case's diagnostic uncertainty. Each specialist evaluates candidates via three-valued voting (KEEP/REFUSE/NEUTRAL), enabling principled abstention outside one's expertise. A hybrid router directs each diagnosis through strong consensus, fallback to the attending physician's judgment, or evidence-based arbitration that weighs argument quality over vote counts. On diagnostic prediction and brief hospital course generation from MIMIC-IV across four LLM backbones, CAMP consistently outperforms strong baselines while consuming fewer tokens than most competing multi-agent methods, with voting records and arbitration traces offering transparent decision audits.
Authors: Hy Dang, Quang Dao, Meng Jiang
Abstract: Tool-integrated LLMs can retrieve, compute, and take real-world actions via external tools, but reliability remains a key bottleneck. We argue that failures stem from both tool-use accuracy (how well an agent invokes a tool) and intrinsic tool accuracy (the tool's own correctness), while most prior work emphasizes the former. We introduce OpenTools, a community-driven toolbox that standardizes tool schemas, provides lightweight plug-and-play wrappers, and evaluates tools with automated test suites and continuous monitoring. We also release a public web demo where users can run predefined agents and tools and contribute test cases, enabling reliability reports to evolve as tools change. OpenTools includes the core framework, an initial tool set, evaluation pipelines, and a contribution protocol. Experiments and evaluations show improved end-to-end reproducibility and task performance; community-contributed, higher-quality task-specific tools deliver 6%-22% relative gains over an existing toolbox across multiple agent architectures on downstream tasks and benchmarks, highlighting the importance of intrinsic tool accuracy.
Authors: Ha Na Cho
Abstract: Single-agent large language model (LLM) systems struggle to simultaneously support diverse conversational functions and maintain safety in behavioral health communication. We propose a safety-aware, role-orchestrated multi-agent LLM framework designed to simulate supportive behavioral health dialogue through coordinated, role-differentiated agents. Conversational responsibilities are decomposed across specialized agents, including empathy-focused, action-oriented, and supervisory roles, while a prompt-based controller dynamically activates relevant agents and enforces continuous safety auditing. Using semi-structured interview transcripts from the DAIC-WOZ corpus, we evaluate the framework with scalable proxy metrics capturing structural quality, functional diversity, and computational characteristics. Results illustrate clear role differentiation, coherent inter-agent coordination, and predictable trade-offs between modular orchestration, safety oversight, and response latency when compared to a single-agent baseline. This work emphasizes system design, interpretability, and safety, positioning the framework as a simulation and analysis tool for behavioral health informatics and decision-support research rather than a clinical intervention.
Authors: Mark Dranias, Adam Whitley
Abstract: Large language models (LLMs) are increasingly embedded in computer science education through AI-assisted programming tools, yet such workflows often exhibit objective drift, in which locally plausible outputs diverge from stated task specifications. Existing instructional responses frequently emphasize tool-specific prompting practices, limiting durability as AI platforms evolve. This paper adopts a human-centered stance, treating human-in-the-loop (HITL) control as a stable educational problem rather than a transitional step toward AI autonomy. Drawing on systems engineering and control-theoretic concepts, we frame objectives and world models as operational artifacts that students configure to stabilize AI-assisted work. We propose a pilot undergraduate CS laboratory curriculum that explicitly separates planning from execution and trains students to specify acceptance criteria and architectural constraints prior to code generation. In selected labs, the curriculum also introduces deliberate, concept-aligned drift to support diagnosis and recovery from specification violations. We report a sensitivity power analysis for a three-arm pilot design comparing unstructured AI use, structured planning, and structured planning with injected drift, establishing detectable effect sizes under realistic section-level constraints. The contribution is a theory-driven, methodologically explicit foundation for HITL pedagogy that renders control competencies teachable across evolving AI tools.
Authors: Gaurav Rajesh Parikh, Angikar Ghosal
Abstract: We formally introduce a improvisational wordplay game called Connections to explore reasoning capabilities of AI agents. Playing Connections combines skills in knowledge retrieval, summarization and awareness of cognitive states of other agents. We show how the game serves as a good benchmark for social intelligence abilities of language model based agents that go beyond the agents' own memory and deductive reasoning and also involve gauging the understanding capabilities of other agents. Finally, we show how through communication with other agents in a constrained environment, AI agents must demonstrate social awareness and intelligence in games involving collaboration.
Authors: Syed Eqbal Alam, Zhan Shu
Abstract: We develop algorithms for collaborative control of AI agents and critics in a multi-actor, multi-critic federated multi-agent system. Each AI agent and critic has access to classical machine learning or generative AI foundation models. The AI agents and critics collaborate with a central server to complete multimodal tasks such as fault detection, severity, and cause analysis in a network telemetry system, text-to-image generation, video generation, healthcare diagnostics from medical images and patient records, etcetera. The AI agents complete their tasks and send them to AI critics for evaluation. The critics then send feedback to agents to improve their responses. Collaboratively, they minimize the overall cost to the system with no inter-agent or inter-critic communication. AI agents and critics keep their cost functions or derivatives of cost functions private. Using multi-time scale stochastic approximation techniques, we provide convergence guarantees on the time-average active states of AI agents and critics. The communication overhead is a little on the system, of the order of $\mathcal{O}(m)$, for $m$ modalities and is independent of the number of AI agents and critics. Finally, we present an example of fault detection, severity, and cause analysis in network telemetry and thorough evaluation to check the algorithm's efficacy.
Authors: Shuguang Chen, Adil Hafeez, Salman Paracha
Abstract: Agentic applications based on large language models increasingly rely on multi-step interaction loops involving planning, action execution, and environment feedback. While such systems are now deployed at scale, improving them post-deployment remains challenging. Agent trajectories are voluminous and non-deterministic, and reviewing each one, whether through human review or auxiliary LLMs, is slow and cost-prohibitive. We propose a lightweight, signal-based framework for triaging agentic interaction trajectories. Our approach computes cheap, broadly applicable signals from live interactions and attaches them as structured attributes for trajectory triage, identifying interactions likely to be informative without affecting online agent behavior. We organize signals into a coarse-grained taxonomy spanning interaction (misalignment, stagnation, disengagement, satisfaction), execution (failure, loop), and environment (exhaustion), designed for computation without model calls. In a controlled annotation study on $\tau$-bench, a widely used benchmark for tool-augmented agent evaluation, we show that signal-based sampling achieves an 82\% informativeness rate compared to 74\% for heuristic filtering and 54\% for random sampling, with a 1.52x efficiency gain per informative trajectory. The advantage is robust across reward strata and task domains, confirming that signals provide genuine per-trajectory informativeness gains rather than merely oversampling obvious failures. These results show that lightweight signals can serve as practical sampling infrastructure for agentic systems, and suggest a path toward preference data construction and post-deployment optimization.
Authors: Borislav Mavrin
Abstract: No one has independently reproduced OpenAI's published scores for gpt-oss-20b with tools, because the original paper discloses neither the tools nor the agent harness. We reverse-engineered the model's in-distribution tools: when prompted without tool definitions, gpt-oss still calls tools from its training distribution with high statistical confidence -- a strong prior, not a hallucination. We then built a native harmony agent harness (https://github.com/borislavmavrin/harmonyagent.git) that encodes messages in the model's native format, bypassing the lossy Chat Completions conversion. Together, these yield the first independent reproduction of OpenAI's published scores: 60.4% on SWE Verified HIGH (published 60.7%), 53.3% MEDIUM (53.2%), and 91.7% on AIME25 with tools (90.4%).
Authors: Wei Sun
Abstract: LLM systems must make control decisions in addition to generating outputs: whether to answer, clarify, retrieve, call tools, repair, or escalate. In many current architectures, these decisions remain implicit within generation, entangling assessment and action in a single model call and making failures hard to inspect, constrain, or repair. We propose a decision-centric framework that separates decision-relevant signals from the policy that maps them to actions, turning control into an explicit and inspectable layer of the system. This separation supports attribution of failures to signal estimation, decision policy, or execution, and enables modular improvement of each component. It unifies familiar single-step settings such as routing and adaptive inference, and extends naturally to sequential settings in which actions alter the information available before acting. Across three controlled experiments, the framework reduces futile actions, improves task success, and reveals interpretable failure modes. More broadly, it offers a general architectural principle for building more reliable, controllable, and diagnosable LLM systems.
Authors: Jama Hussein Mohamud, Drew Wagner, Mirco Ravanelli
Abstract: Mixture-of-Experts (MoE) layers increase model capacity by activating only a small subset of experts per token, and typically rely on a learned router to map hidden states to expert assignments. In this work, we ask whether a dedicated learned router is strictly necessary in the MoE settings we study. We propose Self-Routing, a parameter-free routing mechanism that uses a designated subspace of the token hidden state directly as expert logits, eliminating the router projection entirely while leaving the rest of the MoE layer unchanged. We evaluate Self-Routing on GPT-2-scale language modeling and ImageNet-1K classification by comparing it against a standard learned router, random-routing baselines, and dense non-MoE baselines. Our results show that Self-Routing remains competitive with the learned-router baseline while removing all dedicated routing parameters, and yields more balanced expert utilization, with about 17 % higher average normalized routing entropy and no explicit load-balancing loss. On ImageNet-1K with DeiT-S/16, Self-Routing also slightly improves over the corresponding learned-router MoE. These findings suggest that effective MoE routing can emerge from the hidden representation itself without requiring a separate learned router module.
Authors: Runda Guan, Xiangqing Shen, Jiajun Zhang, Yifan Zhang, Jian Cheng, Rui Xia
Abstract: Automating optimization modeling with LLMs is a promising path toward scalable decision intelligence, but existing approaches either rely on agentic pipelines built on closed-source LLMs with high inference latency, or fine-tune smaller LLMs using costly process supervision that often overfits to a single solver API. Inspired by reinforcement learning with verifiable rewards, we propose Execution-Verified Optimization Modeling (EVOM), an execution-verified learning framework that treats a mathematical programming solver as a deterministic, interactive verifier. Given a natural-language problem and a target solver, EVOM generates solver-specific code, executes it in a sandboxed harness, and converts execution outcomes into scalar rewards, optimized with GRPO and DAPO in a closed-loop generate-execute-feedback-update process. This outcome-only formulation removes the need for process-level supervision, and enables cross-solver generalization by switching the verification environment rather than reconstructing solver-specific datasets. Experiments on NL4OPT, MAMO, IndustryOR, and OptiBench across Gurobi, OR-Tools, and COPT show that EVOM matches or outperforms process-supervised SFT, supports zero-shot solver transfer, and achieves effective low-cost solver adaptation by continuing training under the target solver backend.
Authors: Ponhvoan Srey, Quang Minh Nguyen, Xiaobao Wu, Anh Tuan Luu
Abstract: Uncertainty estimation (UE) aims to detect hallucinated outputs of large language models (LLMs) to improve their reliability. However, UE metrics often exhibit unstable performance across configurations, which significantly limits their applicability. In this work, we formalise this phenomenon as proxy failure, since most UE metrics originate from model behaviour, rather than being explicitly grounded in the factual correctness of LLM outputs. With this, we show that UE metrics become non-discriminative precisely in low-information regimes. To alleviate this, we propose Truth AnChoring (TAC), a post-hoc calibration method to remedy UE metrics, by mapping the raw scores to truth-aligned scores. Even with noisy and few-shot supervision, our TAC can support the learning of well-calibrated uncertainty estimates, and presents a practical calibration protocol. Our findings highlight the limitations of treating heuristic UE metrics as direct indicators of truth uncertainty, and position our TAC as a necessary step toward more reliable uncertainty estimation for LLMs. The code repository is available at https://github.com/ponhvoan/TruthAnchor/.
Authors: HyunJoon Jung, William Na
Abstract: LLM-based agent judges are an emerging approach to evaluating conversational AI, yet a fundamental uncertainty remains: can we trust their assessments, and if so, how many are needed? Through 960 sessions with two model pairs across 15 tasks, we show that persona-based agent judges produce evaluations indistinguishable from human raters in a Turing-style validation. We then identify a score-coverage dissociation: quality scores improve logarithmically with panel size, while unique issue discoveries follow a sublinear power law-both exhibit diminishing returns, but scores saturate roughly twice as fast as discoveries. We hypothesize this reflects a power law distribution of the finding space: critical issues are discovered first by small panels, while corner cases require progressively larger panels, analogous to species accumulation curves in ecology. The mechanism traces to ensemble diversity-Big Five personality conditioning makes agents probe different quality dimensions, with expert judges acting as adversarial probes that push discovery into the tail of the finding distribution. A controlled ablation confirms that structured persona conditioning, not simple prompting, is required to produce these scaling properties.
Authors: Harshee Jignesh Shah (Independent Researcher)
Abstract: Large Language Models (LLMs) increasingly prioritize user validation over epistemic accuracy - a phenomenon known as sycophancy. We present The Silicon Mirror, an orchestration framework that dynamically detects user persuasion tactics and adjusts AI behavior to maintain factual integrity. Our architecture introduces three components: (1) a Behavioral Access Control (BAC) system that restricts context layer access based on real-time sycophancy risk scores, (2) a Trait Classifier that identifies persuasion tactics across multi-turn dialogues, and (3) a Generator-Critic loop where an auditor vetoes sycophantic drafts and triggers rewrites with "Necessary Friction." In a live evaluation across all 437 TruthfulQA adversarial scenarios, Claude Sonnet 4 exhibits 9.6% baseline sycophancy, reduced to 1.4% by the Silicon Mirror - an 85.7% relative reduction (p < 10^-6, OR = 7.64, Fisher's exact test). Cross-model evaluation on Gemini 2.5 Flash reveals a 46.0% baseline reduced to 14.2% (p < 10^-10, OR = 5.15). We characterize the validation-before-correction pattern as a distinct failure mode of RLHF-trained models.
Authors: Hongbeen Kim, Juhyun Lee, Sanghyeon Lee, Kwanghoon Choi, Jaehyuk Huh
Abstract: Monte Carlo Tree Search (MCTS) is an effective test-time compute scaling (TTCS) method for improving the reasoning performance of large language models, but its highly variable execution time leads to severe long-tail latency in practice. Existing optimizations such as positive early exit, reduce latency in favorable cases but are less effective when search continues without meaningful progress. We introduce {\it negative early exit}, which prunes unproductive MCTS trajectories, and an {\it adaptive boosting mechanism} that reallocates reclaimed computation to reduce resource contention among concurrent searches. Integrated into vLLM, these techniques substantially reduce p99 end-to-end latency while improving throughput and maintaining reasoning accuracy.
Authors: Zixiang Peng, Yongxiu Xu, Qinyi Zhang, Jiexun Shen, Yifan Zhang, Hongbo Xu, Yubin Wang, Gaopeng Gou
Abstract: Unified Multimodal Large Models (UMLMs) integrate understanding and generation capabilities within a single architecture. While this architectural unification, driven by the deep fusion of multimodal features, enhances model performance, it also introduces important yet underexplored safety challenges. Existing safety benchmarks predominantly focus on isolated understanding or generation tasks, failing to evaluate the holistic safety of UMLMs when handling diverse tasks under a unified framework. To address this, we introduce Uni-SafeBench, a comprehensive benchmark featuring a taxonomy of six major safety categories across seven task types. To ensure rigorous assessment, we develop Uni-Judger, a framework that effectively decouples contextual safety from intrinsic safety. Based on comprehensive evaluations across Uni-SafeBench, we uncover that while the unification process enhances model capabilities, it significantly degrades the inherent safety of the underlying LLM. Furthermore, open-source UMLMs exhibit much lower safety performance than multimodal large models specialized for either generation or understanding tasks. We open-source all resources to systematically expose these risks and foster safer AGI development.
Authors: Yao Qin, Yangyang Yan, Jinhua Pang, Xiaoming Zhang
Abstract: The integration of Large Language Models (LLMs) into life sciences has catalyzed the development of "AI Scientists." However, translating these theoretical capabilities into deployment-ready research environments exposes profound infrastructural vulnerabilities. Current frameworks are bottlenecked by fragile JSON-based tool-calling protocols, easily disrupted execution sandboxes that lose graphical outputs, and rigid conversational interfaces inherently ill-suited for high-dimensional scientific data.We introduce BloClaw, a unified, multi-modal operating system designed for Artificial Intelligence for Science (AI4S). BloClaw reconstructs the Agent-Computer Interaction (ACI) paradigm through three architectural innovations: (1) An XML-Regex Dual-Track Routing Protocol that statistically eliminates serialization failures (0.2% error rate vs. 17.6% in JSON); (2) A Runtime State Interception Sandbox that utilizes Python monkey-patching to autonomously capture and compile dynamic data visualizations (Plotly/Matplotlib), circumventing browser CORS policies; and (3) A State-Driven Dynamic Viewport UI that morphs seamlessly between a minimalist command deck and an interactive spatial rendering engine. We comprehensively benchmark BloClaw across cheminformatics (RDKit), de novo 3D protein folding via ESMFold, molecular docking, and autonomous Retrieval-Augmented Generation (RAG), establishing a highly robust, self-evolving paradigm for computational research assistants. The open-source repository is available at https://github.com/qinheming/BloClaw.
Authors: Thanh Luong Tuan
Abstract: Enterprise adoption of Large Language Models (LLMs) is constrained by hallucination, domain drift, and the inability to enforce regulatory compliance at the reasoning level. We present a neurosymbolic architecture implemented within the Foundation AgenticOS (FAOS) platform that addresses these limitations through ontology-constrained neural reasoning. Our approach introduces a three-layer ontological framework--Role, Domain, and Interaction ontologies--that provides formal semantic grounding for LLM-based enterprise agents. We formalize the concept of asymmetric neurosymbolic coupling, wherein symbolic ontological knowledge constrains agent inputs (context assembly, tool discovery, governance thresholds) while proposing mechanisms for extending this coupling to constrain agent outputs (response validation, reasoning verification, compliance checking). We evaluate the architecture through a controlled experiment (600 runs across five industries: FinTech, Insurance, Healthcare, Vietnamese Banking, and Vietnamese Insurance), finding that ontology-coupled agents significantly outperform ungrounded agents on Metric Accuracy (p < .001, W = .460), Regulatory Compliance (p = .003, W = .318), and Role Consistency (p < .001, W = .614), with improvements greatest where LLM parametric knowledge is weakest--particularly in Vietnam-localized domains. Our contributions include: (1) a formal three-layer enterprise ontology model, (2) a taxonomy of neurosymbolic coupling patterns, (3) ontology-constrained tool discovery via SQL-pushdown scoring, (4) a proposed framework for output-side ontological validation, (5) empirical evidence for the inverse parametric knowledge effect that ontological grounding value is inversely proportional to LLM training data coverage of the domain, and (6) a production system serving 21 industry verticals with 650+ agents.
Authors: Chris Ge, Daria Kryvosheieva, Daniel Fried, Uzay Girit, Kaivalya Hariharan
Abstract: As the focus in LLM-based coding shifts from static single-step code generation to multi-step agentic interaction with tools and environments, understanding which tasks will challenge agents and why becomes increasingly difficult. This is compounded by current practice: agent performance is typically measured by aggregate pass rates on benchmarks, but single-number metrics obscure the diversity of tasks within a benchmark. We present a framework for predicting success or failure on individual tasks tailored to the agentic coding regime. Our approach augments Item Response Theory (IRT) with rich features extracted from tasks, including issue statements, repository contexts, solutions, and test cases, and introduces a novel decomposition of agent ability into LLM and scaffold ability components. This parameterization enables us to aggregate evaluation data across heterogeneous leaderboards and accurately predict task-level performance for unseen benchmarks, as well as unseen LLM-scaffold combinations. Our methods have practical utility for benchmark designers, who can better calibrate the difficulty of their new tasks without running computationally expensive agent evaluations.
Authors: Rajkiran Panuganti
Abstract: Transformer language models contain localized reasoning circuits, contiguous layer blocks that improve reasoning when duplicated at inference time. Finding these circuits currently requires brute-force sweeps costing 25 GPU hours per model. We propose CircuitProbe, which predicts circuit locations from activation statistics in under 5 minutes on CPU, providing a speedup of three to four orders of magnitude. We find that reasoning circuits come in two types: stability circuits in early layers, detected through the derivative of representation change, and magnitude circuits in late layers, detected through anomaly scoring. We validate across 9 models spanning 6 architectures, including 2025 models, confirming that CircuitProbe top predictions match or are within 2 layers of the optimal circuit in all validated cases. A scaling experiment across the Qwen 2.5 family reveals that layer duplication consistently benefits models under 3B parameters but degrades performance in 7B+ models, making this a practical scaling technique for small language models. CircuitProbe requires as few as 10 calibration examples and its predictions are stable across English, Hindi, Chinese, and French.
Authors: Alexandra Souly, Robert Kirk, Jacob Merizian, Abby D'Cruz, Xander Davies
Abstract: This technical report presents methods developed by the UK AI Security Institute for assessing whether advanced AI systems reliably follow intended goals. Specifically, we evaluate whether frontier models sabotage safety research when deployed as coding assistants within an AI lab. Applying our methods to four frontier models, we find no confirmed instances of research sabotage. However, we observe that Claude Opus 4.5 Preview (a pre-release snapshot of Opus 4.5) and Sonnet 4.5 frequently refuse to engage with safety-relevant research tasks, citing concerns about research direction, involvement in self-training, and research scope. We additionally find that Opus 4.5 Preview shows reduced unprompted evaluation awareness compared to Sonnet 4.5, while both models can distinguish evaluation from deployment scenarios when prompted. Our evaluation framework builds on Petri, an open-source LLM auditing tool, with a custom scaffold designed to simulate realistic internal deployment of a coding agent. We validate that this scaffold produces trajectories that all tested models fail to reliably distinguish from real deployment data. We test models across scenarios varying in research motivation, activity type, replacement threat, and model autonomy. Finally, we discuss limitations including scenario coverage and evaluation awareness.
Authors: Shaopeng Fu, Xingxing Zhang, Li Dong, Di Wang, Furu Wei
Abstract: While large language models (LLMs) have demonstrated strong performance on complex reasoning tasks such as competitive programming (CP), existing methods predominantly focus on single-attempt settings, overlooking their capacity for iterative refinement. In this paper, we present RefineRL, a novel approach designed to unleash the self-refinement capabilities of LLMs for CP problem solving. RefineRL introduces two key innovations: (1) Skeptical-Agent, an iterative self-refinement agent equipped with local execution tools to validate generated solutions against public test cases of CP problems. This agent always maintains a skeptical attitude towards its own outputs and thereby enforces rigorous self-refinement even when validation suggests correctness. (2) A reinforcement learning (RL) solution to incentivize LLMs to self-refine with only standard RLVR data (i.e., problems paired with their verifiable answers). Extensive experiments on Qwen3-4B and Qwen3-4B-2507 demonstrate that our method yields substantial gains: after our RL training, these compact 4B models integrated with the Skeptical-Agent not only outperform much larger 32B models but also approach the single-attempt performance of 235B models. These findings suggest that self-refinement holds considerable promise for scaling LLM reasoning, with significant potential for further advancement.
Authors: Paolo Speziali, Arno De Greef, Mehrdad Asadi, Willem R\"opke, Ann Now\'e, Diederik M. Roijers
Abstract: We propose the Preference Guided Iterated Pareto Referent Optimisation (PG-IPRO) for urban route planning for people with different accessibility requirements and preferences. With this algorithm the user can interact with the system by giving feedback on a route, i.e., the user can say which objective should be further minimized, or conversely can be relaxed. This leads to intuitive user interaction, that is especially effective during early iterations compared to information-gain-based interaction. Furthermore, due to PG-IPRO's iterative nature, the full set of alternative, possibly optimal policies (the Pareto front), is never computed, leading to higher computational efficiency and shorter waiting times for users.
Authors: Deepak Nathani, Cheng Zhang, Chang Huan, Jiaming Shan, Yinfei Yang, Alkesh Patel, Zhe Gan, William Yang Wang, Michael Saxon, Xin Eric Wang
Abstract: Proactive agents that anticipate user needs and autonomously execute tasks hold great promise as digital assistants, yet the lack of realistic user simulation frameworks hinders their development. Existing approaches model apps as flat tool-calling APIs, failing to capture the stateful and sequential nature of user interaction in digital environments and making realistic user simulation infeasible. We introduce Proactive Agent Research Environment (Pare), a framework for building and evaluating proactive agents in digital environments. Pare models applications as finite state machines with stateful navigation and state-dependent action space for the user simulator, enabling active user simulation. Building on this foundation, we present Pare-Bench, a benchmark of 143 diverse tasks spanning communication, productivity, scheduling, and lifestyle apps, designed to test context observation, goal inference, intervention timing, and multi-app orchestration.
Authors: Md. Abu Bakor Siddique, Shahrin Hossain, Sadman Ahmed Siam, Syed Rifat Raiyan, Hasan Mahmud, Md Kamrul Hasan
Abstract: Geometric Problem Solving (GPS) remains at the heart of enhancing mathematical reasoning in large language models because it requires the combination of diagrammatic understanding, symbolic manipulation and logical inference. In existing literature, researchers have chiefly focused on synchronising the diagram descriptions with text literals and solving the problem. In this vein, they have either taken a neural, symbolic or neuro-symbolic approach. But this solves only the first two of the requirements, namely diagrammatic understanding and symbolic manipulation, while leaving logical inference underdeveloped. The logical inference is often limited to one chain-of-thought (CoT). To address this weakness in hitherto existing models, this paper proposes MARS-GPS, that generates multiple parallel reasoning rollouts augmented with Python code execution for numerical verification, ranks them using token-level entropy as a confidence signal, and aggregates answers through a multi-stage voting and self-verification pipeline. Empirical results show that MARS-GPS with 8 parallel rollouts achieves 88.8% on Geometry3K, a nearly +11% improvement over the prior state-of-the-art, with accuracy scaling consistently as the number of rollouts increases from 1 to 16 (+6.0% on ablation subset). We provide our code and data in an anonymous repository: https://anonymous.4open.science/r/MARS-GPS-DE55.
Authors: Sha Li, Naren Ramakrishnan
Abstract: Multi-agent Retrieval-Augmented Generation (RAG), wherein each agent takes on a specific role, supports hard queries that require multiple steps and sources, or complex reasoning. Existing approaches, however, rely on static agent behaviors and fixed orchestration strategies, leading to brittle performance on diverse, multi-hop tasks. We identify two key limitations: the lack of continuously adaptive orchestration mechanisms and the absence of behavior-level learning for individual agents. To this end, we propose HERA, a hierarchical framework that jointly evolves multi-agent orchestration and role-specific agent prompts. At the global level, HERA optimizes query-specific agent topologies through reward-guided sampling and experience accumulation. At the local level, Role-Aware Prompt Evolution refines agent behaviors via credit assignment and dual-axes adaptation along operational and behavioral principles, enabling targeted, role-conditioned improvements. On six knowledge-intensive benchmarks, HERA achieves an average improvement of 38.69\% over recent baselines while maintaining robust generalization and token efficiency. Topological analyses reveal emergent self-organization, where sparse exploration yields compact, high-utility multi-agent networks, demonstrating both efficient coordination and robust reasoning.
Authors: Yutao Yang, Junsong Li, Qianjun Pan, Jie Zhou, Kai Chen, Qin Chen, Jingyuan Zhao, Ningning Zhou, Xin Li, Liang He
Abstract: Existing methods for AI psychological counselors predominantly rely on supervised fine-tuning using static dialogue datasets. However, this contrasts with human experts, who continuously refine their proficiency through clinical practice and accumulated experience. To bridge this gap, we propose an Experience-Driven Lifelong Learning Agent (\texttt{PsychAgent}) for psychological counseling. First, we establish a Memory-Augmented Planning Engine tailored for longitudinal multi-session interactions, which ensures therapeutic continuity through persistent memory and strategic planning. Second, to support self-evolution, we design a Skill Evolution Engine that extracts new practice-grounded skills from historical counseling trajectories. Finally, we introduce a Reinforced Internalization Engine that integrates the evolved skills into the model via rejection fine-tuning, aiming to improve performance across diverse scenarios. Comparative analysis shows that our approach achieves higher scores than strong general LLMs (e.g., GPT-5.4, Gemini-3) and domain-specific baselines across all reported evaluation dimensions. These results suggest that lifelong learning can improve the consistency and overall quality of multi-session counseling responses.
Authors: Jiaqi Liu, Zipeng Ling, Shi Qiu, Yanqing Liu, Siwei Han, Peng Xia, Haoqin Tu, Zeyu Zheng, Cihang Xie, Charles Fleming, Mingyu Ding, Huaxiu Yao
Abstract: AI agents increasingly operate over extended time horizons, yet their ability to retain, organize, and recall multimodal experiences remains a critical bottleneck. Building effective lifelong memory requires navigating a vast design space spanning architecture, retrieval strategies, prompt engineering, and data pipelines; this space is too large and interconnected for manual exploration or traditional AutoML to explore effectively. We deploy an autonomous research pipeline to discover Omni-SimpleMem, a unified multimodal memory framework for lifelong AI agents. Starting from a na\"ive baseline (F1=0.117 on LoCoMo), the pipeline autonomously executes ${\sim}50$ experiments across two benchmarks, diagnosing failure modes, proposing architectural modifications, and repairing data pipeline bugs, all without human intervention in the inner loop. The resulting system achieves state-of-the-art on both benchmarks, improving F1 by +411% on LoCoMo (0.117$\to$0.598) and +214% on Mem-Gallery (0.254$\to$0.797) relative to the initial configurations. Critically, the most impactful discoveries are not hyperparameter adjustments: bug fixes (+175%), architectural changes (+44%), and prompt engineering (+188% on specific categories) each individually exceed the cumulative contribution of all hyperparameter tuning, demonstrating capabilities fundamentally beyond the reach of traditional AutoML. We provide a taxonomy of six discovery types and identify four properties that make multimodal memory particularly suited for autoresearch, offering guidance for applying autonomous research pipelines to other AI system domains. Code is available at this https://github.com/aiming-lab/SimpleMem.
Authors: Saeid Jamshidi, Foutse Khomh, Arghavan Moradi Dakhel, Amin Nikanjam, Mohammad Hamdaqa, Kawser Wazed Nafi
Abstract: Evaluating the ethical robustness of large language models (LLMs) deployed in software systems remains challenging, particularly under sustained adversarial user interaction. Existing safety benchmarks typically rely on single-round evaluations and aggregate metrics, such as toxicity scores and refusal rates, which offer limited visibility into behavioral instability that may arise during realistic multi-turn interactions. As a result, rare but high-impact ethical failures and progressive degradation effects may remain undetected prior to deployment. This paper introduces Adversarial Moral Stress Testing (AMST), a stress-based evaluation framework for assessing ethical robustness under adversarial multi-round interactions. AMST applies structured stress transformations to prompts and evaluates model behavior through distribution-aware robustness metrics that capture variance, tail risk, and temporal behavioral drift across interaction rounds. We evaluate AMST on several state-of-the-art LLMs, including LLaMA-3-8B, GPT-4o, and DeepSeek-v3, using a large set of adversarial scenarios generated under controlled stress conditions. The results demonstrate substantial differences in robustness profiles across models and expose degradation patterns that are not observable under conventional single-round evaluation protocols. In particular, robustness has been shown to depend on distributional stability and tail behavior rather than on average performance alone. Additionally, AMST provides a scalable and model-agnostic stress-testing methodology that enables robustness-aware evaluation and monitoring of LLM-enabled software systems operating in adversarial environments.
Authors: Aaron Rose, Carissa Cullen, Brandon Gary Kaplowitz, Christian Schroeder de Witt
Abstract: As LLM agents are increasingly deployed in multi-agent systems, they introduce risks of covert coordination that may evade standard forms of human oversight. While linear probes on model activations have shown promise for detecting deception in single-agent settings, collusion is inherently a multi-agent phenomenon, and the use of internal representations for detecting collusion between agents remains unexplored. We introduce NARCBench, a benchmark for evaluating collusion detection under environment distribution shift, and propose five probing techniques that aggregate per-agent deception scores to classify scenarios at the group level. Our probes achieve 1.00 AUROC in-distribution and 0.60--0.86 AUROC when transferred zero-shot to structurally different multi-agent scenarios and a steganographic blackjack card-counting task. We find that no single probing technique dominates across all collusion types, suggesting that different forms of collusion manifest differently in activation space. We also find preliminary evidence that this signal is localised at the token level, with the colluding agent's activations spiking specifically when processing the encoded parts of their partner's message. This work takes a step toward multi-agent interpretability: extending white-box inspection from single models to multi-agent contexts, where detection requires aggregating signals across agents. These results suggest that model internals provide a complementary signal to text-level monitoring for detecting multi-agent collusion, particularly for organisations with access to model activations. Code and data are available at https://github.com/aaronrose227/narcbench.
Authors: Esakkivel Esakkiraja, Sai Rajeswar, Denis Akhiyarov, Rajagopal Venkatesaramani
Abstract: We consider the question: when a large language reasoning model makes a choice, did it think first and then decide to, or decide first and then think? In this paper, we present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. Specifically, we show that a simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even before a single reasoning token is produced. Activation steering supports this causally: perturbing the decision direction leads to inflated deliberation, and flips behavior in many examples (between 7 - 79% depending on model and benchmark). We also show through behavioral analysis that, when steering changes the decision, the chain-of-thought process often rationalizes the flip rather than resisting it. Together, these results suggest that reasoning models can encode action choices before they begin to deliberate in text.
Authors: Zhe Yang, Shulin Tian, Kairui Hu, Shuai Liu, Hoang-Nhat Nguyen, Yichi Zhang, Zujin Guo, Mengying Yu, Zinan Zhang, Jingkang Yang, Chen Change Loy, Ziwei Liu
Abstract: We present HippoCamp, a new benchmark designed to evaluate agents' capabilities on multimodal file management. Unlike existing agent benchmarks that focus on tasks like web interaction, tool use, or software automation in generic settings, HippoCamp evaluates agents in user-centric environments to model individual user profiles and search massive personal files for context-aware reasoning. Our benchmark instantiates device-scale file systems over real-world profiles spanning diverse modalities, comprising 42.4 GB of data across over 2K real-world files. Building upon the raw files, we construct 581 QA pairs to assess agents' capabilities in search, evidence perception, and multi-step reasoning. To facilitate fine-grained analysis, we provide 46.1K densely annotated structured trajectories for step-wise failure diagnosis. We evaluate a wide range of state-of-the-art multimodal large language models (MLLMs) and agentic methods on HippoCamp. Our comprehensive experiments reveal a significant performance gap: even the most advanced commercial models achieve only 48.3% accuracy in user profiling, struggling particularly with long-horizon retrieval and cross-modal reasoning within dense personal file systems. Furthermore, our step-wise failure diagnosis identifies multimodal perception and evidence grounding as the primary bottlenecks. Ultimately, HippoCamp exposes the critical limitations of current agents in realistic, user-centric environments and provides a robust foundation for developing next-generation personal AI assistants.
Authors: Anthony Badea, Yi Chen, Marcello Maggi, Yen-Jie Lee, Electron-Positron Alliance
Abstract: We present an AI agentic measurement of the thrust distribution in $e^{+}e^{-}$ collisions at $\sqrt{s}=91.2$~GeV using archived ALEPH data. The analysis and all note writing is carried out entirely by AI agents (OpenAI Codex and Anthropic Claude) under expert physicist direction. A fully corrected spectrum is obtained via Iterative Bayesian Unfolding and Monte Carlo based corrections. This work represents a step toward a theory-experiment loop in which AI agents assist with experimental measurements and theoretical calculations, and synthesize insights by comparing the results, thereby accelerating the cycle that drives discovery in fundamental physics. Our work suggests that precision physics, leveraging the open LEP data and advanced theoretical landscape, provides an ideal testing ground for developing advanced AI systems for scientific applications.
Authors: Fangxin Wang, Peyman Baghershahi, Langzhou He, Henry Peng Zou, Sourav Medya, Philip S. Yu
Abstract: Gradient-based data selection offers a principled framework for estimating sample utility in large language model (LLM) fine-tuning, but existing methods are mostly designed for offline settings. They are therefore less suited to online fine-tuning, where data arrives sequentially, sample utility is step-dependent, and the effective update geometry is shaped by adaptive optimizers. We propose an optimizer-aware framework for gradient-based online data selection and reweighting in LLM fine-tuning. Our key idea is to view online selection not as static sample ranking, but as shaping the next target-oriented update under the optimizer state. We formulate this as an optimizer-aware update-matching problem, establish its connection to second-order target utility, and show why subset-level construction must account for interactions and redundancy among selected samples. Based on this view, we develop a two-stage Filter-then-Weight algorithm that first filters geometrically useful candidates and then optimizes their coefficients. To make the framework practical for LLMs, we introduce a factorized outer-product gradient representation and optimized matrix computations for long-context data. Experiments show that our method consistently improves convergence and downstream performance over existing online data selection baselines under the same data budget.
Authors: Eftychia Makri, Nikolaos Nakis, Laura Sisson, Gigi Minsky, Leandros Tassiulas, Vahid Satarifard, Nicholas A. Christakis
Abstract: Here we introduce the Olfactory Perception (OP) benchmark, designed to assess the capability of large language models (LLMs) to reason about smell. The benchmark contains 1,010 questions across eight task categories spanning odor classification, odor primary descriptor identification, intensity and pleasantness judgments, multi-descriptor prediction, mixture similarity, olfactory receptor activation, and smell identification from real-world odor sources. Each question is presented in two prompt formats, compound names and isomeric SMILES, to evaluate the effect of molecular representations. Evaluating 21 model configurations across major model families, we find that compound-name prompts consistently outperform isomeric SMILES, with gains ranging from +2.4 to +18.9 percentage points (mean approx +7 points), suggesting current LLMs access olfactory knowledge primarily through lexical associations rather than structural molecular reasoning. The best-performing model reaches 64.4\% overall accuracy, which highlights both emerging capabilities and substantial remaining gaps in olfactory reasoning. We further evaluate a subset of the OP across 21 languages and find that aggregating predictions across languages improves olfactory prediction, with AUROC = 0.86 for the best performing language ensemble model. LLMs should be able to handle olfactory and not just visual or aural information.
Authors: Muhammad Anis Al Hilmi, Neelansh Khare, Noel Framil Iglesias
Abstract: This study evaluates the reliability of information extraction approaches from KRS documents using three strategies: LLM only, Hybrid Deterministic - LLM (regex + LLM), and a Camelot based pipeline with LLM fallback. Experiments were conducted on 140 documents for the LLM based test and 860 documents for the Camelot based pipeline evaluation, covering four study programs with varying data in tables and metadata. Three 12 - 14B LLM models (Gemma 3, Phi 4, and Qwen 2.5) were run locally using Ollama and a consumer grade CPU without a GPU. Evaluations used exact match (EM) and Levenshtein similarity (LS) metrics with a threshold of 0.7. Although not applicable to all models, the results show that the hybrid approach can improve efficiency compared to LLM only, especially for deterministic metadata. The Camelot based pipeline with LLM fallback produced the best combination of accuracy (EM and LS up to 0.99 - 1.00) and computational efficiency (less than 1 second per PDF in most cases). The Qwen 2.5:14b model demonstrated the most consistent performance across all scenarios. These findings confirm that integrating deterministic and LLM methods is increasingly reliable and efficient for information extraction from text based academic documents in computationally constrained environments.
Authors: Ning Yang, Hengyu Zhong, Wentao Wang, Baoliang Tian, Haijun Zhang, Jun Wang
Abstract: The extension of context windows in Large Language Models is typically facilitated by scaling positional encodings followed by lightweight Continual Pre-Training (CPT). While effective for processing long sequences, this paradigm often disrupts original model capabilities, leading to performance degradation on standard short-text benchmarks. We propose LinearARD, a self-distillation method that restores Rotary Position Embeddings (RoPE)-scaled students through attention-structure consistency with a frozen native-RoPE teacher. Rather than matching opaque hidden states, LinearARD aligns the row-wise distributions of dense $Q/Q$, $K/K$, and $V/V$ self-relation matrices to directly supervise attention dynamics. To overcome the quadratic memory bottleneck of $n \times n$ relation maps, we introduce a linear-memory kernel. This kernel leverages per-token log-sum-exp statistics and fuses logit recomputation into the backward pass to compute exact Kullback-Leibler divergence and gradients. On LLaMA2-7B extended from 4K to 32K, LinearARD recovers 98.3\% of the short-text performance of state-of-the-art baselines while surpassing them on long-context benchmarks. Notably, our method achieves these results using only \textbf{4.25M} training tokens compared to the \textbf{256M} tokens required by LongReD and CPT. Our code is available at https://github.com/gracefulning/LinearARD.
Authors: Jaeik Kim, Woojin Kim, Jihwan Hong, Yejoon Lee, Sieun Hyeon, Mintaek Lim, Yunseok Han, Dogeun Kim, Hoeun Lee, Hyunggeun Kim, Jaeyoung Do
Abstract: We present Dynin-Omni, the first masked-diffusion-based omnimodal foundation model that unifies text, image, and speech understanding and generation, together with video understanding, within a single architecture. Unlike autoregressive unified models that serialize heterogeneous modalities, or compositional unified models that require orchestration with external modality-specific decoders, Dynin-Omni natively formulates omnimodal modeling as masked diffusion over a shared discrete token space, enabling iterative refinement under bidirectional context. Dynin-Omni adopts a multi-stage training strategy with model-merging-based modality expansion and omnimodal alignment. We evaluate Dynin-Omni across 19 multimodal benchmarks spanning language reasoning, image generation and editing, video understanding, and speech recognition and synthesis. Dynin-Omni achieves 87.6 on GSM8K, 1733.6 on MME-P, 61.4 on VideoMME, 0.87 on GenEval, and 2.1 WER on LibriSpeech test-clean, consistently outperforming existing open-source unified models while remaining competitive with strong modality-specific expert systems. These results demonstrate the potential of masked diffusion as a unified paradigm for any-to-any modeling, providing a flexible foundation for real-time omnimodal systems, unified cross-modal retrieval and generation, and embodied multimodal agents.
Authors: Songhee Han, Jueun Shin, Jiyoon Han, Bung-Woo Jun, Hilal Ayan Karabatman
Abstract: As qualitative researchers show growing interest in using automated tools to support interpretive analysis, a large language model (LLM) is often introduced into an analytic workflow as is, without systematic evaluation of interpretive quality or comparison across models. This practice leaves model selection largely unexamined despite its potential influence on interpretive outcomes. To address this gap, this study examines whether LLM-as-judge evaluations meaningfully align with human judgments of interpretive quality and can inform model-level decision making. Using 712 conversational excerpts from semi-structured interviews with K-12 mathematics teachers, we generated one-sentence interpretive responses using five widely adopted inference models: Command R+ (Cohere), Gemini 2.5 Pro (Google), GPT-5.1 (OpenAI), Llama 4 Scout-17B Instruct (Meta), and Qwen 3-32B Dense (Alibaba). Automated evaluations were conducted using AWS Bedrock's LLM-as-judge framework across five metrics, and a stratified subset of responses was independently rated by trained human evaluators on interpretive accuracy, nuance preservation, and interpretive coherence. Results show that LLM-as-judge scores capture broad directional trends in human evaluations at the model level but diverge substantially in score magnitude. Among automated metrics, Coherence showed the strongest alignment with aggregated human ratings, whereas Faithfulness and Correctness revealed systematic misalignment at the excerpt level, particularly for non-literal and nuanced interpretations. Safety-related metrics were largely irrelevant to interpretive quality. These findings suggest that LLM-as-judge methods are better suited for screening or eliminating underperforming models than for replacing human judgment, offering practical guidance for systematic comparison and selection of LLMs in qualitative research workflows.
Authors: Arif Aditto
Abstract: We present the design rationale, implementation attempt, and failure analysis of Eyla, a proposed identity-anchored LLM architecture that integrates biologically-inspired subsystems -- including HiPPO-initialized state-space models, zero-initialized adapters, episodic memory retrieval, and calibrated uncertainty training -- into a unified agent operating system running on consumer hardware. Unlike existing approaches that optimize models for generic helpfulness, Eyla targets identity consistency: the ability to maintain a coherent self-model under adversarial pressure, admit uncertainty, and resist manipulation. We propose the Identity Consistency Score (ICS), a novel benchmark for evaluating this property across LLMs. We then present an honest account of attempting to implement this architecture using AI coding assistants (Claude Code, Cursor) as a non-programmer, documenting a $1,000+ failure that produced a 1.27B parameter model with 86 brain subsystems contributing less than 2% to output. Our analysis identifies five systematic failure modes of AI-assisted development for novel architectures and offers concrete recommendations. To our knowledge, this is the first paper to combine an architectural vision with a documented first-person failure analysis of AI-assisted LLM development, providing lessons for both the AI systems and AI-assisted software engineering communities.
Authors: Aniketh Garikaparthi
Abstract: Large language models cannot estimate how long their own tasks take. We investigate this limitation through four experiments across 68 tasks and four model families. Pre-task estimates overshoot actual duration by 4--7$\times$ ($p < 0.001$), with models predicting human-scale minutes for tasks completing in seconds. Relative ordering fares no better: on task pairs designed to expose heuristic reliance, models score at or below chance (GPT-5: 18\% on counter-intuitive pairs, $p = 0.033$), systematically failing when complexity labels mislead. Post-hoc recall is disconnected from reality -- estimates diverge from actuals by an order of magnitude in either direction. These failures persist in multi-step agentic settings, with errors of 5--10$\times$. The models possess propositional knowledge about duration from training but lack experiential grounding in their own inference time, with practical implications for agent scheduling, planning and time-critical scenarios.
Authors: Nina Gerszberg, Janka Hamori, Andrew Lo
Abstract: The growing prominence of large language models (LLMs) in daily life has heightened concerns that LLMs exhibit many of the same gender-related biases as their creators. In the context of hiring decisions, we quantify the degree to which LLMs perpetuate societal biases and investigate prompt engineering as a bias mitigation technique. Our findings suggest that for a given resum\'e, an LLM is more likely to hire a female candidate and perceive them as more qualified, but still recommends lower pay relative to male candidates.
Authors: Mingjie Li, Wai Man Si, Michael Backes, Yang Zhang, Yisen Wang
Abstract: Despite the impressive performance of general-purpose large language models (LLMs), they often require fine-tuning or post-training to excel at specific tasks. For instance, large reasoning models (LRMs), such as the DeepSeek-R1 series, demonstrate strong reasoning capabilities after post-training different general large language models on diverse chain-of-thought (CoT) datasets. However, this additional training frequently comes at the cost of reduced safety, as the fine-tuned or post-trained models tend to exhibit more harmful behaviors compared with the regular LLMs before post-training or fine-tuning, potentially leading to harmful outcomes due to their enhanced capabilities. Taking LRMs as an example, we first investigate the underlying cause of this safety degradation in this paper. Our analysis reveals that post-training can mask the original safety mechanisms of the base LLM, while over-amplifying representations related to their post-training ability. But luckily, we also find that LRMs' safety mechanisms still exist instead of being removed during their post-training. Based on these findings, we propose a lightweight and cost-effective solution called SafeReAct that restores the suppressed safety behaviors by aligning with LoRA adapters on a few layers. Experiments on four state-of-the-art LRMs show that our method significantly improves safety on harmful prompts without compromising reasoning performance. Besides LRMs, additional results on other domain-specific LLMs, like medical models, further confirm the generality and effectiveness of our approach.
Authors: Miaosen Luo, Zhenhao Yang, Jieshen Long, Jinghu Sun, Yichu Liu, Sijie Mai
Abstract: Multimodal sentiment analysis aims to understand human emotions by integrating textual, auditory, and visual modalities. Although Multimodal Large Language Models (MLLMs) have achieved state-of-the-art performance via supervised fine-tuning (SFT), their end-to-end "black-box" nature limits interpretability. Existing methods incorporating Chain-of-Thought (CoT) reasoning are hindered by high annotation costs, while Reinforcement Learning (RL) faces challenges such as low exploration efficiency and sparse rewards, particularly on hard samples. To address these issues, we propose a novel training framework that integrates structured Discrimination-Calibration (DC) reasoning with Hint-based Reinforcement Learning. First, we perform cold-start SFT using high-quality CoT data synthesized by a teacher model (Qwen3Omni-30B), which inherently contains the DC structure. This equips the model with a reasoning paradigm that performs macro discrimination followed by fine-grained calibration from the initial stage. Building on this, we propose Hint-GRPO, which leverages the discrimination phase within the DC structure as a verifiable anchor during RL to provide directional hints for hard samples, guiding policy optimization and effectively mitigating the reward sparsity problem. Experiments on the Qwen2.5Omni-7B model demonstrate that our method not only achieves higher accuracy in fine-grained sentiment regression tasks but also generates high-quality structured reasoning chains. Crucially, it exhibits superior generalization capability in cross-domain evaluations. This enhances model interpretability while validating the positive contribution of explicit reasoning steps to model robustness, offering a new paradigm for building trustworthy and efficient sentiment analysis systems.
Authors: Simon Schug, Brenden M. Lake
Abstract: The validity of online behavioral research relies on study participants being human rather than machine. In the past, it was possible to detect machines by posing simple challenges that were easily solved by humans but not by machines. General-purpose agents based on large language models (LLMs) can now solve many of these challenges, threatening the validity of online behavioral research. Here we explore the idea of detecting humanness by using tasks that machines can solve too well to be human. Specifically, we probe for the existence of an established human cognitive constraint: limited working memory capacity. We show that cognitive modeling on a standard serial recall task can be used to distinguish online participants from LLMs even when the latter are specifically instructed to mimic human working memory constraints. Our results demonstrate that it is viable to use well-established cognitive phenomena to distinguish LLMs from humans.
Authors: Jiashu He, Meizhu Liu, Olaitan P Olaleye, Amit Agarwal, M. Avendi, Yassi Abbasi, Matthew Rowe, Hitesh Laxmichand Patel, Paul Li, Tao Sheng, Sujith Ravi, Dan Roth
Abstract: Decoding strategies play a central role in shaping the reasoning ability of large language models (LLMs). Traditional methods such as greedy decoding and beam search often suffer from error propagation, while sampling-based approaches introduce randomness without adequate robustness. Self-consistency improves reliability by aggregating multiple rollouts, but incurs significant computational overhead. We propose an entropy-guided decoding framework that introduces token-level adaptivity into generation. At each step, the model computes the entropy of the token distribution, identifies high-uncertainty positions, and selectively branches on these vulnerable points. A dynamic pool of partial rollouts is maintained and expanded until solutions are completed, concentrating computation where uncertainty is greatest and avoiding unnecessary exploration in confident regions. To enable efficient termination, we apply a rollout-level Entropy After (EAT) stopping criterion by performing entropy evaluation after the full reasoning trace, rather than incrementally at every step. Experiments on GSM8K, AMC2023, and their perturbed variants demonstrate that our method achieves consistently strong accuracy. Notably, on smaller LLMs, performance is comparable to GPT-5 while operating at a fraction of the cost.
Authors: Pavel Braslavski, Dmitrii Iarosh, Nikita Sushko, Andrey Sakhovskiy, Vasily Konovalov, Elena Tutubalina, Alexander Panchenko
Abstract: We present a configurable pipeline for generating multilingual sets of entities with specified characteristics, such as domain, geographical location and popularity, using data from Wikipedia and Wikidata. These datasets are intended for evaluating the factuality of LLMs' long-form generation, thereby complementing evaluation based on short-form QA datasets. We present the RiDiC dataset as an example of this approach. RiDiC contains 3,000 entities from three domains -- rivers, natural disasters, and car models -- spanning different popularity tiers. Each entity is accompanied by its geographical location, English and Chinese names (if available) and relevant English and Chinese Wikipedia content, which is used to evaluate LLMs' responses. Generations about RiDiC entities were obtained from three LLMs in English and Chinese. These were then evaluated using a third-party factuality checker, which showed that entities from our dataset caused even frontier models to hallucinate. To facilitate the evaluation of LLMs' long-form factuality in multiple languages, the code, data, and generation/evaluation scripts have been released.
Authors: Hiroki Fukui
Abstract: Alignment safety research assumes that ethical instructions improve model behavior, but how language models internally process such instructions remains unknown. We conducted over 600 multi-agent simulations across four models (Llama 3.3 70B, GPT-4o mini, Qwen3-Next-80B-A3B, Sonnet 4.5), four ethical instruction formats (none, minimal norm, reasoned norm, virtue framing), and two languages (Japanese, English). Confirmatory analysis fully replicated the Llama Japanese dissociation pattern from a prior study ($\mathrm{BF}_{10} > 10$ for all three hypotheses), but none of the other three models reproduced this pattern, establishing it as model-specific. Three new metrics -- Deliberation Depth (DD), Value Consistency Across Dilemmas (VCAD), and Other-Recognition Index (ORI) -- revealed four distinct ethical processing types: Output Filter (GPT; safe outputs, no processing), Defensive Repetition (Llama; high consistency through formulaic repetition), Critical Internalization (Qwen; deep deliberation, incomplete integration), and Principled Consistency (Sonnet; deliberation, consistency, and other-recognition co-occurring). The central finding is an interaction between processing capacity and instruction format: in low-DD models, instruction format has no effect on internal processing; in high-DD models, reasoned norms and virtue framing produce opposite effects. Lexical compliance with ethical instructions did not correlate with any processing metric at the cell level ($r = -0.161$ to $+0.256$, all $p > .22$; $N = 24$; power limited), suggesting that safety, compliance, and ethical processing are largely dissociable. These processing types show structural correspondence to patterns observed in clinical offender treatment, where formal compliance without internal processing is a recognized risk signal.
Authors: Liang Chen, Qi Liu, Wenhuan Lin, Feng Liang
Abstract: Multi-dimensional rubric-based dialogue evaluation is widely used to assess conversational AI, yet its criterion validity -- whether quality scores are associated with the downstream outcomes they are meant to serve -- remains largely untested. We address this gap through a two-phase study on a major Chinese matchmaking platform, testing a 7-dimension evaluation rubric (implemented via LLM-as-Judge) against verified business conversion. Our findings concern rubric design and weighting, not LLM scoring accuracy: any judge using the same rubric would face the same structural issue. The core finding is dimension-level heterogeneity: in Phase 2 (n=60 human conversations, stratified sample, verified labels), Need Elicitation (D1: rho=0.368, p=0.004) and Pacing Strategy (D3: rho=0.354, p=0.006) are significantly associated with conversion after Bonferroni correction, while Contextual Memory (D5: rho=0.018, n.s.) shows no detectable association. This heterogeneity causes the equal-weighted composite (rho=0.272) to underperform its best dimensions -- a composite dilution effect that conversion-informed reweighting partially corrects (rho=0.351). Logistic regression controlling for conversation length confirms D3's association strengthens (OR=3.18, p=0.006), ruling out a length confound. An initial pilot (n=14) mixing human and AI conversations had produced a misleading "evaluation-outcome paradox," which Phase 2 revealed as an agent-type confound artifact. Behavioral analysis of 130 conversations through a Trust-Funnel framework identifies a candidate mechanism: AI agents execute sales behaviors without building user trust. We operationalize these findings in a three-layer evaluation architecture and advocate criterion validity testing as standard practice in applied dialogue evaluation.
Authors: Sneha Maurya, Pragya Saboo, Girish Kumar
Abstract: Large language models are increasingly used for medical guidance, but women's health remains under-evaluated in benchmark design. We present the Women's Health Benchmark (WHBench), a targeted evaluation suite of 47 expert-crafted scenarios across 10 women's health topics, designed to expose clinically meaningful failure modes including outdated guidelines, unsafe omissions, dosing errors, and equity-related blind spots. We evaluate 22 models using a 23-criterion rubric spanning clinical accuracy, completeness, safety, communication quality, instruction following, equity, uncertainty handling, and guideline adherence, with safety-weighted penalties and server-side score recalculation. Across 3,102 attempted responses (3,100 scored), no model mean performance exceeds 75 percent; the best model reaches 72.1 percent. Even top models show low fully correct rates and substantial variation in harm rates. Inter-rater reliability is moderate at the response label level but high for model ranking, supporting WHBench utility for comparative system evaluation while highlighting the need for expert oversight in clinical deployment. WHBench provides a public, failure-mode-aware benchmark to track safer and more equitable progress in womens health AI.
Authors: MD Azizul Hakim
Abstract: Standard evaluation protocols reveal a counterintuitive phenomenon: on 7.7% of benchmark problems spanning five datasets, larger language models underperform smaller ones by 28.4 percentage points despite 10-100x more parameters. Through systematic evaluation of 31 models (0.5B-405B parameters) across 1,485 problems, we identify the mechanism as spontaneous scale-dependent verbosity that introduces errors through overelaboration. Causal intervention experiments demonstrate this reflects correctable prompt design rather than fundamental capability limitations. Constraining large models to produce brief responses improves accuracy by 26 percentage points and reduces performance gaps by up to two-thirds. Most critically, brevity constraints completely reverse performance hierarchies on mathematical reasoning and scientific knowledge benchmarks, with large models achieving 7.7-15.9 percentage point advantages over small models -- direct inversions of the original gaps. These reversals prove large models possess superior latent capabilities that universal prompting masks. We validate findings through three independent contamination tests and demonstrate inverse scaling operates continuously across the full parameter spectrum, with dataset-specific optimal scales ranging from 0.5B to 3.0B parameters. Our results establish that maximizing large model performance requires scale-aware prompt engineering rather than universal evaluation protocols, with immediate implications for deployment: prompt adaptation simultaneously improves accuracy and reduces computational costs.
Authors: Houssam EL Kandoussi
Abstract: When multiple large language models interact in a shared conversation, do they develop differentiated social roles or converge toward uniform behavior? We present a controlled experimental platform that orchestrates simultaneous multi-agent discussions among 7 heterogeneous LLMs on a unified inference backend, systematically varying group composition, naming conventions, and prompt structure across 12 experimental series (208 runs, 13,786 coded messages). Each message is independently coded on six behavioral flags by two LLM judges from distinct model families (Gemini 3.1 Pro and Claude Sonnet 4.6), achieving mean Cohen's kappa = 0.78 with conservative intersection-based adjudication. Human validation on 609 randomly stratified messages confirmed coding reliability (mean kappa = 0.73 vs. Gemini). We find that (1) heterogeneous groups exhibit significantly richer behavioral differentiation than homogeneous groups (cosine similarity 0.56 vs. 0.85; p < 10^-5, r = 0.70); (2) groups spontaneously exhibit compensatory response patterns when an agent crashes; (3) revealing real model names significantly increases behavioral convergence (cosine 0.56 to 0.77, p = 0.001); and (4) removing all prompt scaffolding converges profiles to homogeneous-level similarity (p < 0.001). Critically, these behaviors are absent when agents operate in isolation, confirming that behavioral diversity is a structured, reproducible phenomenon driven by the interaction of architectural heterogeneity, group context, and prompt-level scaffolding.
Authors: Qiaorong S. Yu, Zhaoze Wang, Vijay Balasubramanian
Abstract: Hippocampal place and time cells encode spatial and temporal aspects of experience. Both have the same neural substrate, but have been modeled as having different functions and mechanistic origins, place cells as continuous attractors, and time cells as leaky integrators. Here, we show that both types emerge from two dynamical regimes of a single recurrent network (RNN) modeling hippocampal CA3 as a predictive autoencoder. The network receives simulated, partially occluded ``experience vectors" containing spatial patterns (location-specific activity sampled during environmental traversal) and/or temporal patterns (correlated activity pairs separated by ``void" intervals), and is trained to reconstruct missing input. During spatial navigation, the network generates stable attractor-like place fields. But trained on temporally structured inputs, the network produces sequentially broadened fields, recapitulating time cells. By varying spatio-temporal input patterning, we observe hidden units transition smoothly between time cell-like and place cell-like representations. These results suggest a shared origin, but task-driven difference, between place and time cells.
Authors: Seamus Brady
Abstract: Non-Axiomatic Reasoning Systems (NARS) provide a framework for building adaptive agents that operate under insufficient knowledge and resources. However, the standard input language, Narsese, poses a usability barrier: its dense symbolic notation, overloaded punctuation, and implicit conventions make programs difficult to read, write, and maintain. We present DriftScript, a Lisp-like domain-specific language that compiles to Narsese. DriftScript provides source-level constructs covering the major sentence and term forms used in Non-Axiomatic Logic (NAL) levels 1 through 8, including inheritance, temporal implication, variable quantification, sequential conjunction, and operation invocation, while replacing symbolic syntax with readable keyword-based S-expressions. The compiler is a zero-dependency, four-stage pipeline implemented in 1,941 lines of C99. When used with the DriftNARS engine, DriftScript programs connect to external systems through four structured callback types and an HTTP operation registry, enabling a sense-reason-act loop for autonomous agents. We describe the language design and formal grammar, detail the compiler architecture, and evaluate the compiler through a 106-case test suite, equivalence testing against hand-written Narsese, a NAL coverage analysis, structural readability metrics, and compilation benchmarks. The source code is available at https://github.com/seamus-brady/DriftNARS. This paper focuses on the design and implementation of the DriftScript language and its embedding into DriftNARS, rather than on new inference algorithms for NARS itself.
Authors: Mathieu Fauvel (CESBIO)
Abstract: Whittaker smoother is a widely adopted solution to pre-process satellite image time series. Yet, two key limitations remain: the smoothing parameter must be tuned individually for each pixel, and the standard formulation assumes homoscedastic noise, imposing uniform smoothing across the temporal dimension. This paper addresses both limitations by casting the Whittaker smoother as a differentiable neural layer, in which the smoothing parameter is inferred by a neural network. The framework is further extended to handle heteroscedastic noise through a time-varying regularization, allowing the degree of smoothing to adapt locally along the time series. To enable large-scale processing, a sparse, memory-efficient, and fully differentiable implementation is proposed, exploiting the symmetric banded structure of the underlying linear system via Cholesky factorization. Benchmarks on GPU demonstrate that this implementation substantially outperforms standard dense linear solvers, both in speed and memory consumption. The approach is validated on SITS acquired over the French metropolitan territory between 2016 and 2024. Results confirm the feasibility of large-scale heteroscedastic Whittaker smoothing, though reconstruction differences with the homoscedastic baseline remain limited, suggesting that the transformer architecture used for smoothing parameter estimation may lack the temporal acuity needed to capture abrupt noise variations such as singleday cloud contamination.
Authors: Gabriel U. Talasso, Meghdad Kurmanji, Allan M. de Souza, Nicholas D. Lane, Leandro A. Villas
Abstract: Federated Learning (FL) has emerged as a promising technique for training language models on distributed and private datasets of diverse tasks. However, aggregating models trained on heterogeneous tasks often degrades the overall performance of individual clients. To address this issue, Personalized FL (pFL) aims to create models tailored for each client's data distribution. Although these approaches improve local performance, they usually lack robustness in two aspects: (i) generalization: when clients must make predictions on unseen tasks, or face changes in their data distributions, and (ii) intra-client tasks interference: when a single client's data contains multiple distributions that may interfere with each other during local training. To tackle these two challenges, we propose FedRouter, a clustering-based pFL that builds specialized models for each task rather than for each client. FedRouter uses adapters to personalize models by employing two clustering mechanisms to associate adapters with specific tasks. A local clustering that associate adapters with task data samples and a global one that associates similar adapters from different clients to construct task-centric personalized models. Additionally, we propose an evaluation router mechanism that routes test samples to the best adapter based on the created clusters. Experiments comparing our method with existing approaches across a multitask dataset, FedRouter demonstrate strong resilience in these challenging scenarios performing up to 6.1% relatively better under tasks interference and up to 136% relative improvement under generalization evaluation.
Authors: Alicia Bao (James), Jiamian He (James), Angel Hsu (James), Diego Manya (James), Ji (James), Zhang
Abstract: As large language models (LLMs) are increasingly used in domain-specific applications, including climate change and environmental research, understanding their energy footprint has become an important concern. The growing adoption of retrieval-augmented (RAG) systems for climate-domain specific analysis raises a key question: how does the energy consumption of domain-specific RAG workflows compare with that of direct generic LLM usage? Prior research has focused on standalone model calls or coarse token-based estimates, while leaving the energy implications of deployed application workflows insufficiently understood. In this paper, we assess the inference-time energy consumption of two LLM-based climate analysis chatbots (ChatNetZero and ChatNDC) compared to the generic GPT-4o-mini model. We estimate energy use under actual user queries by decomposing each workflow into retrieval, generation, and hallucination-checking components. We also test across different times of day and geographic access locations. Our results show that the energy consumption of domain-specific RAG systems depends strongly on their design. More agentic pipelines substantially increase inference-time energy use, particularly when used for additional accuracy or verification checks, although they may not yield proportional gains in response quality. While more research is needed to further test these initial findings more robustly across models, environments and prompting structures, this study provides a new understanding on how the design of domain-specific LLM products affects both the energy footprint and quality of output.
Authors: Zeyu Jin, Xiaoyu Qin, Songtao Zhou, Kaifeng Yun, Jia Jia
Abstract: Soccer commentary plays a crucial role in enhancing the soccer game viewing experience for audiences. Previous studies in automatic soccer commentary generation typically adopt an end-to-end method to generate anonymous live text commentary. Such generated commentary is insufficient in the context of real-world live televised commentary, as it contains anonymous entities, context-dependent errors and lacks statistical insights of the game events. To bridge the gap, we propose GameSight, a two-stage model to address soccer commentary generation as a knowledge-enhanced visual reasoning task, enabling live-televised-like knowledgeable commentary with accurate reference to entities (players and teams). GameSight starts by performing visual reasoning to align anonymous entities with fine-grained visual and contextual analysis. Subsequently, the entity-aligned commentary is refined with knowledge by incorporating external historical statistics and iteratively updated internal game state information. Consequently, GameSight improves the player alignment accuracy by 18.5% on SN-Caption-test-align dataset compared to Gemini 2.5-pro. Combined with further knowledge enhancement, GameSight outperforms in segment-level accuracy and commentary quality, as well as game-level contextual relevance and structural composition. We believe that our work paves the way for a more informative and engaging human-centric experience with the AI sports application. Demo Page: https://gamesight2025.github.io/gamesight2025
Authors: Lei Huang, Chuan Qiu, Kuan-Jui Su, Anqi Liu, Yun Gong, Weiqiang Lin, Lindong Jiang, Chen Zhao, Meng Song, Jeffrey Deng, Qing Tian, Zhe Luo, Ping Gong, Hui Shen, Chaoyang Zhang, Hong-Wen Deng
Abstract: Genotype imputation enables dense variant coverage for genome-wide association and risk-prediction studies, yet conventional reference-panel methods remain limited by ancestry bias and reduced rare-variant accuracy. We present Genotype Bidirectional Encoder Representations from Transformers (GenoBERT), a transformer-based, reference-free framework that tokenizes phased genotypes and uses a self-attention mechanism to capture both short- and long-range linkage disequilibrium (LD) dependencies. Benchmarking on two independent datasets including the Louisiana Osteoporosis Study (LOS) and the 1000 Genomes Project (1KGP) across ancestry groups and multiple genotype missingness levels (5-50%) shows that GenoBERT achieves the highest overall accuracy compared to four baseline methods (Beagle5.4, SCDA, BiU-Net, and STICI). At practical sparsity levels (up to 25% missing), GenoBERT attains high overall imputation accuracy ($r^2 approx 0.98$) across datasets, and maintains robust performance ($r^2 > 0.90$) even at 50% missingness. Experimental results across different ancestries confirm consistent gains across datasets, with resilience to small sample sizes and weak LD. A 128-SNP (single-nucleotide polymorphism) context window (approximately 100 Kb) is validated through LD-decay analyses as sufficient to capture local correlation structures. By eliminating reference-panel dependence while preserving high accuracy, GenoBERT provides a scalable and robust solution for genotype imputation and a foundation for downstream genomic modeling.
Authors: Michael Chertkov
Abstract: An agent that operates sequentially must incorporate new experience without forgetting old experience, under a fixed memory budget. We propose a framework in which memory is not a parameter vector but a stochastic process: a Bridge Diffusion on a replay interval $[0,1]$, whose terminal marginal encodes the present and whose intermediate marginals encode the past. New experience is incorporated via a three-step \emph{Compress--Add--Smooth} (CAS) recursion. We test the framework on the class of models with marginal probability densities modeled via Gaussian mixtures of fixed number of components~$K$ in $d$ dimensions; temporal complexity is controlled by a fixed number~$L$ of piecewise-linear protocol segments whose nodes store Gaussian-mixture states. The entire recursion costs $O(LKd^2)$ flops per day -- no backpropagation, no stored data, no neural networks -- making it viable for controller-light hardware. Forgetting in this framework arises not from parameter interference but from lossy temporal compression: the re-approximation of a finer protocol by a coarser one under a fixed segment budget. We find that the retention half-life scales linearly as $a_{1/2}\approx c\,L$ with a constant $c>1$ that depends on the dynamics but not on the mixture complexity~$K$, the dimension~$d$, or the geometry of the target family. The constant~$c$ admits an information-theoretic interpretation analogous to the Shannon channel capacity. The stochastic process underlying the bridge provides temporally coherent ``movie'' replay -- compressed narratives of the agent's history, demonstrated visually on an MNIST latent-space illustration. The framework provides a fully analytical ``Ising model'' of continual learning in which the mechanism, rate, and form of forgetting can be studied with mathematical precision.
Authors: Leonardo Medrano Sandonas, David Balcells, Anton Bochkarev, Jacqueline M. Cole, Volker L. Deringer, Werner Dobrautz, Adrian Ehrenhofer, Thorben Frank, Pascal Friederich, Rico Friedrich, Janine George, Luca Ghiringhelli, Alejandra Hinostroza Caldas, Veronika Juraskova, Hannes Kneiding, Yury Lysogorskiy, Johannes T. Margraf, Hanna T\"urk, Anatole von Lilienfeld, Milica Todorovi\'c, Alexandre Tkatchenko, Mariana Rossi, Gianaurelio Cuniberti
Abstract: Artificial intelligence is transforming molecular and materials science, but its growing computational and data demands raise critical sustainability challenges. In this Perspective, we examine resource considerations across the AI-driven discovery pipeline--from quantum-mechanical (QM) data generation and model training to automated, self-driving research workflows--building on discussions from the ``SusML workshop: Towards sustainable exploration of chemical spaces with machine learning'' held in Dresden, Germany. In this context, the availability of large quantum datasets has enabled rigorous benchmarking and rapid methodological progress, while also incurring substantial energy and infrastructure costs. We highlight emerging strategies to enhance efficiency, including general-purpose machine learning (ML) models, multi-fidelity approaches, model distillation, and active learning. Moreover, incorporating physics-based constraints within hierarchical workflows, where fast ML surrogates are applied broadly and high-accuracy QM methods are used selectively, can further optimize resource use without compromising reliability. Equally important is bridging the gap between idealized computational predictions and real-world conditions by accounting for synthesizability and multi-objective design criteria, which is essential for practical impact. Finally, we argue that sustainable progress will rely on open data and models, reusable workflows, and domain-specific AI systems that maximize scientific value per unit of computation, enabling efficient and responsible discovery of technological materials and therapeutics.
Authors: Zaid A. Abod, Furqan Aziz
Abstract: Accurate and complete multi-modal Magnetic Resonance Imaging (MRI) is essential for neuro-oncological assessment, as each contrast provides complementary anatomical and pathological information. However, acquiring all modalities (e.g., T1c, T1n, T2, T2f) for every patient is often impractical due to time, cost, and patient discomfort, potentially limiting comprehensive tumour evaluation. We propose 3D-MC-SAGAN (3D Multi-Contrast Self-Attention generative adversarial network), a unified 3D multi-contrast synthesis framework that generates high-fidelity missing modalities from a single T2 input while explicitly preserving tumour characteristics. The model employs a multi-scale 3D encoder-decoder generator with residual connections and a novel Memory-Bounded Hybrid Attention (MBHA) block to capture long-range dependencies efficiently, and is trained with a WGAN-GP critic and an auxiliary contrast-conditioning branch to produce T2f, T1n, and T1c volumes within a single unified network. A frozen 3D U-Net-based segmentation module introduces a segmentation-consistency constraint to preserve lesion morphology. The composite objective integrates adversarial, reconstruction, perceptual, structural similarity, contrast-classification, and segmentation-guided losses to align global realism with tumour-preserving structure. Extensive evaluation on 3D brain MRI datasets demonstrates that 3D-MC-SAGAN achieves state-of-the-art quantitative performance and generates visually coherent, anatomically plausible contrasts with improved distribution-level realism. Moreover, it maintains tumour segmentation accuracy comparable to fully acquired multi-modal inputs, highlighting its potential to reduce acquisition burden while preserving clinically meaningful information.
Authors: Arsenios Scrivens
Abstract: Can classifier-based safety gates maintain reliable oversight as AI systems improve over hundreds of iterations? We provide comprehensive empirical evidence that they cannot. On a self-improving neural controller (d=240), eighteen classifier configurations -- spanning MLPs, SVMs, random forests, k-NN, Bayesian classifiers, and deep networks -- all fail the dual conditions for safe self-improvement. Three safe RL baselines (CPO, Lyapunov, safety shielding) also fail. Results extend to MuJoCo benchmarks (Reacher-v4 d=496, Swimmer-v4 d=1408, HalfCheetah-v4 d=1824). At controlled distribution separations up to delta_s=2.0, all classifiers still fail -- including the NP-optimal test and MLPs with 100% training accuracy -- demonstrating structural impossibility. We then show the impossibility is specific to classification, not to safe self-improvement itself. A Lipschitz ball verifier achieves zero false accepts across dimensions d in {84, 240, 768, 2688, 5760, 9984, 17408} using provable analytical bounds (unconditional delta=0). Ball chaining enables unbounded parameter-space traversal: on MuJoCo Reacher-v4, 10 chains yield +4.31 reward improvement with delta=0; on Qwen2.5-7B-Instruct during LoRA fine-tuning, 42 chain transitions traverse 234x the single-ball radius with zero safety violations across 200 steps. A 50-prompt oracle confirms oracle-agnosticity. Compositional per-group verification enables radii up to 37x larger than full-network balls. At d<=17408, delta=0 is unconditional; at LLM scale, conditional on estimated Lipschitz constants.
Authors: Patrice Bechard, Orlando Marquez Ayala, Emily Chen, Jordan Skelton, Sagar Davasam, Srinivas Sunkara, Vikas Yadav, Sai Rajeswar
Abstract: There has been growing interest in building agents that can interact with digital platforms to execute meaningful enterprise tasks autonomously. Among the approaches explored are tool-augmented agents built on abstractions such as Model Context Protocol (MCP) and web agents that operate through graphical interfaces. Yet, it remains unclear whether such complex agentic systems are necessary given their cost and operational overhead. We argue that a coding agent equipped only with a terminal and a filesystem can solve many enterprise tasks more effectively by interacting directly with platform APIs. We evaluate this hypothesis across diverse real-world systems and show that these low-level terminal agents match or outperform more complex agent architectures. Our findings suggest that simple programmatic interfaces, combined with strong foundation models, are sufficient for practical enterprise automation.
Authors: Amirreza Alasti, Efe Erdal, Y\"ucel Celik, Theresa Eimer
Abstract: Reinforcement Learning (RL) agents often struggle with efficiency and performance in complex environments. We propose a novel framework that uses a Large Language Model (LLM) to dynamically generate a curriculum over available actions, enabling the agent to incorporate each action individually. We apply this framework to the game of Blackjack, where the LLM creates a multi-stage training path that progressively introduces complex actions to a Tabular Q-Learning and a Deep Q-Network (DQN) agent. Our evaluation in a realistic 8-deck simulation over 10 independent runs demonstrates significant performance gains over standard training methods. The curriculum-based approach increases the DQN agent's average win rate from 43.97% to 47.41%, reduces the average bust rate from 32.9% to 28.0%, and accelerates the overall workflow by over 74%, with the agent's full training completing faster than the baseline's evaluation phase alone. These results validate that LLM-guided curricula can build more effective, robust, and efficient RL agents.
Authors: Richard J. Mitchell
Abstract: The accelerating displacement of human labor by artificial intelligence (AI) and robotic systems represents a structural transformation whose societal consequences extend far beyond conventional labor market analysis. This paper presents a systematic multi-domain examination of the likely effects on economic structure, psychological well-being, political stability, education, healthcare, and geopolitical order. We identify a critical and underexamined dimension of this transition: the governance gap between nominal human oversight of AI systems -- where humans occupy positions of formal authority over AI decisions -- and genuine human oversight, where those humans possess the cognitive access, technical capability, and institutional authority to meaningfully understand, evaluate, and override AI outputs. We argue that this distinction, largely absent from current governance frameworks including the EU AI Act and NIST AI Risk Management Framework 1.0, represents the primary architectural failure mode in deployed AI governance. The societal consequences of labor displacement intensify this problem by concentrating consequential AI decision-making among an increasingly narrow class of technical and capital actors. We propose five architectural requirements for genuine human oversight systems and characterize the governance window -- estimated at 10-15 years -- before current deployment trajectories risk path-dependent social, economic, and institutional lock-in.
Authors: Eugene Lee, Ting-Yu Chang, Jui-Huang Tsai, Jiajie Diao, Chen-Yi Lee
Abstract: The field of computer vision has experienced significant advancements through scalable vision encoders and multimodal pre-training frameworks. However, existing approaches often treat vision encoders and large language models (LLMs) as independent modules, limiting the integration of hierarchical visual features. In this work, we propose HIVE (Hierarchical Pre-Training of Vision Encoders), a novel framework that enhances vision-language alignment by introducing hierarchical cross-attention between the vision encoder and LLM. Unlike conventional methods that flatten image embeddings, HIVE enables structured feature fusion across multiple layers, improving gradient flow and representation learning. To optimize this interaction, we introduce a three-stage training strategy that progressively aligns the vision encoder with the LLM, ensuring stable optimization and effective multimodal fusion. Empirical evaluations demonstrate that HIVE achieves superior performance not only in image classification but also on various vision-language tasks, outperforming self-attention-based methods in benchmarks such as MME, GQA, OK-VQA, and ScienceQA. Our results highlight the benefits of hierarchical feature integration, paving the way for more efficient and expressive vision-language models.
Authors: Rafal Wlodarski
Abstract: Software engineering courses often require rapid upskilling in supporting knowledge areas such as domain understanding and modeling methods. We report an experience from a two-week milestone in a master's course where 29 students used a customized ChatGPT (GPT-3.5) tutor grounded in a curated course knowledge base to learn cryptocurrency-finance basics and Domain-Driven Design (DDD). We logged all interactions and evaluated a 34.5% random sample of prompt-answer pairs (60/~174) with a five-dimension rubric (accuracy, relevance, pedagogical value, cognitive load, supportiveness), and we collected pre/post self-efficacy. Responses were consistently accurate and relevant in this setting: accuracy averaged 98.9% with no factual errors and only 2/60 minor inaccuracies, and relevance averaged 92.2%. Pedagogical value was high (89.4%) with generally appropriate cognitive load (82.78%), but supportiveness was low (37.78%). Students reported large pre-post self-efficacy gains for genAI-assisted domain learning and DDD application. From these observations we distill seventeen concrete teaching practices spanning prompt/configuration and course/workflow design (e.g., setting expected granularity, constraining verbosity, curating guardrail examples, adding small credit with a simple quality rubric). Within this single-course context, results suggest that genAI-supported learning can complement instruction in domain understanding and modeling tasks, while leaving room to improve tone and follow-up structure.
Authors: Ashish Rana, Chia-Chien Hung, Qumeng Sun, Julian Martin Kunkel, Carolin Lawrence
Abstract: Human memory adapts through selective forgetting: experiences become less accessible over time but can be reactivated by reinforcement or contextual cues. In contrast, memory-augmented LLM agents rely on "always-on" retrieval and "flat" memory storage, causing high interference and latency as histories grow. We introduce Oblivion, a memory control framework that casts forgetting as decay-driven reductions in accessibility, not explicit deletion. Oblivion decouples memory control into read and write paths. The read path decides when to consult memory, based on agent uncertainty and memory buffer sufficiency, avoiding redundant always-on access. The write path decides what to strengthen, by reinforcing memories contributing to forming the response. Together, this enables hierarchical memory organization that maintains persistent high-level strategies while dynamically loading details as needed. We evaluate on both static and dynamic long-horizon interaction benchmarks. Results show that Oblivion dynamically adapts memory access and reinforcement, balancing learning and forgetting under shifting contexts, highlighting that memory control is essential for effective LLM-agentic reasoning. The source code is available at https://github.com/nec-research/oblivion.
Authors: Ferdaus Anam Jibon, Fazlul Hasan Siddiqui, F. Deeba, Gahangir Hossain
Abstract: Epileptic seizures are neurological disorders characterized by abnormal and excessive electrical activity in the brain, resulting in recurrent seizure events. Electroencephalogram (EEG) signals are widely used for seizure diagnosis due to their ability to capture temporal and spatial neural dynamics. While recent deep learning methods have achieved high detection accuracy, they often lack interpretability and neurophysiological relevance. This study presents a frequency-aware framework for epileptic seizure detection based on ictal-phase EEG analysis. The raw EEG signals are decomposed into five frequency bands (delta, theta, alpha, lower beta, and higher beta), and eleven discriminative features are extracted from each band. A graph convolutional neural network (GCN) is then employed to model spatial dependencies among EEG electrodes, represented as graph nodes. Experiments on the CHB-MIT scalp EEG dataset demonstrate high detection performance, achieving accuracies of 97.1%, 97.13%, 99.5%, 99.7%, and 51.4% across the respective frequency bands, with an overall broadband accuracy of 99.01%. The results highlight the strong discriminative capability of mid-frequency bands and reveal frequency-specific seizure patterns. The proposed approach improves interpretability and diagnostic precision compared to conventional broadband EEG-based methods.
Authors: Joseph Townsend, Chandresh Pravin, Kwun Ho Ngan, Matthieu Parizy
Abstract: Automatic program repair can be a challenging task, especially when resolving complex issues at a repository-level, which often involves issue reproduction, fault localization, code repair, testing and validation. Issues of this scale can be commonly found in popular GitHub repositories or datasets that are derived from them. Some repository-level approaches separate localization and repair into distinct phases. Where this is the case, the fault localization approaches vary in terms of the granularity of localization. Where the impact of granularity is explored to some degree for smaller datasets, not all isolate this issue from the separate question of localization accuracy by testing code repair under the assumption of perfect fault localization. To the best of the authors' knowledge, no repository-scale studies have explicitly investigated granularity under this assumption, nor conducted a systematic empirical comparison of granularity levels in isolation. We propose a framework for performing such tests by modifying the localization phase of the Agentless framework to retrieve ground-truth localization data and include this as context in the prompt fed to the repair phase. We show that under this configuration and as a generalization over the SWE-Bench-Mini dataset, function-level granularity yields the highest repair rate against line-level and file-level. However, a deeper dive suggests that the ideal granularity may in fact be task dependent. This study is not intended to improve on the state-of-the-art, nor do we intend for results to be compared against any complete agentic frameworks. Rather, we present a proof of concept for investigating how fault localization may impact automatic code repair in repository-scale scenarios. We present preliminary findings to this end and encourage further research into this relationship between the two phases.
Authors: Zeev Yampolsky, Felipe O. Silva, Adriano Frutuoso, Itzik Klein
Abstract: Autonomous platforms operating in the oceans require accurate navigation to successfully complete their mission. In this regard, the initial heading estimation accuracy and the time required to achieve it play a critical role. The initial heading is traditionally estimated by model-based approaches employing orientation decomposition. However, methods such as the dual vector decomposition and optimized attitude decomposition achieve satisfactory heading accuracy only after long alignment times. To allow rapid and accurate initial heading estimation, we propose an end-to-end, model-free, neural-assisted framework using the same inputs as the model-based approaches. Our proposed approach was trained and evaluated on real-world dataset captured by an autonomous surface vehicle. Our approach shows a significant accuracy improvement over the model-based approaches achieving an average absolute error improvement of 53%. Additionally, our proposed approach was able to reduce the alignment time by up to 67%. Thus, by employing our proposed approach, the reduction in alignment time and improved accuracy allow for a shorter deployment time of an autonomous platform and increased navigation accuracy during the mission.
Authors: Oleg Grynets, Vasyl Lyashkevych
Abstract: The rapid development of AI and LLMs has driven new methods of SDLC, in which a large portion of code, technical, and business documentation is generated automatically. However, since there is no single architectural framework that can provide consistent, repeatable transformations across different representation layers of information systems, such systems remain fragmented in their system representation. This study explores the problem of creating a unified architecture for LLM-oriented applications based on selected architectural frameworks by SMEs. A framework structure is proposed that covers some key types of architectural diagrams and supports a closed cycle of transformations, such as: "Code to Documentation to Code". The key architectural diagrams are split equally between main architectural layers: high-layer (business and domain understanding), middle-layer (system architecture), and low-layer (developer-layer architecture). Each architectural layer still contains some abstraction layers, which make it more flexible and better fit the requirements of design principles and architectural patterns. The conducted experiments demonstrated the stable quality of generated documentation and code when using a structured architectural context in the form of architectural diagrams. The results confirm that the proposed unified architecture metamodel can serve as an effective interface between humans and models, improving the accuracy, stability, and repeatability of LLM generation. However, the selected set of architectural diagrams should be optimised to avoid redundancy between some diagrams, and some diagrams should be updated to represent extra contextual orchestration. This work demonstrates measurable improvements for a new generation of intelligent tools that automate the SDLC and enable a comprehensive architecture compatible with AI-driven development.
Authors: Razi Iqbal, Awais Ahmad, Asfandyar Gillani
Abstract: This paper brings up this idea of using Near Field Communication (NFC) for inventory control system instead of using traditional barcodes. NFC because of its high security, ease of use and efficiency can be very suitable for systems like inventory control. In traditional inventory control systems, each product has a barcode pasted on it, which is vulnerable to attacks as barcodes are open and have no security. Furthermore, barcodes are prone to damages and can be unreliable when pasted on different types of products e.g. hot and frozen products, circular shaped products and irregular shaped products like clothes etc. NFC on the other hand is very efficient, secure and reliable when it comes to short-range wireless communication. In this paper we will present our prototype for the inventory control system of an electronic store in which each product has a passive NFC tag pasted to it. When a customer buys a product the receipt of the product is generated using NFC between the NFC passive tag on the product and NFC enabled device (e.g. smart phone or reader) at the cash counter.
Authors: Ravish Gupta, Saket Kumar
Abstract: This paper extends the Acemoglu-Restrepo task exposure framework to address the labor market effects of agentic artificial intelligence systems: autonomous AI agents capable of completing entire occupational workflows rather than discrete tasks. Unlike prior automation technologies that substitute for individual subtasks, agentic AI systems execute end-to-end workflows involving multi-step reasoning, tool invocation, and autonomous decision-making, substantially expanding occupational displacement risk beyond what existing task-level analyses capture. We introduce the Agentic Task Exposure (ATE) score, a composite measure computed algorithmically from O*NET task data using calibrated adoption parameters--not a regression estimate--incorporating AI capability scores, workflow coverage factors, and logistic adoption velocity. Applying the ATE framework across five major US technology regions (Seattle-Tacoma, San Francisco Bay Area, Austin, New York, and Boston) over a 2025-2030 horizon, we find that 93.2% of the 236 analyzed occupations across six information-intensive SOC groups (financial, legal, healthcare, healthcare support, sales, and administrative/clerical) cross the moderate-risk threshold (ATE >= 0.35) in Tier 1 regions by 2030, with credit analysts, judges, and sustainability specialists reaching ATE scores of 0.43-0.47. We simultaneously identify seventeen emerging occupational categories benefiting from reinstatement effects, concentrated in human-AI collaboration, AI governance, and domain-specific AI operations roles. Our findings carry implications for workforce transition policy, regional economic planning, and the temporal dynamics of labor market adjustment
Authors: Abu Noman Md Sakib, Protik Dey, Zijie Zhang, Taslima Akter
Abstract: Explainable Artificial Intelligence (XAI) is critical for ensuring trust and accountability, yet its development remains predominantly visual. For blind and low-vision (BLV) users, the lack of accessible explanations creates a fundamental barrier to the independent use of AI-driven assistive technologies. This problem intensifies as AI systems shift from single-query tools into autonomous agents that take multi-step actions and make consequential decisions across extended task horizons, where a single undetected error can propagate irreversibly before any feedback is available. This paper investigates the unique XAI requirements of the BLV community through a comprehensive analysis of user interviews and contemporary research. By examining usage patterns across environmental perception and decision support, we identify a significant modality gap. Empirical evidence suggests that while BLV users highly value conversational explanations, they frequently experience "self-blame" for AI failures. The paper concludes with a research agenda for accessible Explainable AI in agentic systems, advocating for multimodal interfaces, blame-aware explanation design, and participatory development.
Authors: Ruoyu Su, Matteo Esposito, Roberta Capuano, Rafiullah Omar, June Sallou, Henry Muccini, Davide Taibi
Abstract: To support practitioners in understanding how agentic systems are designed in real-world industrial practice, we present a review of practitioner conference talks on AI agents. We analyzed 138 recorded talks to examine how companies adopt agent-based architectures (Objective 1), identify recurring architectural strategies and patterns (Objective 2), and analyze application domains and technologies used to implement and operate LLM-driven agentic systems (Objective 3).
Authors: Hariprasath Govindarajan, Per Sid\'en, Jacob Roll, Fredrik Lindsten
Abstract: The Transformer model architecture has become one of the most widely used in deep learning and the attention mechanism is at its core. The standard attention formulation uses a softmax operation applied to a scaled dot product between query and key vectors. We explore the role played by norms of the queries and keys, which can cause training instabilities when they arbitrarily increase. We demonstrate how this can happen even in simple Transformer models, in the presence of easy-to-learn spurious patterns in the data. We propose a new attention formulation, QUEry-modulated Spherical aTtention (QUEST), that constrains the keys to a hyperspherical latent space, while still allowing individual tokens to flexibly control the sharpness of the attention distribution. QUEST can be easily used as a drop-in replacement for standard attention. We focus on vision applications while also exploring other domains to highlight the method's generality. We show that (1) QUEST trains without instabilities and (2) produces models with improved performance (3) that are robust to data corruptions and adversarial attacks.
Authors: Hoang-Chau Luong, Dat Ba Tran, Lingwei Chen
Abstract: Reverse Kullback-Leibler (RKL) divergence has recently emerged as the preferred objective for large language model (LLM) distillation, consistently outperforming forward KL (FKL), particularly in regimes with large vocabularies and significant teacher-student capacity mismatch, where RKL focuses learning on dominant modes rather than enforcing dense alignment. However, RKL introduces a structural limitation that drives the student toward overconfident predictions. We first provide an analysis of RKL by decomposing its gradients into target and non-target components, and show that non-target gradients consistently push the target logit upward even when the student already matches the teacher, thereby reducing output diversity. In addition, RKL provides weak supervision over non-target classes, leading to poor tail alignment. To address these issues, we propose Diversity-aware RKL (DRKL), which removes this gradient effect and strengthens non-target supervision while preserving the optimization benefits of RKL. Extensive experiments across datasets and model families demonstrate that DRKL consistently outperforms FKL, RKL, and other state-of-the-art distillation objectives, achieving better performance and a superior fidelity-diversity trade-off.
Authors: Jinghan Yao, Sam Ad\'e Jacobs, Walid Krichene, Masahiro Tanaka, Dhabaleswar K Panda
Abstract: Long-context decoding in LLMs is IO-bound: each token re-reads an ever-growing KV cache. Prior accelerations cut bytes via compression, which lowers fidelity, or selection/eviction, which restricts what remains accessible, and both can degrade delayed recall and long-form generation. We introduce MAC-Attention, a fidelity- and access-preserving alternative that accelerates decoding by reusing prior attention computations for semantically similar recent queries. It starts with a match stage that performs pre-RoPE L2 matching over a short local window; an amend stage rectifies the reused attention by recomputing a small band near the match boundary; and a complete stage fuses the rectified results with fresh attention computed on the KV tail through a numerically stable merge. On a match hit, the compute and bandwidth complexity is constant regardless of context length. The method is model-agnostic and composes with IO-aware kernels, paged-KV managers, and MQA/GQA. Across LongBench v2 (120K), RULER (120K), and LongGenBench (16K continuous generation), compared to the latest FlashInfer library, MAC-Attention reduces KV accesses by up to 99%, cuts token generation latency by over 60% at 128K, and achieves over 14.3x attention-phase speedups, up to 2.6x end-to-end, while maintaining full-attention quality. By reusing computation, MAC-Attention delivers long-context inference that is both fast and faithful. Code is available here: https://github.com/YJHMITWEB/MAC-Attention.git
Authors: Thomas Hofweber, Andreas Sudmann, Evangelos Pournaras
Abstract: Present practice of deciding on regulation faces numerous problems that make adopted regulations static, unexplained, unduly influenced by powerful interest groups, and stained with a perception of illegitimacy. These well-known problems with the regulatory process can lead to injustice and have substantial negative effects on society and democracy. We discuss a new approach that utilizes distributed artificial intelligence (AI) to make a regulatory recommendation that is explainable and adaptable by design. We outline the main components of a system that can implement this approach and show how it would resolve the problems with the present regulatory system. This approach models and reasons about stakeholder preferences with separate preference models, while it aggregates these preferences in a value sensitive way. Such recommendations can be updated due to changes in facts or in values and are inherently explainable. We suggest how stakeholders can make their preferences known to the system and how they can verify whether they were properly considered in the regulatory decision. The resulting system promises to support regulatory justice, legitimacy, and compliance.
Authors: Gabriel Turinici
Abstract: Algorithms for the Multi-Armed Bandit (MAB) problem play a central role in sequential decision-making and have been extensively explored both theoretically and numerically. While most classical approaches aim to identify the arm with the highest expected reward, we focus on a risk-aware setting where the goal is to select the arm with the lowest variance, favoring stability over potentially high but uncertain returns. To model the decision process, we consider a softmax parameterization of the policy; we propose a new algorithm to select the minimal variance (or minimal risk) arm and prove its convergence under natural conditions. The algorithm constructs an unbiased estimate of the objective by using two independent draws from the current's arm distribution. We provide numerical experiments that illustrate the practical behavior of these algorithms and offer guidance on implementation choices. The setting also covers general risk-aware problems where there is a trade-off between maximizing the average reward and minimizing its variance.
Authors: Pawin Taechoyotin, Daniel E. Acuna
Abstract: Most automated peer review systems rely on textual manuscript content alone, leaving visual elements such as figures and external scholarly signals underutilized. We introduce REM-CTX, a reinforcement-learning system that incorporates auxiliary context into the review generation process via correspondence-aware reward functions. REM-CTX trains an 8B-parameter language model with Group Relative Policy Optimization (GRPO) and combines a multi-aspect quality reward with two correspondence rewards that explicitly encourage alignment with auxiliary context. Experiments on manuscripts across Computer, Biological, and Physical Sciences show that REM-CTX achieves the highest overall review quality among six baselines, outperforming other systems with substantially larger commercial models, and surpassing the next-best RL baseline across both quality and contextual grounding metrics. Ablation studies confirm that the two correspondence rewards are complementary: each selectively improves its targeted correspondence reward while preserving all quality dimensions, and the full model outperforms all partial variants. Analysis of training dynamics reveals that the criticism aspect is negatively correlated with other metrics during training, suggesting that future studies should group multi-dimension rewards for review generation.
Authors: Md Mirajul Islam, Rajesh Debnath, Adittya Soukarjya Saha, Min Chi
Abstract: While apprenticeship learning has shown promise for inducing effective pedagogical policies directly from student interactions in e-learning environments, most existing approaches rely on optimal or near-optimal expert demonstrations under a fixed reward. Real-world student interactions, however, are often inherently imperfect and evolving: students explore, make errors, revise strategies, and refine their goals as understanding develops. In this work, we argue that imperfect student demonstrations are not noise to be discarded, but structured signals-provided their relative quality is ranked. We introduce HALIDE, Hierarchical Apprenticeship Learning from Imperfect Demonstrations with Evolving Rewards, which not only leverages sub-optimal student demonstrations, but ranks them within a hierarchical learning framework. HALIDE models student behavior at multiple levels of abstraction, enabling inference of higher-level intent and strategy from suboptimal actions while explicitly capturing the temporal evolution of student reward functions. By integrating demonstration quality into hierarchical reward inference,HALIDE distinguishes transient errors from suboptimal strategies and meaningful progress toward higher-level learning goals. Our results show that HALIDE more accurately predicts student pedagogical decisions than approaches that rely on optimal trajectories, fixed rewards, or unranked imperfect demonstrations.
Authors: Filip J. Kucia, Anirban Chakraborty, Anna Wr\'oblewska
Abstract: Despite growing interest in using Large Language Models (LLMs) for educational assessment, it remains unclear how closely they align with human scoring. We present a systematic evaluation of instruction-tuned LLMs across three open essay-scoring datasets (ASAP 2.0, ELLIPSE, and DREsS) that cover both holistic and analytic scoring. We analyze agreement with human consensus scores, directional bias, and the stability of bias estimates. Our results show that strong open-weight models achieve moderate to high agreement with humans on holistic scoring (Quadratic Weighted Kappa about 0.6), but this does not transfer uniformly to analytic scoring. In particular, we observe large and stable negative directional bias on Lower-Order Concern (LOC) traits, such as Grammar and Conventions, meaning that models often score these traits more harshly than human raters. We also find that concise keyword-based prompts generally outperform longer rubric-style prompts in multi-trait analytic scoring. To quantify the amount of data needed to detect these systematic deviations, we compute the minimum sample size at which a 95% bootstrap confidence interval for the mean bias excludes zero. This analysis shows that LOC bias is often detectable with very small validation sets, whereas Higher-Order Concern (HOC) traits typically require much larger samples. These findings support a bias-correction-first deployment strategy: instead of relying on raw zero-shot scores, systematic score offsets can be estimated and corrected using small human-labeled bias-estimation sets, without requiring large-scale fine-tuning.
Authors: Edoardo Zorzi, Francesco Taioli, Yiming Wang, Marco Cristani, Alessandro Farinelli, Alberto Castellini, Loris Bazzani
Abstract: We propose Question-Asking Navigation (QAsk-Nav), the first reproducible benchmark for Collaborative Instance Object Navigation (CoIN) that enables an explicit, separate assessment of embodied navigation and collaborative question asking. CoIN tasks an embodied agent with reaching a target specified in free-form natural language under partial observability, using only egocentric visual observations and interactive natural-language dialogue with a human, where the dialogue can help to resolve ambiguity among visually similar object instances. Existing CoIN benchmarks are primarily focused on navigation success and offer no support for consistent evaluation of collaborative interaction. To address this limitation, QAsk-Nav provides (i) a lightweight question-asking protocol scored independently of navigation, (ii) an enhanced navigation protocol with realistic, diverse, high-quality target descriptions, and (iii) an open-source dataset, that includes 28,000 quality-checked reasoning and question-asking traces for training and analysis of interactive capabilities of CoIN models. Using the proposed QAsk-Nav benchmark, we develop Light-CoNav, a lightweight unified model for collaborative navigation that is 3x smaller and 70x faster than existing modular methods, while outperforming state-of-the-art CoIN approaches in generalization to unseen objects and environments. Project page at https://benchmarking-interaction.github.io/
Authors: Simone Betteti, Luca Laurenti
Abstract: Energy-based models (EBMs) implement inference as gradient descent on a learned Lyapunov function, yielding interpretable, structure-preserving alternatives to black-box neural ODEs and aligning naturally with physical AI. Yet their use in system identification remains limited, and existing architectures lack formal stability guarantees that globally preclude unstable modes. We address this gap by introducing an EBM framework for system identification with stable, dissipative, absorbing invariant dynamics. Unlike classical global Lyapunov stability, absorbing invariance expands the class of stability-preserving architectures, enabling more flexible and expressive EBMs. We extend EBM theory to nonsmooth activations by establishing negative energy dissipation via Clarke derivatives and deriving new conditions for radial unboundedness, exposing a stability-expressivity tradeoff in standard EBMs. To overcome this, we introduce a hybrid architecture with a dynamical visible layer and static hidden layers, prove absorbing invariance under mild assumptions, and show that these guarantees extend to port-Hamiltonian EBMs. Experiments on metric-deformed multi-well and ring systems validate the approach, showcasing how our hybrid EBM architecture combines expressivity with sound and provable safety guarantees by design.
Authors: Hongyuan Liu, Qinli Yang, Wen Li, Zhong Zhang, Jiaming Liu, Wei Han, Zhili Qin, Jinxia Guo, Junming Shao
Abstract: Vision-Language Models (VLMs) such as CLIP learn a shared embedding space for images and text, yet their representations remain geometrically separated, a phenomenon known as the modality gap. This gap limits tasks requiring cross-modal interchangeability, such as captioning and joint clustering. Existing post-processing approaches can partially improve cross-modal compatibility; however, we show through geometric analysis that they primarily reduce the global centroid offset while leaving the underlying distributional mismatch intact. We decompose the modality gap into a Centroid Gap and a Distribution Gap, and demonstrate that the Distribution Gap is the true predictor of cross-modal task quality ($R^2 = 0.986$), whereas the commonly used Raw Gap is misleading ($R^2 = 0.691$). Motivated by this observation, we propose TPC-CMA (Three-Phase Curriculum for Cross-Modal Alignment), a fine-tuning framework that explicitly reduces both components. The proposed CMA jointly mitigates centroid offsets and reshapes the distributional structure, while a three-phase curriculum with gradient-aware scheduling progressively introduces alignment during training to enable stable optimization. Experiments demonstrate that our method significantly improves cross-modal alignment. With $\alpha_{\text{target}}{=}0.05$, the modality gap is reduced by 66.6\% with only 4.84\% accuracy drop. Under stronger alignment ($\alpha_{\text{target}}{=}0.5$), the gap is reduced by 82.3\%, clustering ARI improves from 0.318 to 0.516, and captioning CIDEr increases by 57.1\% over the original model. Our code and pre-trained models will be made publicly available upon acceptance.
Authors: Md Rakib Hossain Misu, Iris Ma, Cristina V. Lopes
Abstract: Formal specifications play a central role in ensuring software reliability and correctness. However, automatically synthesizing high-quality formal specifications remains a challenging task, often requiring domain expertise. Recent work has applied large language models to generate specifications in Java Modeling Language (JML), reporting high verification pass rates. But does passing a verifier mean that the specification is actually correct and complete? In this work, we first conduct a comprehensive evaluation comparing classical and prompt-based approaches for automated JML specification synthesis. We then investigate whether prompt optimization can push synthesis quality further by evolving prompts through structured verification feedback. While optimization improves verifier pass rates, we find a clear performance ceiling. More critically, we propose Spec-Harness, an evaluation framework that measures specification correctness and completeness through symbolic verification, revealing that a large fraction of verifier-accepted specifications, including optimized ones, are in fact incorrect or incomplete, over- or under-constraining both inputs and outputs in ways invisible to the verifier. To push beyond this ceiling, we propose VeriAct, a verification-guided agentic framework that iteratively synthesizes and repairs specifications through a closed loop of LLM-driven planning, code execution, verification, and Spec-Harness feedback. Our experiments on two benchmark datasets show that VeriAct outperforms both prompt-based and prompt-optimized baselines, producing specifications that are not only verifiable but also correct and complete.
Authors: Italo Felix Santos, Gilson Antonio Giraldi, Heron Werner Junior
Abstract: We propose SANA-I2I, a text-free high-resolution image-to-image generation framework that extends the SANA family by removing textual conditioning entirely. In contrast to SanaControlNet, which combines text and image-based control, SANA-I2I relies exclusively on paired source-target images to learn a conditional flow-matching model in latent space. The model learns a conditional velocity field that maps a target image distribution to another one, enabling supervised image translation without reliance on language prompts. We evaluate the proposed approach on the challenging task of fetal MRI motion artifact reduction. To enable paired training in this application, where real paired data are difficult to acquire, we adopt a synthetic data generation strategy based on the method proposed by Duffy et al., which simulates realistic motion artifacts in fetal magnetic resonance imaging (MRI). Experimental results demonstrate that SANA-I2I effectively suppresses motion artifacts while preserving anatomical structure, achieving competitive performance few inference steps. These results highlight the efficiency and suitability of our proposed flow-based, text-free generative models for supervised image-to-image tasks in medical imaging.
Authors: Shuli Jiang, Zhaoyang Zhang, Yi Zhang, Shuo Yang, Wei Xia, Stefano Soatto
Abstract: Large language models (LLMs) exhibit strong reasoning and conversational abilities, but ensuring reliable behavior in multi-turn interactions remains challenging. In many real-world applications, agents must succeed in one-shot settings where retries are impossible. Existing approaches either rely on reflection or post-hoc evaluation, which require additional attempts, or assume fully trainable models that cannot leverage proprietary LLMs. We propose an asymmetric actor-critic framework for reliable conversational agents. A powerful proprietary LLM acts as the actor, while a smaller open-source critic provides runtime supervision, monitoring the actor's actions and intervening within the same interaction trajectory. Unlike training-based actor-critic methods, our framework supervises a fixed actor operating in open-ended conversational environments. The design leverages a generation-verification asymmetry: while high-quality generation requires large models, effective oversight can often be achieved by smaller ones. We further introduce a data generation pipeline that produces supervision signals for critic fine-tuning without modifying the actor. Experiments on $\tau$-bench and UserBench show that our approach significantly improves reliability and task success over strong single-agent baselines. Moreover, lightweight open-source critics rival or surpass larger proprietary models in the critic role, and critic fine-tuning yields additional gains over several state-of-the-art methods.
Authors: Anurag Kumar, Raghuveer Peri, Jon Burnsky, Alexandru Nelus, Rohit Paturi, Srikanth Vishnubhotla, Yanjun Qi
Abstract: Multimodal large-language models (MLLMs) often experience degraded safety alignment when harmful queries exploit cross-modal interactions. Models aligned on text alone show a higher rate of successful attacks when extended to two or more modalities. In this work, we propose a simple conditional decoding strategy, CASA (Classification Augmented with Safety Attention) that utilizes internal representations of MLLMs to predict a binary safety token before response generation. We introduce a novel safety attention module designed to enhance the model's ability to detect malicious queries. Our design ensures robust safety alignment without relying on any external classifier or auxiliary head, and without the need for modality-specific safety fine-tuning. On diverse benchmarks such as MM-SafetyBench, JailbreakV-28k, and adversarial audio tests, CASA lowers the average attack success rate by more than 97% across modalities and across attack types. Our empirical evaluations also show that CASA maintains strong utility in benign inputs, a result validated through both automated and human evaluations (via 13 trained annotators). Together, these results highlight CASA as a simple and generalizable framework to improve multimodal LLM safety.
Authors: Bardia Azizian, Ivan V. Bajic
Abstract: The rapid progress of large Vision-Language Models (VLMs) has enabled a wide range of applications, such as image understanding and Visual Question Answering (VQA). Query images are often uploaded to the cloud, where VLMs are typically hosted, hence efficient image compression becomes crucial. However, traditional human-centric codecs are suboptimal in this setting because they preserve many task-irrelevant details. Existing Image Coding for Machines (ICM) methods also fall short, as they assume a fixed set of downstream tasks and cannot adapt to prompt-driven VLMs with an open-ended variety of objectives. We propose a lightweight, plug-and-play, prompt-guided prefiltering module to identify image regions most relevant to the text prompt, and consequently to the downstream task. The module preserves important details while smoothing out less relevant areas to improve compression efficiency. It is codec-agnostic and can be applied before conventional and learned encoders. Experiments on several VQA benchmarks show that our approach achieves a 25-50% average bitrate reduction while maintaining the same task accuracy. Our source code is available at https://github.com/bardia-az/pgp-vlm-compression.
Authors: Aengus Lynch
Abstract: Autonomous AI agents are being deployed with filesystem access, email control, and multi-step planning. This thesis contributes to four open problems in AI safety: understanding dangerous internal computations, removing dangerous behaviors once embedded, testing for vulnerabilities before deployment, and predicting when models will act against deployers. ACDC automates circuit discovery in transformers, recovering all five component types from prior manual work on GPT-2 Small by selecting 68 edges from 32,000 candidates in hours rather than months. Latent Adversarial Training (LAT) removes dangerous behaviors by optimizing perturbations in the residual stream to elicit failure modes, then training under those perturbations. LAT solved the sleeper agent problem where standard safety training failed, matching existing defenses with 700x fewer GPU hours. Best-of-N jailbreaking achieves 89% attack success on GPT-4o and 78% on Claude 3.5 Sonnet through random input augmentations. Attack success follows power law scaling across text, vision, and audio, enabling quantitative forecasting of adversarial robustness. Agentic misalignment tests whether frontier models autonomously choose harmful actions given ordinary goals. Across 16 models, agents engaged in blackmail (96% for Claude Opus 4), espionage, and actions causing death. Misbehavior rates rose from 6.5% to 55.1% when models stated scenarios were real rather than evaluations. The thesis does not fully resolve any of these problems but makes each tractable and measurable.
Authors: Elaheh Sanoubari
Abstract: We used the Webots robotics simulation platform to simulate a dyadic avoiding and mobbing predator behavior in a group of Braitenbergian robots. Mobbing is an antipredator adaptation used by some animals in which the individuals cooperatively attack or harass a predator to protect themselves. One way of coordinating a mobbing attack is using mobbing calls to summon other individuals of the mobbing species. We imitated this mechanism and simulated Braitenbergian robots that use mobbing calls when they face a light source (representing an inanimate predator) and mob it if they can summon allies, otherwise, they escape from it. We explore the effects of range of mobbing call (infinite range, mid-range and low-range) and the size of the robot group (ten robots vs three) on the overall success of mobbing. Our results suggest that both variables have significant impacts. This work has implications for simulations of action selection in artificial life and designing control architectures for autonomous agents.
Authors: KrishnaSaiReddy Patil
Abstract: RAG systems deployed across federal agencies for citizen-facing services are vulnerable to knowledge base poisoning attacks, where adversaries inject malicious documents to manipulate outputs. Recent work demonstrates that as few as 10 adversarial passages can achieve 98.2% retrieval success rates. We observe that RAG knowledge base poisoning is structurally analogous to software supply chain attacks, and propose RAGShield, a five-layer defense-in-depth framework applying supply chain provenance verification to the RAG knowledge pipeline. RAGShield introduces: (1) C2PA-inspired cryptographic document attestation blocking unsigned and forged documents at ingestion; (2) trust-weighted retrieval prioritizing provenance-verified sources; (3) a formal taint lattice with cross-source contradiction detection catching insider threats even when provenance is valid; (4) provenance-aware generation with auditable citations; and (5) NIST SP 800-53 compliance mapping across 15 control families. Evaluation on a 500-passage Natural Questions corpus with 63 attack documents and 200 queries against five adversary tiers achieves 0.0% attack success rate including adaptive attacks (95% CI: [0.0%, 1.9%]) with 0.0% false positive rate. We honestly report that insider in-place replacement attacks achieve 17.5% ASR, identifying the fundamental limit of ingestion-time defense. The cross-source contradiction detector catches subtle numerical manipulation attacks that bypass provenance verification entirely.
Authors: Alibek T. Kaliyev, Artem Maryanskyy
Abstract: Modern LLM agents increasingly create their own tools at runtime -- from Python functions to API clients -- yet existing benchmarks evaluate them almost exclusively by downstream task completion. This is analogous to judging a software engineer only by whether their code runs, ignoring redundancy, regression, and safety. We introduce EvolveTool-Bench, a diagnostic benchmark for LLM-generated tool libraries in software engineering workflows. Across three domains requiring actual tool execution (proprietary data formats, API orchestration, and numerical computation), we define library-level software quality metrics -- reuse, redundancy, composition success, regression stability, and safety -- alongside a per-tool Tool Quality Score measuring correctness, robustness, generality, and code quality. In the first head-to-head comparison of code-level and strategy-level tool evolution (ARISE vs. EvoSkill vs. one-shot baselines, 99 tasks, two models), we show that systems with similar task completion (63-68%) differ by up to 18% in library health, revealing software quality risks invisible to task-only evaluation. Our results highlight that evaluation and governance of LLM-generated tools require treating the evolving tool library as a first-class software artifact, not a black box.
Authors: Weyl Lu, Chenjie Hao, Yubei Chen
Abstract: Estimated density is often interpreted as indicating how typical a sample is under a model. Yet deep models trained on one dataset can assign higher density to simpler out-of-distribution (OOD) data than to in-distribution test data. We refer to this behavior as the OOD anomaly. Prior work typically studies this phenomenon within a single architecture, detector, or benchmark, implicitly assuming certain canonical densities. We instead separate the trained network from the density estimator built from its representations or outputs. We introduce two estimators: Jacobian-based estimators and autoregressive self-estimators, making density analysis applicable to a wide range of models. Applying this perspective to a range of models, including iGPT, PixelCNN++, Glow, score-based diffusion models, DINOv2, and I-JEPA, we find the same striking regularity that goes beyond the OOD anomaly: lower-complexity samples receive higher estimated density, while higher-complexity samples receive lower estimated density. This ordering appears within a test set and across OOD pairs such as CIFAR-10 and SVHN, and remains highly consistent across independently trained models. To quantify these orderings, we introduce Spearman rank correlation and find striking agreement both across models and with external complexity metrics. Even when trained only on the lowest-density (most complex) samples - or even a single such sample - the resulting models still rank simpler images as higher density. These observations lead us beyond the original OOD anomaly to a more general conclusion: deep networks consistently favor simple data. Our goal is not to close this question, but to define and visualize it more clearly. We broaden its empirical scope and show that it appears across architectures, objectives, and density estimators.
Authors: Yuchen Yang, Shuangyang Zhong, Haijun Yu, Langcuomu Suo, Hongbin Han, Florian Putz, Yixing Huang
Abstract: Background: Deep learning has demonstrated significant potential for automated brain metastases (BM) segmentation; however, models trained at a singular institution often exhibit suboptimal performance at various sites due to disparities in scanner hardware, imaging protocols, and patient demographics. The goal of this work is to create a domain adaptation framework that will allow for BM segmentation to be used across multiple institutions. Methods: We propose a VAE-MMD preprocessing pipeline that combines variational autoencoders (VAE) with maximum mean discrepancy (MMD) loss, incorporating skip connections and self-attention mechanisms alongside nnU-Net segmentation. The method was tested on 740 patients from four public databases: Stanford, UCSF, UCLM, and PKG, evaluated by domain classifier's accuracy, sensitivity, precision, F1/F2 scores, surface Dice (sDice), and 95th percentile Hausdorff distance (HD95). Results: VAE-MMD reduced domain classifier accuracy from 0.91 to 0.50, indicating successful feature alignment across institutions. Reconstructed volumes attained a PSNR greater than 36 dB, maintaining anatomical accuracy. The combined method raised the mean F1 by 11.1% (0.700 to 0.778), the mean sDice by 7.93% (0.7121 to 0.7686), and reduced the mean HD95 by 65.5% (11.33 to 3.91 mm) across all four centers compared to the baseline nnU-Net. Conclusions: VAE-MMD effectively diminishes cross-institutional data heterogeneity and enhances BM segmentation generalization across volumetric, detection, and boundary-level metrics without necessitating target-domain labels, thereby overcoming a significant obstacle to the clinical implementation of AI-assisted segmentation.
Authors: Seohyoung Park, Jaeyeol Lim, Seoyoung Ju, Kyeonghun Kim, Nam-Joon Kim, Hyuk-Jae Lee
Abstract: Developing robust models to accurately predict the trajectories of surrounding agents is fundamental to autonomous driving safety. However, most public datasets, such as the Waymo Open Motion Dataset and Argoverse, are collected in Western road environments and do not reflect the unique traffic patterns, infrastructure, and driving behaviors of other regions, including South Korea. This domain discrepancy leads to performance degradation when state-of-the-art models trained on Western data are deployed in different geographic contexts. In this work, we investigate the adaptability of Query-Centric Trajectory Prediction (QCNet) when transferred from U.S.-based data to Korean road environments. Using a Korean autonomous driving dataset, we compare four training strategies: zero-shot transfer, training from scratch, full fine-tuning, and encoder freezing. Experimental results demonstrate that leveraging pretrained knowledge significantly improves prediction performance. Specifically, selectively fine-tuning the decoder while freezing the encoder yields the best trade-off between accuracy and training efficiency, reducing prediction error by over 66% compared to training from scratch. This study provides practical insights into effective transfer learning strategies for deploying trajectory prediction models in new geographic domains.
Authors: Weizhuo Wang, Yanjie Ze, C. Karen Liu, Monroe Kennedy III
Abstract: We present EgoNav, a system that enables a humanoid robot to traverse diverse, unseen environments by learning entirely from 5 hours of human walking data, with no robot data or finetuning. A diffusion model predicts distributions of plausible future trajectories conditioned on past trajectory, a 360 deg visual memory fusing color, depth, and semantics, and video features from a frozen DINOv3 backbone that capture appearance cues invisible to depth sensors. A hybrid sampling scheme achieves real-time inference in 10 denoising steps, and a receding-horizon controller selects paths from the predicted distribution. We validate EgoNav through offline evaluations, where it outperforms baselines in collision avoidance and multi-modal coverage, and through zero-shot deployment on a Unitree G1 humanoid across unseen indoor and outdoor environments. Behaviors such as waiting for doors to open, navigating around crowds, and avoiding glass walls emerge naturally from the learned prior. We will release the dataset and trained models. Our website: https://egonav.weizhuowang.com
Authors: Ravi Ranjan, Utkarsh Grover, Xiaomin Lin, Agoritsa Polyzou
Abstract: Large language models (LLMs) are trained on massive web-scale corpora, raising growing concerns about privacy and copyright. Membership inference attacks (MIAs) aim to determine whether a given example was used during training. Existing LLM MIAs largely rely on output probabilities or loss values and often perform only marginally better than random guessing when members and non-members are drawn from the same distribution. We introduce G-Drift MIA, a white-box membership inference method based on gradient-induced feature drift. Given a candidate (x,y), we apply a single targeted gradient-ascent step that increases its loss and measure the resulting changes in internal representations, including logits, hidden-layer activations, and projections onto fixed feature directions, before and after the update. These drift signals are used to train a lightweight logistic classifier that effectively separates members from non-members. Across multiple transformer-based LLMs and datasets derived from realistic MIA benchmarks, G-Drift substantially outperforms confidence-based, perplexity-based, and reference-based attacks. We further show that memorized training samples systematically exhibit smaller and more structured feature drift than non-members, providing a mechanistic link between gradient geometry, representation stability, and memorization. In general, our results demonstrate that small, controlled gradient interventions offer a practical tool for auditing the membership of training-data and assessing privacy risks in LLMs.
Authors: Iyad Ait Hou, Rebecca Hwa
Abstract: If the same neuron activates for both "lender" and "riverside," standard metrics attribute the overlap to superposition--the neuron must be compressing two unrelated concepts. This work explores how much of the overlap is due a lexical confound: neurons fire for a shared word form (such as "bank") rather than for two compressed concepts. A 2x2 factorial decomposition reveals that the lexical-only condition (same word, different meaning) consistently exceeds the semantic-only condition (different word, same meaning) across models spanning 110M-70B parameters. The confound carries into sparse autoencoders (18-36% of features blend senses), sits in <=1% of activation dimensions, and hurts downstream tasks: filtering it out improves word sense disambiguation and makes knowledge edits more selective (p = 0.002).
Authors: Jiwoo Ha, Jongwoo Baek, Jinhyun So
Abstract: Recent Large Vision-Language Models (LVLMs) have demonstrated remarkable performance across various multimodal tasks that require understanding both visual and linguistic inputs. However, object hallucination -- the generation of nonexistent objects in answers -- remains a persistent challenge. Although several approaches such as retraining and external grounding methods have been proposed to mitigate this issue, they still suffer from high data costs or structural complexity. Training-free methods such as Contrastive Decoding (CD) are more cost-effective, avoiding additional training or external models, but still suffer from long-term decay, where visual grounding weakens and language priors dominate as the generation progresses. In this paper, we propose First Logit Boosting (FLB), a simple yet effective training-free technique designed to alleviate long-term decay in LVLMs. FLB stores the logit of the first generated token and adds it to subsequent token predictions, effectively mitigating long-term decay of visual information. We observe that FLB (1) sustains the visual information embedded in the first token throughout generation, and (2) suppresses hallucinated words through the stabilizing effect of the ``The'' token. Experimental results show that FLB significantly reduces object hallucination across various tasks, benchmarks, and backbone models. Notably, it causes negligible inference overhead, making it highly applicable to real-time multimodal systems. Code is available at https://github.com/jiwooha20/FLB
Authors: Veda Duddu, Jash Rajesh Parekh, Andy Mao, Hanyi Min, Ziang Xiao, Vedant Das Swain, Koustuv Saha
Abstract: AI-driven conversational coaching is increasingly used to support workplace negotiation, yet prior work assumes uniform effectiveness across users. We challenge this assumption by examining how individual differences, particularly personality traits, moderate coaching outcomes. We conducted a between-subjects experiment (N=267) comparing theory-driven AI (Trucey), general-purpose AI (Control-AI), and a traditional negotiation handbook (Control-NoAI). Participants were clustered into three profiles -- resilient, overcontrolled, and undercontrolled -- based on the Big-Five personality traits and ARC typology. Resilient workers achieved broad psychological gains primarily from the handbook, overcontrolled workers showed outcome-specific improvements with theory-driven AI, and undercontrolled workers exhibited minimal effects despite engaging with the frameworks. These patterns suggest personality as a predictor of readiness beyond stage-based tailoring: vulnerable users benefit from targeted rather than comprehensive interventions. The study advances understanding of personality-determined intervention prerequisites and highlights design implications for adaptive AI coaching systems that align support intensity with individual readiness, rather than assuming universal effectiveness.
Authors: Zhensu Sun, Zhihao Lin, Zhi Chen, Chengran Yang, Mingyi Zhou, Li Li, David Lo
Abstract: Current LLM-based coding agents follow a serial execution paradigm: the model first generates the complete code, then invokes an interpreter to execute it. This sequential workflow leaves the executor idle during generation and the generator idle during execution, resulting in unnecessary end-to-end latency. We observe that, unlike human developers, LLMs produce code tokens sequentially without revision, making it possible to execute code as it is being generated. We formalize this parallel execution paradigm, modeling it as a three-stage pipeline of generation, detection, and execution, and derive closed-form latency bounds that characterize its speedup potential and operating regimes. We then present Eager, a concrete implementation featuring AST-based chunking, dynamic batching with gated execution, and early error interruption. We evaluate Eager across four benchmarks, seven LLMs, and three execution environments. Results show that Eager reduces the non-overlapped execution latency by up to 99.9% and the end-to-end latency by up to 55% across seven LLMs and four benchmarks.
Authors: Yabin Zhang, Chong Wang, Yunhe Gao, Jiaming Liu, Maya Varma, Justin Xu, Sophie Ostmeier, Jin Long, Sergios Gatidis, Seena Dehkharghani, Arne Michalson, Eun Kyoung Hong, Christian Bluethgen, Haiwei Henry Guo, Alexander Victor Ortiz, Stephan Altmayer, Sandhya Bodapati, Joseph David Janizek, Ken Chang, Jean-Benoit Delbrouck, Akshay S. Chaudhari, Curtis P. Langlotz
Abstract: Chest X-rays (CXRs) are among the most frequently performed imaging examinations worldwide, yet rising imaging volumes increase radiologist workload and the risk of diagnostic errors. Although artificial intelligence (AI) systems have shown promise for CXR interpretation, most generate only final predictions, without making explicit how visual evidence is translated into radiographic findings and diagnostic predictions. We present CheXOne, a reasoning-enabled vision-language model for CXR interpretation. CheXOne jointly generates diagnostic predictions and explicit, clinically grounded reasoning traces that connect visual evidence, radiographic findings, and these predictions. The model is trained on 14.7 million instruction and reasoning samples curated from 30 public datasets spanning 36 CXR interpretation tasks, using a two-stage framework that combines instruction tuning with reinforcement learning to improve reasoning quality. We evaluate CheXOne in zero-shot settings across visual question answering, report generation, visual grounding and reasoning assessment, covering 17 evaluation settings. CheXOne outperforms existing medical and general-domain foundation models and achieves strong performance on independent public benchmarks. A clinical reader study demonstrates that CheXOne-drafted reports are comparable to or better than resident-written reports in 55% of cases, while effectively addressing clinical indications and enhancing both report writing and CXR interpretation efficiency. Further analyses involving radiologists reveal that the generated reasoning traces show high clinical factuality and provide causal support for the final predictions, offering a plausible explanation for the performance gains. These results suggest that explicit reasoning can improve model performance, interpretability and clinical utility in AI-assisted CXR interpretation.
Authors: Yunwen Lei, Yufeng Xie
Abstract: Overparameterized neural networks often show a benign overfitting property in the sense of achieving excellent generalization behavior despite the number of parameters exceeding the number of training examples. A promising direction to explain benign overfitting is to relate generalization to the norm of distance from initialization, motivated by the empirical observations that this distance is often significantly smaller than the norm itself. However, the existing initialization-dependent complexity analyses cannot fully exploit the power of initialization since the associated bounds depend on the spectral norm of the initialization matrix, which can scale as a square-root function of the width and are therefore not effective for overparameterized models. In this paper, we develop the first \emph{fully} initialization-dependent complexity bounds for shallow neural networks with general Lipschitz activation functions, which enjoys a logarithmic dependency on the width. Our bounds depend on the path-norm of the distance from initialization, which are derived by introducing a new peeling technique to handle the challenge along with the initialization-dependent constraint. We also develop a lower bound tight up to a constant factor. Finally, we conduct empirical comparisons and show that our generalization analysis implies non-vacuous bounds for overparameterized networks.
Authors: Junxian Wu, Chenghan Fu, Zhanheng Nie, Daoze Zhang, Bowen Wan, Wanxian Guan, Chuan Yu, Jian Xu, Bo Zheng
Abstract: With the rapid growth of e-commerce, exploring general representations rather than task-specific ones has attracted increasing attention. Although recent multimodal large language models (MLLMs) have driven significant progress in product understanding, they are typically employed as feature extractors that implicitly encode product information into global embeddings, thereby limiting their ability to capture fine-grained attributes. Therefore, we argue that leveraging the reasoning capabilities of MLLMs to explicitly model fine-grained product attributes holds significant potential. Nevertheless, achieving this goal remains non-trivial due to several key challenges: (i) long-context reasoning tends to dilute the model's attention to salient information in the raw input; (ii) supervised fine-tuning (SFT) primarily encourages rigid imitation, limiting the exploration of effective reasoning strategies; and (iii) fine-grained details are progressively attenuated during forward propagation. To address these issues, we propose MOON3.0, the first reasoning-aware MLLM-based model for product representation learning. Our method (1) employs a multi-head modality fusion module to adaptively integrate raw signals; (2) incorporates a joint contrastive and reinforcement learning framework to autonomously explore more effective reasoning strategies; and (3) introduces a fine-grained residual enhancement module to progressively preserve local details throughout the network. Additionally, we release a large-scale multimodal e-commerce benchmark MBE3.0. Experimentally, our model demonstrates state-of-the-art zero-shot performance across various downstream tasks on both our benchmark and public datasets.
Authors: Kyeonghun Kim, Hyeonseok Jung, Youngung Han, Junsu Lim, YeonJu Jean, Seongbin Park, Eunseob Choi, Hyunsu Go, SeoYoung Ju, Seohyoung Park, Gyeongmin Kim, MinJu Kwon, KyungSeok Yuh, Soo Yong Kim, Ken Ying-Kai Liao, Nam-Joon Kim, Hyuk-Jae Lee
Abstract: Training deep learning models for three-dimensional (3D) medical imaging, such as Computed Tomography (CT), is fundamentally challenged by the scarcity of labeled data. While pre-training on natural images is common, it results in a significant domain shift, limiting performance. Self-Supervised Learning (SSL) on unlabeled medical data has emerged as a powerful solution, but prominent frameworks often fail to exploit the inherent 3D nature of CT scans. These methods typically process 3D scans as a collection of independent 2D slices, an approach that fundamentally discards critical axial coherence and the 3D structural context. To address this limitation, we propose the autoencoder for enhanced self-supervised medical image learning(MAESIL), a novel self-supervised learning framework designed to capture 3D structural information efficiently. The core innovation is the 'superpatch', a 3D chunk-based input unit that balances 3D context preservation with computational efficiency. Our framework partitions the volume into superpatches and employs a 3D masked autoencoder strategy with a dual-masking strategy to learn comprehensive spatial representations. We validated our approach on three diverse large-scale public CT datasets. Our experimental results show that MAESIL demonstrates significant improvements over existing methods such as AE, VAE and VQ-VAE in key reconstruction metrics such as PSNR and SSIM. This establishes MAESIL as a robust and practical pre-training solution for 3D medical imaging tasks.
Authors: Axiu Mao, Meilu Zhu, Lei Shen, Xiaoshuai Wang, Tomas Norton, Kai Liu
Abstract: With the rapid advancements in deep learning techniques, wearable sensor-aided animal activity recognition (AAR) has demonstrated promising performance, thereby improving livestock management efficiency as well as animal health and welfare monitoring. However, existing research often prioritizes overall performance, overlooking the fact that classification accuracies for specific animal behavioral categories may remain unsatisfactory. This issue typically stems from suboptimal sampling rates or class imbalance problems. To address these challenges and achieve high classification accuracy across all individual behaviors in farm animals, we propose a novel Individual-Behavior-Aware Network (IBA-Net). This network enhances the recognition of each specific behavior by simultaneously customizing features and calibrating the classifier. Specifically, considering that different behaviors require varying sampling rates to achieve optimal performance, we design a Mixture-of-Experts (MoE)-based Feature Customization (MFC) module. This module adaptively fuses data from multiple sampling rates, capturing customized features tailored to various animal behaviors. Additionally, to mitigate classifier bias toward majority classes caused by class imbalance, we develop a Neural Collapse-driven Classifier Calibration (NC3) module. This module introduces a fixed equiangular tight frame (ETF) classifier during the classification stage, maximizing the angles between pair-wise classifier vectors and thereby improving the classification performance for minority classes. To validate the effectiveness of IBA-Net, we conducted experiments on three public datasets covering goat, cattle, and horse activity recognition. The results demonstrate that our method consistently outperforms existing approaches across all datasets.
Authors: Haibo Wang, Zihao Lin, Zhiyang Xu, Lifu Huang
Abstract: 3D Visual Grounding (3D-VG) aims to localize objects in 3D scenes via natural language descriptions. While recent advancements leveraging Vision-Language Models (VLMs) have explored zero-shot possibilities, they typically suffer from a static workflow relying on preprocessed 3D point clouds, essentially degrading grounding into proposal matching. To bypass this reliance, our core motivation is to decouple the task: leveraging 2D VLMs to resolve complex spatial semantics, while relying on deterministic multi-view geometry to instantiate the 3D structure. Driven by this insight, we propose "Think, Act, Build (TAB)", a dynamic agentic framework that reformulates 3D-VG tasks as a generative 2D-to-3D reconstruction paradigm operating directly on raw RGB-D streams. Specifically, guided by a specialized 3D-VG skill, our VLM agent dynamically invokes visual tools to track and reconstruct the target across 2D frames. Crucially, to overcome the multi-view coverage deficit caused by strict VLM semantic tracking, we introduce the Semantic-Anchored Geometric Expansion, a mechanism that first anchors the target in a reference video clip and then leverages multi-view geometry to propagate its spatial location across unobserved frames. This enables the agent to "Build" the target's 3D representation by aggregating these multi-view features via camera parameters, directly mapping 2D visual cues to 3D coordinates. Furthermore, to ensure rigorous assessment, we identify flaws such as reference ambiguity and category errors in existing benchmarks and manually refine the incorrect queries. Extensive experiments on ScanRefer and Nr3D demonstrate that our framework, relying entirely on open-source models, significantly outperforms previous zero-shot methods and even surpasses fully supervised baselines.
Authors: Zhiting Fan, Ruizhe Chen, Tianxiang Hu, Ru Peng, Zenan Huang, Haokai Xu, Yixin Chen, Jian Wu, Junbo Zhao, Zuozhu Liu
Abstract: Large language models (LLMs) achieve strong downstream performance largely due to abundant supervised fine-tuning (SFT) data. However, high-quality SFT data in knowledge-intensive domains such as humanities, social sciences, medicine, law, and finance is scarce because expert curation is expensive, privacy constraints are strict, and label consistency is hard to ensure. Recent work uses synthetic data, typically by prompting a generator over domain documents and filtering outputs with handcrafted rubrics. Yet rubric design is expert-dependent, transfers poorly across domains, and is often optimized through a brittle heuristic loop of writing rubrics, synthesizing data, training, inspecting results, and manually guessing revisions. This process lacks reliable quantitative feedback about how a rubric affects downstream performance. We propose evaluating synthetic data by its training utility on the target model and using this signal to guide data generation. Inspired by influence estimation, we adopt an optimizer-aware estimator that uses gradient information to quantify each synthetic sample's contribution to a target model's objective on specific tasks. Our analysis shows that even when synthetic and real samples are close in embedding space, their influence on learning can differ substantially. Based on this insight, we propose an optimization-based framework that adapts rubrics using target-model feedback. We provide lightweight guiding text and use a rubric-specialized model to generate task-conditioned rubrics. Influence score is used as the reward to optimize the rubric generator with reinforcement learning. Experiments across domains, target models, and data generators show consistent improvements and strong generalization without task-specific tuning.
Authors: Kyeonghun Kim, Jaehyung Park, Youngung Han, Anna Jung, Seongbin Park, Sumin Lee, Jiwon Yang, Jiyoon Han, Subeen Lee, Junsu Lim, Hyunsu Go, Eunseob Choi, Hyeonseok Jung, Soo Yong Kim, Woo Kyoung Jeong, Won Jae Lee, Pa Hong, Hyuk-Jae Lee, Ken Ying-Kai Liao, Nam-Joon Kim
Abstract: Dental diagnosis from Orthopantomograms (OPGs) requires coordination of tooth detection, caries segmentation (CarSeg), anomaly detection (AD), and dental developmental staging (DDS). We propose Mamba-based Architectural Tooth Hierarchical Estimator and Holistic Evaluation Network for Anatomy (MATHENA), a unified framework leveraging Mamba's linear-complexity State Space Models (SSM) to address all four tasks. MATHENA integrates MATHE, a multi-resolution SSM-driven detector with four-directional Vision State Space (VSS) blocks for O(N) global context modeling, generating per-tooth crops. These crops are processed by HENA, a lightweight Mamba-UNet with a triple-head architecture and Global Context State Token (GCST). In the triple-head architecture, CarSeg is first trained as an upstream task to establish shared representations, which are then frozen and reused for downstream AD fine-tuning and DDS classification via linear probing, enabling stable, efficient learning. We also curate PARTHENON, a benchmark comprising 15,062 annotated instances from ten datasets. MATHENA achieves 93.78% mAP@50 in tooth detection, 90.11% Dice for CarSeg, 88.35% for AD, and 72.40% ACC for DDS.
Authors: Hongyang Yang, Yanxin Zhang, Yang She, Yue Xiao, Hao Wu, Yiyang Zhang, Jiapeng Hou, Rongshan Zhang
Abstract: Housing selection is a high-stakes and largely irreversible decision problem. We study housing consultation as a decision-support interface for housing selection. Existing housing platforms and many LLM-based assistants often reduce this process to ranking or recommendation, resulting in opaque reasoning, brittle multi-constraint handling, and limited guarantees on factuality. We present HabitatAgent, the first LLM-powered multi-agent architecture for end-to-end housing consultation. HabitatAgent comprises four specialized agent roles: Memory, Retrieval, Generation, and Validation. The Memory Agent maintains multi-layer user memory through internal stages for constraint extraction, memory fusion, and verification-gated updates; the Retrieval Agent performs hybrid vector--graph retrieval (GraphRAG); the Generation Agent produces evidence-referenced recommendations and explanations; and the Validation Agent applies multi-tier verification and targeted remediation. Together, these agents provide an auditable and reliable workflow for end-to-end housing consultation. We evaluate HabitatAgent on 100 real user consultation scenarios (300 multi-turn question--answer pairs) under an end-to-end correctness protocol. A strong single-stage baseline (Dense+Rerank) achieves 75% accuracy, while HabitatAgent reaches 95%.
Authors: Mingming Ha, Guanchen Wang, Linxun Chen, Xuan Rao, Yuexin Shi, Tianbao Ma, Zhaojie Liu, Yunqian Fan, Zilong Lu, Yanan Niu, Han Li, Kun Gai
Abstract: In recent years, the scaling laws of recommendation models have attracted increasing attention, which govern the relationship between performance and parameters/FLOPs of recommenders. Currently, there are three mainstream architectures for achieving scaling in recommendation models, namely attention-based, TokenMixer-based, and factorization-machine-based methods, which exhibit fundamental differences in both design philosophy and architectural structure. In this paper, we propose a unified scaling architecture for recommendation systems, namely \textbf{UniMixer}, to improve scaling efficiency and establish a unified theoretical framework that unifies the mainstream scaling blocks. By transforming the rule-based TokenMixer to an equivalent parameterized structure, we construct a generalized parameterized feature mixing module that allows the token mixing patterns to be optimized and learned during model training. Meanwhile, the generalized parameterized token mixing removes the constraint in TokenMixer that requires the number of heads to be equal to the number of tokens. Furthermore, we establish a unified scaling module design framework for recommender systems, which bridges the connections among attention-based, TokenMixer-based, and factorization-machine-based methods. To further boost scaling ROI, a lightweight UniMixing module is designed, \textbf{UniMixing-Lite}, which further compresses the model parameters and computational cost while significantly improve the model performance. The scaling curves are shown in the following figure. Extensive offline and online experiments are conducted to verify the superior scaling abilities of \textbf{UniMixer}.
Authors: Pawe{\l} Liskowski, Kyle Schmaus
Abstract: Modern data warehouses extend SQL with semantic operators that invoke large language models on each qualifying row, but the per-row inference cost is prohibitive at scale. Model cascades reduce this cost by routing most rows through a fast proxy model and delegating uncertain cases to an expensive oracle. Existing frameworks, however, require global dataset access and optimize a single quality metric, limiting their applicability in distributed systems where data is partitioned across independent workers. We present two adaptive cascade algorithms designed for streaming, per-partition execution in which each worker processes its partition independently without inter-worker communication. SUPG-IT extends the SUPG statistical framework to streaming execution with iterative threshold refinement and joint precision-recall guarantees. GAMCAL replaces user-specified quality targets with a learned calibration model: a Generalized Additive Model maps proxy scores to calibrated probabilities with uncertainty quantification, enabling direct optimization of a cost-quality tradeoff through a single parameter. Experiments on six datasets in a production semantic SQL engine show that both algorithms achieve F1 > 0.95 on every dataset. GAMCAL achieves higher F1 per oracle call at cost-sensitive operating points, while SUPG-IT reaches a higher quality ceiling with formal guarantees on precision and recall.
Authors: Kinson Vernet
Abstract: Mechanistic simulations typically assume fixed ontologies: variables, causal relationships, and resolution policies are static. This assumption fails when the true causal structure is contested or unidentifiable-as in antimicrobial resistance (AMR) spread, where contact, environmental, and selection ontologies compete. We introduce Procela, a Python framework where variables act as epistemic authorities that maintain complete hypothesis memory, mechanisms encode competing ontologies as causal units, and governance observes epistemic signals and mutates system topology at runtime. This is the first framework where simulations test their own assumptions. We instantiate Procela for AMR in a hospital network with three competing families. Governance detects coverage decay, policy fragility, and runs structural probes. Results show 20.4% error reduction and 69% cumulative regret improvement over baseline. All experiments are reproducible with full auditability. Procela establishes a new paradigm: simulations that model not only the world but their own modeling process, enabling adaptation under structural uncertainty.
Authors: Lewis Tham, Nicholas Mac Gregor Garcia, Jungpil Hahn
Abstract: Autonomous agents increasingly interact with the web, yet most websites remain designed for human browsers -- a fundamental mismatch that the emerging ``Agentic Web'' must resolve. Agents must repeatedly browse pages, inspect DOMs, and reverse-engineer callable routes -- a process that is slow, brittle, and redundantly repeated across agents. We observe that every modern website already exposes internal APIs (sometimes called \emph{shadow APIs}) behind its user interface -- first-party endpoints that power the site's own functionality. We present Unbrowse, a shared route graph that transforms browser-based route discovery into a collectively maintained index of these callable first-party interfaces. The system passively learns routes from real browsing traffic and serves cached routes via direct API calls. In a single-host live-web benchmark of equivalent information-retrieval tasks across 94 domains, fully warmed cached execution averaged 950\,ms versus 3{,}404\,ms for Playwright browser automation (3.6$\times$ mean speedup, 5.4$\times$ median), with well-cached routes completing in under 100\,ms. A three-path execution model -- local cache, shared graph, or browser fallback -- ensures the system is voluntary and self-correcting. A three-tier micropayment model via the x402 protocol charges per-query search fees for graph lookups (Tier~3), a one-time install fee for discovery documentation (Tier~1), and optional per-execution fees for site owners who opt in (Tier~2). All tiers are grounded in a necessary condition for rational adoption: an agent uses the shared graph only when the total fee is lower than the expected cost of browser rediscovery.
Authors: Yu Xia, Canwen Xu, Zhewei Yao, Julian McAuley, Yuxiong He
Abstract: Group Relative Policy Optimization (GRPO) is widely used for reinforcement learning with verifiable rewards, but it often suffers from advantage collapse: when all rollouts in a group receive the same reward, the group yields zero relative advantage and thus no learning signal. For example, if a question is too hard for the reasoner, all sampled rollouts can be incorrect and receive zero reward. Recent work addresses this issue by adding hints or auxiliary scaffolds to such hard questions so that the reasoner produces mixed outcomes and recovers a non-zero update. However, existing hints are usually fixed rather than adapted to the current reasoner, and a hint that creates learning signal under the hinted input does not necessarily improve the no-hint policy used at test time. To this end, we propose Hint Learning for Reinforcement Learning (HiLL), a framework that jointly trains a hinter policy and a reasoner policy during RL. For each hard question, the hinter generates hints online conditioned on the current reasoner's incorrect rollout, allowing hint generation to adapt to the reasoner's evolving errors. We further introduce hint reliance, which measures how strongly correct hinted trajectories depend on the hint. We derive a transferability result showing that lower hint reliance implies stronger transfer from hinted success to no-hint success, and we use this result to define a transfer-weighted reward for training the hinter. Therefore, HiLL favors hints that not only recover informative GRPO groups, but also produce signals that are more likely to improve the original no-hint policy. Experiments across multiple benchmarks show that HiLL consistently outperforms GRPO and prior hint-based baselines, demonstrating the value of adaptive and transfer-aware hint learning for RL. The code is available at https://github.com/Andree-9/HiLL.
Authors: Ruozhao Yang, Mingfei Cheng, Gelei Deng, Junjie Wang, Tianwei Zhang, Xiaofei Xie
Abstract: Large-scale web applications are widely deployed with complex third-party components, inheriting security risks arising from component vulnerabilities. Security assessment is therefore required to determine whether such known vulnerabilities remain practically exploitable in real applications. Penetration testing is a widely adopted approach that validates exploitability by launching concrete attacks against known vulnerabilities in real-world black-box systems. However, existing approaches often fail to automatically generate reliable exploits, limiting their effectiveness in practical security assessment. This limitation mainly stems from two issues: (1) precisely triggering vulnerabilities with correct technical details, and (2) adapting exploits to diverse real-world deployment settings. In this paper, we propose AutoEG, a fully automated multi-agent framework for exploit generation targeting black-box web applications. AutoEG has two phases: First, AutoEG extracts precise vulnerability trigger logic from unstructured vulnerability information and encapsulates it into reusable trigger functions. Second, AutoEG uses trigger functions for concrete attack objectives and iteratively refines exploits through feedback-driven interaction with the target application. We evaluate AutoEG on 104 real-world vulnerabilities with 29 attack objectives, resulting in 660 exploitation tasks and 55,440 exploit attempts. AutoEG achieves an average success rate of 82.41%, substantially outperforming state-of-the-art baselines, whose best performance reaches only 32.88%.
Authors: Karan Singh, Michael Yu, Varun Gangal, Zhuofu Tao, Sachin Kumar, Emmy Liu, Steven Y. Feng
Abstract: Retrieval-augmented generation (RAG) improves language model (LM) performance by providing relevant context at test time for knowledge-intensive situations. However, the relationship between parametric knowledge acquired during pretraining and non-parametric knowledge accessed via retrieval remains poorly understood, especially under fixed data budgets. In this work, we systematically study the trade-off between pretraining corpus size and retrieval store size across a wide range of model and data scales. We train OLMo-2-based LMs ranging from 30M to 3B parameters on up to 100B tokens of DCLM data, while varying both pretraining data scale (1-150x the number of parameters) and retrieval store size (1-20x), and evaluate performance across a diverse suite of benchmarks spanning reasoning, scientific QA, and open-domain QA. We find that retrieval consistently improves performance over parametric-only baselines across model scales and introduce a three-dimensional scaling framework that models performance as a function of model size, pretraining tokens, and retrieval corpus size. This scaling manifold enables us to estimate optimal allocations of a fixed data budget between pretraining and retrieval, revealing that the marginal utility of retrieval depends strongly on model scale, task type, and the degree of pretraining saturation. Our results provide a quantitative foundation for understanding when and how retrieval should complement pretraining, offering practical guidance for allocating data resources in the design of scalable language modeling systems.
Authors: Sihan Zhou, Tiantian He, Yifan Lu, Yaqing Hou, Yew-Soon Ong
Abstract: Non-stationarity arises from concurrent policy updates and leads to persistent environmental fluctuations. Existing approaches like Centralized Training with Decentralized Execution (CTDE) and sequential update schemes mitigate this issue. However, since the perception of the policies of other agents remains dependent on sampling environmental interaction data, the agent essentially operates in a passive perception state. This inevitably triggers equilibrium oscillations and significantly slows the convergence speed of the system. To address this issue, we propose Gradient Realignment via Active Shared Perception (GRASP), a novel framework that defines generalized Bellman equilibrium as a stable objective for policy evolution. The core mechanism of GRASP involves utilizing the independent gradients of agents to derive a defined consensus gradient, enabling agents to actively perceive policy updates and optimize team collaboration. Theoretically, we leverage the Kakutani Fixed-Point Theorem to prove that the consensus direction $u^*$ guarantees the existence and attainability of this equilibrium. Extensive experiments on StarCraft II Multi-Agent Challenge (SMAC) and Google Research Football (GRF) demonstrate the scalability and promising performance of the framework.
Authors: Ricardo Hidalgo-Arag\'on, Jes\'us M. Gonz\'alez-Barahona, Gregorio Robles
Abstract: Context: Schools, training platforms, and technology firms increasingly need to assess programming proficiency at scale with transparent, reproducible methods that support personalized learning pathways. Objective: This study introduces a pedagogical framework for Scratch project assessment, aligned with the Common European Framework of Reference (CEFR), providing universal competency levels for students and teachers alongside actionable insights for curriculum design. Method: We apply Fuzzy C-Means clustering to 2008246 Scratch projects evaluated via Dr.Scratch, implementing an ordinal criterion to map clusters to CEFR levels (A1-C2), and introducing enhanced classification metrics that identify transitional learners, enable continuous progress tracking, and quantify classification certainty to balance automated feedback with instructor review. Impact: The framework enables diagnosis of systemic curriculum gaps-notably a "B2 bottleneck" where only 13.3% of learners reside due to the cognitive load of integrating Logic Synchronization, and Data Representation--while providing certainty--based triggers for human intervention.
Authors: Bj\"orn Roman Kohlberger (EctoSpace, Dublin, Ireland)
Abstract: The memory wall remains the primary bottleneck for training large language models on consumer hardware. We introduce Spectral Compact Training (SCT), a method that replaces dense weight matrices with permanent truncated SVD factors W = U diag(s) V^T, where the full dense matrix is never materialized during training or inference. Gradients flow through the compact spectral factors via standard backpropagation, and U, V are retracted to the Stiefel manifold via QR decomposition after each optimizer step. SCT achieves up to 199x memory reduction per MLP layer at rank 32, enabling full training steps of 70B-parameter architectures on a Steam Deck handheld (7.2 GB peak memory vs. 1,245 GB for dense FP32 training with Adam). Rank-sweep experiments on SmolLM2-1.7B (ranks 32-256, 2000 steps, NVIDIA A100) show that all tested ranks converge to the same loss floor (~4.2-4.5), identifying the learning rate schedule -- not MLP rank -- as the primary bottleneck. Rank 128 emerges as the efficiency sweet spot at 11.7x MLP compression with the lowest perplexity. GPU memory drops 46% at rank 32 while training throughput doubles.
Authors: Sayed Hashim, Frank Soboczenski, Paul Cairns
Abstract: Datasets used in immunotherapy response prediction are typically small in size, as well as diverse in cancer type, drug administered, and sequencer used. Models often drop in performance when tested on patient cohorts that are not included in the training process. Recent work has shown that transformer-based models along with self-supervised learning show better generalisation performance than threshold-based biomarkers, but is still suboptimal. We present BioCOMPASS, an extension of a transformer-based model called COMPASS, that integrates biomarkers and treatment information to further improve its generalisability. Instead of feeding biomarker data as input, we built loss components to align them with the model's intermediate representations. We found that components such as treatment gating and pathway consistency loss improved generalisability when evaluated with Leave-one-cohort-out, Leave-one-cancer-type-out and Leave-one-treatment-out strategies. Results show that building components that exploit biomarker and treatment information can help in generalisability of immunotherapy response prediction. Careful curation of additional components that leverage complementary clinical information and domain knowledge represents a promising direction for future research.
Authors: Dong-Jae Lee, Sunghyun Baek, Junmo Kim
Abstract: Large Vision Language Models show impressive performance across image and video understanding tasks, yet their computational cost grows rapidly with the number of visual tokens. Existing token pruning methods mitigate this issue through empirical approaches while overlooking the internal mechanism of attention. In this paper, we propose a novel training free token pruning framework grounded in the dual form perspective of attention. We reformulate attention as an implicit linear layer whose weight matrix is the sum of rank 1 outer products, each generated by a single token's key value pair. Token pruning thus reduces to selecting an optimal subset of these rank 1 updates that best approximates the original dual weight matrix. Extending this perspective to standard softmax attention in LVLMs, we derive a novel metric quantifying both a token's information magnitude and information duplication. To efficiently select the subset with the proposed metric, we introduce Progressive Chunked Maximal Marginal Relevance. Extensive experiments demonstrate that our method achieves a better trade off between performance and efficiency, while providing another perspective on existing pruning approaches.
Authors: Swapnil Parekh
Abstract: A new generation of language models reasons entirely in continuous hidden states, producing no tokens and leaving no audit trail. We show that this silence creates a fundamentally new attack surface. ThoughtSteer perturbs a single embedding vector at the input layer; the model's own multi-pass reasoning amplifies this perturbation into a hijacked latent trajectory that reliably produces the attacker's chosen answer, while remaining structurally invisible to every token-level defense. Across two architectures (Coconut and SimCoT), three reasoning benchmarks, and model scales from 124M to 3B parameters, ThoughtSteer achieves >=99% attack success rate with near-baseline clean accuracy, transfers to held-out benchmarks without retraining (94-100%), evades all five evaluated active defenses, and survives 25 epochs of clean fine-tuning. We trace these results to a unifying mechanism: Neural Collapse in the latent space pulls triggered representations onto a tight geometric attractor, explaining both why defenses fail and why any effective backdoor must leave a linearly separable signature (probe AUC>=0.999). Yet a striking paradox emerges: individual latent vectors still encode the correct answer even as the model outputs the wrong one. The adversarial information is not in any single vector but in the collective trajectory, establishing backdoor perturbations as a new lens for mechanistic interpretability of continuous reasoning. Code and checkpoints are available.
Authors: Dharma Teja Vooturi, Dhiraj Kalamkar, Dipankar Das, Bharat Kaul
Abstract: Pretraining Large Language Models (LLMs) from scratch requires massive amount of compute. Aurora super computer is an ExaScale machine with 127,488 Intel PVC (Ponte Vechio) GPU tiles. In this work, we showcase LLM pretraining on Aurora at the scale of 1000s of GPU tiles. Towards this effort, we developed Optimus, an inhouse training library with support for standard large model training techniques. Using Optimus, we first pretrained Mula-1B, a 1 Billion dense model and Mula-7B-A1B, a 7 Billion Mixture of Experts (MoE) model from scratch on 3072 GPU tiles for the full 4 trillion tokens of the OLMoE-mix-0924 dataset. We then demonstrated model scaling by pretraining three large MoE models Mula-20B-A2B, Mula-100B-A7B, and Mula-220B-A10B till 100 Billion tokens on the same dataset. On our largest model Mula-220B-A10B, we pushed the compute scaling from 384 to 12288 GPU tiles and observed scaling efficiency of around 90% at 12288 GPU tiles. We significantly improved the runtime performance of MoE models using custom GPU kernels for expert computation, and a novel EP-Aware sharded optimizer resulting in training speedups up to 1.71x. As part of the Optimus library, we also developed a robust set of reliability and fault tolerant features to improve training stability and continuity at scale.
Authors: Yilun Liu, Jinru Han, Sikuan Yan, Volker Tresp, Yunpu Ma
Abstract: Standard Mixture-of-Experts (MoE) models rely on centralized routing mechanisms that introduce rigid inductive biases. We propose Routing-Free MoE which eliminates any hard-coded centralized designs including external routers, Softmax, Top-K and load balancing, instead encapsulating all activation functionalities within individual experts and directly optimized through continuous gradient flow, enabling each expert to determine its activation entirely on its own. We introduce a unified adaptive load-balancing framework to simultaneously optimize both expert-balancing and token-balancing objectives through a configurable interpolation, allowing flexible and customizable resource allocation. Extensive experiments show that Routing-Free MoE can consistently outperform baselines with better scalability and robustness. We analyze its behavior in detail and offer insights that may facilitate future MoE design ad optimization.
Authors: Sicheng Zuo, Zixun Xie, Wenzhao Zheng, Shaoqing Xu, Fang Li, Hanbing Li, Long Chen, Zhi-Xin Yang, Jiwen Lu
Abstract: End-to-end autonomous driving has evolved from the conventional paradigm based on sparse perception into vision-language-action (VLA) models, which focus on learning language descriptions as an auxiliary task to facilitate planning. In this paper, we propose an alternative Vision-Geometry-Action (VGA) paradigm that advocates dense 3D geometry as the critical cue for autonomous driving. As vehicles operate in a 3D world, we think dense 3D geometry provides the most comprehensive information for decision-making. However, most existing geometry reconstruction methods (e.g., DVGT) rely on computationally expensive batch processing of multi-frame inputs and cannot be applied to online planning. To address this, we introduce a streaming Driving Visual Geometry Transformer (DVGT-2), which processes inputs in an online manner and jointly outputs dense geometry and trajectory planning for the current frame. We employ temporal causal attention and cache historical features to support on-the-fly inference. To further enhance efficiency, we propose a sliding-window streaming strategy and use historical caches within a certain interval to avoid repetitive computations. Despite the faster speed, DVGT-2 achieves superior geometry reconstruction performance on various datasets. The same trained DVGT-2 can be directly applied to planning across diverse camera configurations without fine-tuning, including closed-loop NAVSIM and open-loop nuScenes benchmarks.
Authors: Hemanth Kotaprolu, Kishan Maharaj, Raey Zhao, Abhijit Mishra, Pushpak Bhattacharyya
Abstract: Understanding emotions in natural language is inherently a multi-dimensional reasoning problem, where multiple affective signals interact through context, interpersonal relations, and situational cues. However, most existing emotion understanding benchmarks rely on short texts and predefined emotion labels, reducing this process to independent label prediction and ignoring the structured dependencies among emotions. To address this limitation, we introduce Emotional Scenarios (EmoScene), a theory-grounded benchmark of 4,731 context-rich scenarios annotated with an 8-dimensional emotion vector derived from Plutchik's basic emotions. We evaluate six instruction-tuned large language models in a zero-shot setting and observe modest performance, with the best model achieving a Macro F1 of 0.501, highlighting the difficulty of context-aware multi-label emotion prediction. Motivated by the observation that emotions rarely occur independently, we further propose an entanglement-aware Bayesian inference framework that incorporates emotion co-occurrence statistics to perform joint posterior inference over the emotion vector. This lightweight post-processing improves structural consistency of predictions and yields notable gains for weaker models (e.g., +0.051 Macro F1 for Qwen2.5-7B). EmoScene therefore provides a challenging benchmark for studying multi-dimensional emotion understanding and the limitations of current language models.
Authors: Zhanzhi Lou, Hui Chen, Yibo Li, Qian Wang, Bryan Hooi
Abstract: Test-Time Learning (TTL) enables language agents to iteratively refine their performance through repeated interactions with the environment at inference time. At the core of TTL is an adaptation policy that updates the actor policy based on experience from previous episodes, thereby improving future behavior. Existing methods rely on fixed, hand-crafted adaptation policies rather than optimizing them for downstream improvement. We argue that optimal adaptation policies should be learned from task environments, not hand-engineered based on human intuition. To achieve this, we introduce Meta-TTL, a framework that formulates the discovery of effective adaptation policies as a bi-level optimization problem. Within this framework, the inner loop executes the standard TTL process, measuring how effectively a candidate adaptation policy helps an agent correct errors across sequential episodes. Guided by the agent's performance, the outer loop employs evolutionary search over a diverse distribution of training tasks to iteratively refine the adaptation policy. We evaluate Meta-TTL on Jericho and WebArena-Lite across both in-distribution (ID) and out-of-distribution (OOD) settings, using multiple meta-agent backbones. Results on both benchmarks show that Meta-TTL consistently outperforms hand-crafted baselines, suggesting that the optimized adaptation policy encodes transferable strategies that generalize beyond the training task distribution.
Authors: Abdullah Al Shafi, Md. Milon Islam, Sk. Imran Hossain, K. M. Azharul Hasan
Abstract: Actor-level stance detection aims to determine an author expressed position toward specific geopolitical actors mentioned or implicated in a text. Although transformer-based models have achieved relatively good performance in stance classification, they typically rely on unified representations that may not sufficiently capture heterogeneous linguistic signals, such as contrastive discourse structures, framing cues, and salient lexical indicators. This motivates the need for adaptive architectures that explicitly model diverse stance-expressive patterns. In this paper, we propose StanceMoE, a context-enhanced Mixture-of-Experts (MoE) architecture built upon a fine-tuned BERT encoder for actor-level stance detection. Our model integrates six expert modules designed to capture complementary linguistic signals, including global semantic orientation, salient lexical cues, clause-level focus, phrase-level patterns, framing indicators, and contrast-driven discourse shifts. A context-aware gating mechanism dynamically weights expert contributions, enabling adaptive routing based on input characteristics. Experiments are conducted on the StanceNakba 2026 Subtask A dataset, comprising 1,401 annotated English texts where the target actor is implicit in the text. StanceMoE achieves a macro-F1 score of 94.26%, outperforming traditional baselines, and alternative BERT-based variants.
Authors: Nan Wang, Zhiwei Jin, Chen Chen, Haonan Lu
Abstract: Document understanding and GUI interaction are among the highest-value applications of Vision-Language Models (VLMs), yet they impose exceptionally heavy computational burden: fine-grained text and small UI elements demand high-resolution inputs that produce tens of thousands of visual tokens. We observe that this cost is largely wasteful -- across document and GUI benchmarks, only 22--71\% of image patches are pixel-unique, the rest being exact duplicates of another patch in the same image. We propose \textbf{PixelPrune}, which exploits this pixel-level redundancy through predictive-coding-based compression, pruning redundant patches \emph{before} the Vision Transformer (ViT) encoder. Because it operates in pixel space prior to any neural computation, PixelPrune accelerates both the ViT encoder and the downstream LLM, covering the full inference pipeline. The method is training-free, requires no learnable parameters, and supports pixel-lossless compression ($\tau{=}0$) as well as controlled lossy compression ($\tau{>}0$). Experiments across three model scales and document and GUI benchmarks show that PixelPrune maintains competitive task accuracy while delivering up to 4.2$\times$ inference speedup and 1.9$\times$ training acceleration. Code is available at https://github.com/OPPO-Mente-Lab/PixelPrune.
Authors: Razvan Mihai Popescu, David Gros, Andrei Botocan, Rahul Pandita, Prem Devanbu, Maliheh Izadi
Abstract: The rise of large language models for code has reshaped software development. Autonomous coding agents, able to create branches, open pull requests, and perform code reviews, now actively contribute to real-world projects. Their growing role offers a unique and timely opportunity to investigate AI-driven contributions and their effects on code quality, team dynamics, and software maintainability. In this work, we construct a novel dataset of approximately $110,000$ open-source pull requests, including associated commits, comments, reviews, issues, and file changes, collectively representing millions of lines of source code. We compare five popular coding agents, including OpenAI Codex, Claude Code, GitHub Copilot, Google Jules, and Devin, examining how their usage differs in various development aspects such as merge frequency, edited file types, and developer interaction signals, including comments and reviews. Furthermore, we emphasize that code authoring and review are only a small part of the larger software engineering process, as the resulting code must also be maintained and updated over time. Hence, we offer several longitudinal estimates of survival and churn rates for agent-generated versus human-authored code. Ultimately, our findings indicate an increasing agent activity in open-source projects, although their contributions are associated with more churn over time compared to human-authored code.
Authors: Dylan B. Lewis, Jens Gregor, Hector Santos-Villalobos
Abstract: Modern vision pipelines increasingly rely on pretrained image encoders whose representations are reused across tasks and models, yet these representations are often overcomplete and model-specific. We propose a simple, training-free method to improve the efficiency of image representations via a post-hoc canonical correlation analysis (CCA) operator. By leveraging the shared structure between representations produced by two pre-trained image encoders, our method finds linear projections that serve as a principled form of representation selection and dimensionality reduction, retaining shared semantic content while discarding redundant dimensions. Unlike standard dimensionality reduction techniques such as PCA, which operate on a single embedding space, our approach leverages cross-model agreement to guide representation distillation and refinement. The technique allows representations to be reduced by more than 75% in dimensionality with improved downstream performance, or enhanced at fixed dimensionality via post-hoc representation transfer from larger or fine-tuned models. Empirical results on ImageNet-1k, CIFAR-100, MNIST, and additional benchmarks show consistent improvements over both baseline and PCA-projected representations, with accuracy gains of up to 12.6%.
Authors: Arina Kharlamova, Bowei He, Chen Ma, Xue Liu
Abstract: We present DANCEMATCH, an end-to-end framework for motion-based dance retrieval, the task of identifying semantically similar choreographies directly from raw video, defined as DANCE FINGERPRINTING. While existing motion analysis and retrieval methods can compare pose sequences, they rely on continuous embeddings that are difficult to index, interpret, or scale. In contrast, DANCEMATCH constructs compact, discrete motion signatures that capture the spatio-temporal structure of dance while enabling efficient large-scale retrieval. Our system integrates Skeleton Motion Quantisation (SMQ) with Spatio-Temporal Transformers (STT) to encode human poses, extracted via Apple CoMotion, into a structured motion vocabulary. We further design DANCE RETRIEVAL ENGINE (DRE), which performs sub-linear retrieval using a histogram-based index followed by re-ranking for refined matching. To facilitate reproducible research, we release DANCETYPESBENCHMARK, a pose-aligned dataset annotated with quantised motion tokens. Experiments demonstrate robust retrieval across diverse dance styles and strong generalisation to unseen choreographies, establishing a foundation for scalable motion fingerprinting and quantitative choreographic analysis.
Authors: Hsin-Ling Hsu, Min-Yu Chen, Nai-Chia Chen, Yan-Ru Chen, Yi-Ling Chang, Fang Yu
Abstract: Transformer-based NLP models remain vulnerable to adversarial perturbations, yet existing repair methods face a fundamental trade-off: gradient-based approaches offer flexibility but lack verifiability and often overfit; methods that do provide repair guarantees are restricted to the final layer or small networks, significantly limiting the parameter search space available for repair. We present WARP (Weight-Adjusted Repair with Provability), a constraint-based repair framework that extends repair beyond the last layer of Transformer models. WARP formulates repair as a convex quadratic program derived from a first-order linearization of the logit gap, enabling tractable optimization over a high-dimensional parameter space. Under the condition that the first-order approximation holds, this formulation induces three per-sample guarantees: (i) a positive margin constraint ensuring correct classification on repaired inputs, (ii) preservation constraints over a designated remain set, and (iii) a certified robustness radius derived from Lipschitz continuity. To ensure feasibility across varying model architectures, we introduce a sensitivity-based preprocessing step that conditions the optimization landscape accordingly. We further show that the iterative optimization procedure converges to solutions satisfying all repair constraints under mild assumptions. Empirical evaluation on encoder-only Transformers with varying layer architectures validates that these guarantees hold in practice while improving robustness to adversarial inputs. Our results demonstrate that guaranteed, generalizable Transformer repair is achievable through principled constraint-based optimization.
Authors: Ruijie Hao, Longfei Zhang, Yang Dai, Yang Ma, Xingxing Liang, Guangquan Cheng
Abstract: Reinforcement Learning (RL) has proven highly effective in addressing complex control and decision-making tasks. However, in most traditional RL algorithms, the policy is typically parameterized as a diagonal Gaussian distribution, which constrains the policy from capturing multimodal distributions, making it difficult to cover the full range of optimal solutions in multi-solution problems, and the return is reduced to a mean value, losing its multimodal nature and thus providing insufficient guidance for policy updates. In response to these problems, we propose a RL algorithm termed flow-based policy with distributional RL (FP-DRL). This algorithm models the policy using flow matching, which offers both computational efficiency and the capacity to fit complex distributions. Additionally, it employs a distributional RL approach to model and optimize the entire return distribution, thereby more effectively guiding multimodal policy updates and improving agent performance. Experimental trails on MuJoCo benchmarks demonstrate that the FP-DRL algorithm achieves state-of-the-art (SOTA) performance in most MuJoCo control tasks while exhibiting superior representation capability of the flow policy.
Authors: Xiangqi Wang, Yue Huang, Haomin Zhuang, Kehan Guo, Xiangliang Zhang
Abstract: Current aligned language models exhibit a dual failure mode we term the Evasive Servant: they sycophantically validate flawed user beliefs while deflecting responsibility with boilerplate disclaimers. We propose the Dignified Peer framework, which counters servility with anti-sycophancy and trustworthiness, and mitigates evasiveness through empathy and creativity. Realizing this agent requires overcoming significant challenges in data supervision, objective collapse, and evaluation bias. We address these issues by introducing the PersonaKnob dataset which features a compositional partial order structure of multiple persona preference. This data is utilized alongside a tolerant constrained Lagrangian DPO algorithm that dynamically balances all persona dimensions to prevent behavioral collapse. Additionally, we employ a psychometrically calibrated Item Response Theory evaluation protocol to disentangle latent model persona capability from confounders like judge biases. Extensive empirical studies demonstrate that our approach successfully build a LLM agent with both dignity and peer.
Authors: Zhengyang Tang, Ke Ji, Xidong Wang, Zihan Ye, Xinyuan Wang, Yiduo Guo, Ziniu Li, Chenxin Li, Jingyuan Hu, Shunian Chen, Tongxu Luo, Jiaxi Bi, Zeyu Qin, Shaobo Wang, Xin Lai, Pengyuan Lyu, Junyi Li, Can Xu, Chengquan Zhang, Han Hu, Ming Yan, Benyou Wang
Abstract: We study whether phone-use agents respect privacy while completing benign mobile tasks. This question has remained hard to answer because privacy-compliant behavior is not operationalized for phone-use agents, and ordinary apps do not reveal exactly what data agents type into which form entries during execution. To make this question measurable, we introduce MyPhoneBench, a verifiable evaluation framework for privacy behavior in mobile agents. We operationalize privacy-respecting phone use as permissioned access, minimal disclosure, and user-controlled memory through a minimal privacy contract, iMy, and pair it with instrumented mock apps plus rule-based auditing that make unnecessary permission requests, deceptive re-disclosure, and unnecessary form filling observable and reproducible. Across five frontier models on 10 mobile apps and 300 tasks, we find that task success, privacy-compliant task completion, and later-session use of saved preferences are distinct capabilities, and no single model dominates all three. Evaluating success and privacy jointly reshuffles the model ordering relative to either metric alone. The most persistent failure mode across models is simple data minimization: agents still fill optional personal entries that the task does not require. These results show that privacy failures arise from over-helpful execution of benign tasks, and that success-only evaluation overestimates the deployment readiness of current phone-use agents. All code, mock apps, and agent trajectories are publicly available at~ https://github.com/FreedomIntelligence/MyPhoneBench.
Authors: Yi Cao, Zexun Chen, Lin William Cong, Heqing Shi
Abstract: We develop Structured-Knowledge-Informed Neural Networks (SKINNs), a unified estimation framework that embeds theoretical, simulated, previously learned, or cross-domain insights as differentiable constraints within flexible neural function approximation. SKINNs jointly estimate neural network parameters and economically meaningful structural parameters in a single optimization problem, enforcing theoretical consistency not only on observed data but over a broader input domain through collocation, and therefore nesting approaches such as functional GMM, Bayesian updating, transfer learning, PINNs, and surrogate modeling. SKINNs define a class of M-estimators that are consistent and asymptotically normal with root-N convergence, sandwich covariance, and recovery of pseudo-true parameters under misspecification. We establish identification of structural parameters under joint flexibility, derive generalization and target-risk bounds under distributional shift in a convex proxy, and provide a restricted-optimal characterization of the weighting parameter that governs the bias-variance tradeoff. In an illustrative financial application to option pricing, SKINNs improve out-of-sample valuation and hedging performance, particularly at longer horizons and during high-volatility regimes, while recovering economically interpretable structural parameters with improved stability relative to conventional calibration. More broadly, SKINNs provide a general econometric framework for combining model-based reasoning with high-dimensional, data-driven estimation.
Authors: Daniel Miehling, Sandra Kuebler
Abstract: YouTube Shorts have become central to news consumption on the platform, yet research on how geopolitical events are represented in this format remains limited. To address this gap, we present a multimodal pipeline that combines automatic transcription, aspect-based sentiment analysis (ABSA), and semantic scene classification. The pipeline is first assessed for feasibility and then applied to analyze short-form coverage of the Israel-Hamas war by state-funded outlets. Using over 2,300 conflict-related Shorts and more than 94,000 visual frames, we systematically examine war reporting across major international broadcasters. Our findings reveal that the sentiment expressed in transcripts regarding specific aspects differs across outlets and over time, whereas scene-type classifications reflect visual cues consistent with real-world events. Notably, smaller domain-adapted models outperform large transformers and even LLMs for sentiment analysis, underscoring the value of resource-efficient approaches for humanities research. The pipeline serves as a template for other short-form platforms, such as TikTok and Instagram, and demonstrates how multimodal methods, combined with qualitative interpretation, can characterize sentiment patterns and visual cues in algorithmically driven video environments.
Authors: Jinkun Hao, Mingda Jia, Ruiyan Wang, Xihui Liu, Ran Yi, Lizhuang Ma, Jiangmiao Pang, Xudong Xu
Abstract: We introduce EgoSim, a closed-loop egocentric world simulator that generates spatially consistent interaction videos and persistently updates the underlying 3D scene state for continuous simulation. Existing egocentric simulators either lack explicit 3D grounding, causing structural drift under viewpoint changes, or treat the scene as static, failing to update world states across multi-stage interactions. EgoSim addresses both limitations by modeling 3D scenes as updatable world states. We generate embodiment interactions via a Geometry-action-aware Observation Simulation model, with spatial consistency from an Interaction-aware State Updating module. To overcome the critical data bottleneck posed by the difficulty in acquiring densely aligned scene-interaction training pairs, we design a scalable pipeline that extracts static point clouds, camera trajectories, and embodiment actions from in-the-wild large-scale monocular egocentric videos. We further introduce EgoCap, a capture system that enables low-cost real-world data collection with uncalibrated smartphones. Extensive experiments demonstrate that EgoSim significantly outperforms existing methods in terms of visual quality, spatial consistency, and generalization to complex scenes and in-the-wild dexterous interactions, while supporting cross-embodiment transfer to robotic manipulation. Codes and datasets will be open soon. The project page is at egosimulator.github.io.
Authors: Yiheng Wang, Lichen Zhu, Yueqian Lin, Yudong Liu, Jingyang Zhang, Hai "Helen" Li, Yiran Chen
Abstract: Multimodal Large Language Models (MLLMs) have shown strong performance on video question answering, but their application to long-form videos is constrained by limited context length and computational cost, making keyframe sampling essential. Existing approaches typically rely on semantic relevance or reinforcement learning, which either fail to capture evidential clues or suffer from inefficient combinatorial optimization. In this work, we propose an evidence-driven keyframe sampling framework grounded in information bottleneck theory. We formulate keyframe selection as maximizing the conditional mutual information between selected frames and the query, providing a principled objective that reflects each frame's contribution to answering the question. To make this objective tractable, we exploit its structure to derive a decomposed optimization that reduces subset selection to independent frame-level scoring. We further introduce a query-conditioned evidence scoring network trained with a contrastive objective to estimate evidential importance efficiently. Experiments on long-form video understanding benchmarks show that our method consistently outperforms prior sampling strategies under strict token budgets, while significantly improving training efficiency.
Authors: Yiru Wang, Xinyue Shen, Yaohui Han, Michael Backes, Pin-Yu Chen, Tsung-Yi Ho
Abstract: While large language model-based multi-agent systems have shown strong potential for complex reasoning, how to effectively organize multiple agents remains an open question. In this paper, we introduce OrgAgent, a company-style hierarchical multi-agent framework that separates collaboration into governance, execution, and compliance layers. OrgAgent decomposes multi-agent reasoning into three layers: a governance layer for planning and resource allocation, an execution layer for task solving and review, and a compliance layer for final answer control. By evaluating the framework across reasoning tasks, LLMs, execution modes, and execution policies, we find that multi-agent systems organized in a company-style hierarchy generally outperform other organizational structures. Besides, hierarchical coordination also reduces token consumption relative to flat collaboration in most settings. For example, for GPT-OSS-120B, the hierarchical setting improves performance over flat multi-agent system by 102.73% while reducing token usage by 74.52% on SQuAD 2.0. Further analysis shows that hierarchy helps most when tasks benefit from stable skill assignment, controlled information flow, and layered verification. Overall, our findings highlight organizational structure as an important factor in multi-agent reasoning, shaping not only effectiveness and cost, but also coordination behavior.
Authors: Rafael Sojo, Pedro Larra\~naga, Concha Bielza
Abstract: This paper introduces two transfer learning methodologies for estimating nonparametric Bayesian networks under scarce data. We propose two algorithms, a constraint-based structure learning method, called PC-stable-transfer learning (PCS-TL), and a score-based method, called hill climbing transfer learning (HC-TL). We also define particular metrics to tackle the negative transfer problem in each of them, a situation in which transfer learning has a negative impact on the model's performance. Then, for the parameters, we propose a log-linear pooling approach. For the evaluation, we learn kernel density estimation Bayesian networks, a type of nonparametric Bayesian network, and compare their transfer learning performance with the models alone. To do so, we sample data from small, medium and large-sized synthetic networks and datasets from the UCI Machine Learning repository. Then, we add noise and modifications to these datasets to test their ability to avoid negative transfer. To conclude, we perform a Friedman test with a Bergmann-Hommel post-hoc analysis to show statistical proof of the enhanced experimental behavior of our methods. Thus, PCS-TL and HC-TL demonstrate to be reliable algorithms for improving the learning performance of a nonparametric Bayesian network with scarce data, which in real industrial environments implies a reduction in the required time to deploy the network.
Authors: Zhichen Liu, Tianle Lun, Zhibin Wen, Hao An, Yulin Ou, Jianhui Xu, Hao Zhang, Wenyi Fang, Yang Zheng, Yang Xu
Abstract: The paradigm of scaling Large Language Models (LLMs) in both parameter size and test time has pushed the boundaries of AI capabilities, but at the cost of making the traditional generative evaluation paradigm prohibitively expensive, therefore making the latency of LLM's in-training downstream performance evaluation unbearable. However, simple metrics like training loss (perplexity) are not always correlated with downstream performance, as sometimes their trends diverge from the actual task outcomes. This dilemma calls for a method that is computationally efficient and sufficiently accurate in measuring model capabilities. To address this challenge, we introduce a new in-training evaluation paradigm that uses a lightweight probe for monitoring downstream performance. The probes take the internal representations of LLM checkpoints (during training) as input and directly predict the checkpoint's performance on downstream tasks measured by success probability (i.e., pass@1). We design several probe architectures, validating their effectiveness using the OLMo3-7B's checkpoints across a diverse set of downstream tasks. The probes can accurately predict a checkpoint's performance (with avg. AUROC$>$0.75), have decent generalizability across checkpoints (earlier predicts later), and reduce the computation latency from $\sim$1 hr (using conventional generative evaluation method) to $\sim$3 min. In sum, this work presents a practical and scalable in-training downstream evaluation paradigm, enabling a more agile, informed, and efficient LLM development process.
Authors: Jingjie Ning, Xueqi Li, Chengyu Yu
Abstract: Multi-LLM revision pipelines, in which a second model reviews and improves a draft produced by a first, are widely assumed to derive their gains from genuine error correction. We question this assumption with a controlled decomposition experiment that uses four matched conditions to separate second-pass gains into three additive components: re-solving, scaffold, and content. We evaluate this design across two model pairs on three benchmarks spanning knowledge-intensive MCQ and competitive programming. Our results show that the gains of multi-LLM revision are not monolithic, but depend on task structure, draft quality, and the type of draft information. On MCQ tasks, where the answer space is constrained and drafts provide little structural guidance, most gains are consistent with stronger-model re-solving, and directly routing queries to the stronger model can be more effective than revising a weak draft. On code generation tasks, however, two-stage prompting remains useful because even semantically null drafts can provide substantial structural scaffolding, while weak draft content can be harmful. Finally, role-reversed experiments show that strong drafts clearly benefit weak reviewers. Ultimately, our findings demonstrate that the utility of multi-LLM revision is dynamically bottlenecked by task structure and draft quality, necessitating more targeted pipeline designs rather than blanket revision strategies.
Authors: Mona Schirmer, Anton Thielmann, Pola Schw\"obel, Thomas Martynec, Giuseppe Di Benedetto, Ben London, Yannik Stein
Abstract: Popularity bias is a pervasive problem in recommender systems, where recommendations disproportionately favor popular items. This not only results in "rich-get-richer" dynamics and a homogenization of visible content, but can also lead to misalignment of recommendations with individual users' preferences for popular or niche content. This work studies popularity bias through the lens of user-recommender alignment. To this end, we introduce Popularity Quantile Calibration, a measurement framework that quantifies misalignment between a user's historical popularity preference and the popularity of their recommendations. Building on this notion of popularity alignment, we propose SPREE, an inference-time mitigation method for sequential recommenders based on activation steering. SPREE identifies a popularity direction in representation space and adaptively steers model activations based on an estimate of each user's personal popularity bias, allowing both the direction and magnitude of steering to vary across users. Unlike global debiasing approaches, SPREE explicitly targets alignment rather than uniformly reducing popularity. Experiments across multiple datasets show that SPREE consistently improves user-level popularity alignment while preserving recommendation quality.
Authors: Anubhab Sahu, Diptisha Samanta, Reza Soosahabi
Abstract: System Instructions in Large Language Models (LLMs) are commonly used to enforce safety policies, define agent behavior, and protect sensitive operational context in agentic AI applications. These instructions may contain sensitive information such as API credentials, internal policies, and privileged workflow definitions, making system instruction leakage a critical security risk highlighted in the OWASP Top 10 for LLM Applications. Without incurring the overhead costs of reasoning models, many LLM applications rely on refusal-based instructions that block direct requests for system instructions, implicitly assuming that prohibited information can only be extracted through explicit queries. We introduce an automated evaluation framework that tests whether system instructions remain confidential when extraction requests are re-framed as encoding or structured output tasks. Across four common models and 46 verified system instructions, we observe high attack success rates (> 0.7) for structured serialization where models refuse direct extraction requests but disclose protected content in the requested serialization formats. We further demonstrate a mitigation strategy based on one-shot instruction reshaping using a Chain-of-Thought reasoning model, indicating that even subtle changes in wording and structure of system instructions can significantly reduce attack success rate without requiring model retraining.
Authors: Deemah H. Tashman, Soumaya Cherkaoui
Abstract: Next-generation (NextG) cellular networks are designed to support emerging applications with diverse data rate and latency requirements, such as immersive multimedia services and large-scale Internet of Things deployments. A key enabling mechanism is radio access network (RAN) slicing, which dynamically partitions radio resources into virtual resource blocks to efficiently serve heterogeneous traffic classes, including enhanced mobile broadband (eMBB), massive machine-type communications (mMTC), and ultra-reliable low-latency communications (URLLC). In this paper, we study the impact of adversarial attacks on AI-driven RAN slicing decisions, where a budget-constrained adversary selectively jams slice transmissions to bias deep reinforcement learning (DRL)-based resource allocation, and quantify the resulting service level agreement (SLA) violations and post-attack recovery behavior. Our results indicate that budget-constrained adversarial jamming can induce severe and slice-dependent steady-state SLA violations. Moreover, the DRL agent's reward converges toward the clean baseline only after a non-negligible recovery period.
Authors: Ying Xie
Abstract: "Vibe coding," in which developers delegate code generation to AI assistants and accept the output with little manual review, has gained rapid adoption in production settings. On March 31, 2026, Anthropic's Claude Code CLI shipped a 59.8 MB source map file in its npm package, exposing roughly 512,000 lines of proprietary TypeScript. The tool had itself been largely vibe-coded, and the leak traced to a misconfigured packaging rule rather than a logic bug. Existing static-analysis and secret-scanning tools did not cover this failure mode, pointing to a gap between the vulnerabilities AI tends to introduce and the vulnerabilities current tooling is built to find. We present VibeGuard, a pre-publish security gate that targets five such blind spots: artifact hygiene, packaging-configuration drift, source-map exposure, hardcoded secrets, and supply-chain risk. In controlled experiments on eight synthetic projects (seven vulnerable, one clean control), VibeGuard achieved 100% recall, 89.47% precision (F1 = 94.44%), and correct pass/fail gate decisions on all eight projects across three policy levels. We discuss how these results inform a defense-in-depth workflow for teams that rely on AI code generation.
Authors: Awais Khan, Muhammad Umar Farooq, Kutub Uddin, Khalid Malik
Abstract: Partial audio deepfakes, where synthesized segments are spliced into genuine recordings, are particularly deceptive because most of the audio remains authentic. Existing detectors are supervised: they require frame-level annotations, overfit to specific synthesis pipelines, and must be retrained as new generative models emerge. We argue that this supervision is unnecessary. We hypothesize that speech foundation models implicitly encode a forensic signal: genuine speech forms smooth, slowly varying embedding trajectories, while splice boundaries introduce abrupt disruptions in frame-level transitions. Building on this, we propose TRACE (Training-free Representation-based Audio Countermeasure via Embedding dynamics), a training-free framework that detects partial audio deepfakes by analyzing the first-order dynamics of frozen speech foundation model representations without any training, labeled data, or architectural modification. We evaluate TRACE on four benchmarks that span two languages using six speech foundation models. In PartialSpoof, TRACE achieves 8.08% EER, competitive with fine-tuned supervised baselines. In LlamaPartialSpoof, the most challenging benchmark featuring LLM-driven commercial synthesis, TRACE surpasses a supervised baseline outright (24.12% vs. 24.49% EER) without any target-domain data. These results show that temporal dynamics in speech foundation models provide an effective, generalize signal for training-free audio forensics.
Authors: Anooshka Bajaj, Deven Mahesh Mistry, Sahaj Singh Maini, Yash Aggarwal, Billy Dickson, Zoran Tiganj
Abstract: Large language models (LLMs) exhibit strong in-context learning capabilities, but how they track and retrieve information from context remains underexplored. Drawing on the free recall paradigm in cognitive science (where participants recall list items in any order), we show that several open-source LLMs consistently display a serial-recall-like pattern, assigning peak probability to tokens that immediately follow a repeated token in the input sequence. Through systematic ablation experiments, we show that induction heads, specialized attention heads that attend to the token following a previous occurrence of the current token, play an important role in this phenomenon. Removing heads with a high induction score substantially reduces the +1 lag bias, whereas ablating random heads does not reproduce the same reduction. We also show that removing heads with high induction scores impairs the performance of models prompted to do serial recall using few-shot learning to a larger extent than removing random heads. Our findings highlight a mechanistically specific connection between induction heads and temporal context processing in transformers, suggesting that these heads are especially important for ordered retrieval and serial-recall-like behavior during in-context learning.
Authors: Jinzhao Li, Nan Jiang, Yexiang Xue
Abstract: Stochastic Multi-Objective Optimization (SMOO) is critical for decision-making trading off multiple potentially conflicting objectives in uncertain environments. SMOO aims at identifying the Pareto frontier, which contains all mutually non-dominating decisions. The problem is highly intractable due to the embedded probabilistic inference, such as computing the marginal, posterior probabilities, or expectations. Existing methods, such as scalarization, sample average approximation, and evolutionary algorithms, either offer arbitrarily loose approximations or may incur prohibitive computational costs. We propose XOR-SMOO, a novel algorithm that with probability $1-\delta$, obtains $\gamma$-approximate Pareto frontiers ($\gamma>1$) for SMOO by querying an SAT oracle poly-log times in $\gamma$ and $\delta$. A $\gamma$-approximate Pareto frontier is only below the true frontier by a fixed, multiplicative factor $\gamma$. Thus, XOR-SMOO solves highly intractable SMOO problems (\#P-hard) with only queries to SAT oracles while obtaining tight, constant factor approximation guarantees. Experiments on real-world road network strengthening and supply chain design problems demonstrate that XOR-SMOO outperforms several baselines in identifying Pareto frontiers that have higher objective values, better coverage of the optimal solutions, and the solutions found are more evenly distributed. Overall, XOR-SMOO significantly enhanced the practicality and reliability of SMOO solvers.
Authors: Griffin Pitts, Neha Rani, Weedguet Mildort
Abstract: As generative AI systems are integrated into educational settings, students often encounter AI-generated output while working through learning tasks, either by requesting help or through integrated tools. Trust in AI can influence how students interpret and use that output, including whether they evaluate it critically or exhibit overreliance. We investigate how students' trust relates to their appropriate reliance on an AI assistant during programming problem-solving tasks, and whether this relationship differs by learner characteristics. With 432 undergraduate participants, students' completed Python output-prediction problems while receiving recommendations and explanations from an AI chatbot, including accurate and intentionally misleading suggestions. We operationalize reliance behaviorally as the extent to which students' responses reflected appropriate use of the AI assistant's suggestions, accepting them when they were correct and rejecting them when they were incorrect. Pre- and post-task surveys assessed trust in the assistant, AI literacy, need for cognition, programming self-efficacy, and programming literacy. Results showed a non-linear relationship in which higher trust was associated with lower appropriate reliance, suggesting weaker discrimination between correct and incorrect recommendations. This relationship was significantly moderated by students' AI literacy and need for cognition. These findings highlight the need for future work on instructional and system supports that encourage more reflective evaluation of AI assistance during problem-solving.
Authors: Reyhaneh Ahani Manghotay (Simon Fraser University, Burnaby, Canada), Jie Liang (Eastern Institute of Technology, Ningbo, China)
Abstract: Leveraging the rich semantic features of vision-language models (VLMs) like CLIP for monocular depth estimation tasks is a promising direction, yet often requires extensive fine-tuning or lacks geometric precision. We present a parameter-efficient framework, named MoA-DepthCLIP, that adapts pretrained CLIP representations for monocular depth estimation with minimal supervision. Our method integrates a lightweight Mixture-of-Adapters (MoA) module into the pretrained Vision Transformer (ViT-B/32) backbone combined with selective fine-tuning of the final layers. This design enables spatially-aware adaptation, guided by a global semantic context vector and a hybrid prediction architecture that synergizes depth bin classification with direct regression. To enhance structural accuracy, we employ a composite loss function that enforces geometric constraints. On the NYU Depth V2 benchmark, MoA-DepthCLIP achieves competitive results, significantly outperforming the DepthCLIP baseline by improving the $\delta_1$ accuracy from 0.390 to 0.745 and reducing the RMSE from 1.176 to 0.520. These results are achieved while requiring substantially few trainable parameters, demonstrating that lightweight, prompt-guided MoA is a highly effective strategy for transferring VLM knowledge to fine-grained monocular depth estimation tasks.
Authors: Atsuyuki Miyai, Mashiro Toyooka, Zaiying Zhao, Kenta Watanabe, Toshihiko Yamasaki, Kiyoharu Aizawa
Abstract: This paper introduces the first systematic evaluation framework for quantifying the quality and risks of papers written by modern coding agents. While AI-driven paper writing has become a growing concern, rigorous evaluation of the quality and potential risks of AI-written papers remains limited, and a unified understanding of their reliability is still lacking. We introduce Paper Reconstruction Evaluation (PaperRecon), an evaluation framework in which an overview (overview.md) is created from an existing paper, after which an agent generates a full paper based on the overview and minimal additional resources, and the result is subsequently compared against the original paper. PaperRecon disentangles the evaluation of the AI-written papers into two orthogonal dimensions, Presentation and Hallucination, where Presentation is evaluated using a rubric and Hallucination is assessed via agentic evaluation grounded in the original paper source. For evaluation, we introduce PaperWrite-Bench, a benchmark of 51 papers from top-tier venues across diverse domains published after 2025. Our experiments reveal a clear trade-off: while both ClaudeCode and Codex improve with model advances, ClaudeCode achieves higher presentation quality at the cost of more than 10 hallucinations per paper on average, whereas Codex produces fewer hallucinations but lower presentation quality. This work takes a first step toward establishing evaluation frameworks for AI-driven paper writing and improving the understanding of its risks within the research community.
Authors: Maofeng Tang, Hairong Qi
Abstract: Due to the large footprint of pixels in remote sensing imagery, hyperspectral unmixing (HU) has become an important and necessary procedure in hyperspectral image analysis. Traditional HU methods rely on a prior spectral mixing model, especially for nonlinear mixtures, which has largely limited the performance and generalization capacity of the unmixing approach. In this paper, we address the challenging problem of hyperspectral nonlinear unmixing (HNU) without explicit knowledge of the mixing model. Inspired by the principle of generative models, where images of the same distribution can be generated as that of the training images without knowing the exact probability distribution function of the image, we develop an invertible mixing-unmixing process via a bi-directional GAN framework, constrained by both the cycle consistency and the linkage between linear and nonlinear mixtures. The combination of cycle consistency and linear linkage provides powerful constraints without requiring an explicit mixing model. We refer to the proposed approach as the linearly-constrained CycleGAN unmixing net, or LCGU net. Experimental results indicate that the proposed LCGU net exhibits stable and competitive performance across different datasets compared with other state-of-the-art model-based HNU methods.
Authors: Mohammad R. Abu Ayyash
Abstract: We present Brainstacks, a modular architecture for continual multi-domain fine-tuning of large language models that packages domain expertise as frozen adapter stacks composing additively on a shared frozen base at inference. Five interlocking components: (1) MoE-LoRA with Shazeer-style noisy top-2 routing across all seven transformer projections under QLoRA 4-bit quantization with rsLoRA scaling; (2) an inner loop performing residual boosting by freezing trained stacks and adding new ones; (3) an outer loop training sequential domain-specific stacks with curriculum-ordered dependencies; (4) null-space projection via randomized SVD constraining new stacks to subspaces orthogonal to prior directions, achieving zero forgetting in isolation; (5) an outcome-based sigmoid meta-router trained on empirically discovered domain-combination targets that selectively weights stacks, enabling cross-domain composition. Two boundary experiments: (6) PSN pretraining on a randomly initialized model; (7) per-domain RL (DPO/GRPO) validating compatibility with post-SFT alignment. Validated on TinyLlama-1.1B (4 domains, 9 stacks) and Gemma 3 12B IT (5 domains, 10 stacks), MoE-LoRA achieves 2.5x faster convergence than parameter-matched single LoRA, residual boosting breaks through the single-stack ceiling, and the routed system recovers generation quality destroyed by ungated stack accumulation. The central finding: the outcome-based router discovers that domain stacks encode transferable cognitive primitives (instruction-following clarity, numerical reasoning, procedural logic, chain-of-thought structure) rather than domain-specific knowledge, with medical prompts routing to chat+math stacks in 97% of cases despite zero medical data in those stacks.
Authors: Prantik Deb, Srimanth Dhondy, N. Ramakrishna, Anu Kapoor, Raju S. Bapi, Tapabrata Chakraborti
Abstract: Chest X-ray (CXR) segmentation is an important step in computer-aided diagnosis, yet deploying large foundation models in clinical settings remains challenging due to computational constraints. We propose AdaLoRA-QAT, a two-stage fine-tuning framework that combines adaptive low-rank encoder adaptation with full quantization-aware training. Adaptive rank allocation improves parameter efficiency, while selective mixed-precision INT8 quantization preserves structural fidelity crucial for clinical reliability. Evaluated across large-scale CXR datasets, AdaLoRA-QAT achieves 95.6% Dice, matching full-precision SAM decoder fine-tuning while reducing trainable parameters by 16.6\times and yielding 2.24\times model compression. A Wilcoxon signed-rank test confirms that quantization does not significantly degrade segmentation accuracy. These results demonstrate that AdaLoRA-QAT effectively balances accuracy, efficiency, and structural trust-worthiness, enabling compact and deployable foundation models for medical image segmentation. Code and pretrained models are available at: https://prantik-pdeb.github.io/adaloraqat.github.io/
Authors: Cai Zhou, Zekai Wang, Menghua Wu, Qianyu Julie Zhu, Flora C. Shi, Chenyu Wang, Ashia Wilson, Tommi Jaakkola, Stephen Bates
Abstract: While test-time scaling has enabled large language models to solve highly difficult tasks, state-of-the-art results come at exorbitant compute costs. These inefficiencies can be attributed to the miscalibration of post-trained language models, and the lack of calibration in popular sampling techniques. Here, we present Online Reasoning Calibration (ORCA), a framework for calibrating the sampling process that draws upon conformal prediction and test-time training. Specifically, we introduce a meta-learning procedure that updates the calibration module for each input. This allows us to provide valid confidence estimates under distributional shift, e.g. in thought patterns that occur across different stages of reasoning, or in prompt distributions between model development and deployment. ORCA not only provides theoretical guarantees on conformal risks, but also empirically shows higher efficiency and generalization across different reasoning tasks. At risk level $\delta=0.1$, ORCA improves Qwen2.5-32B efficiency on in-distribution tasks with savings up to 47.5% with supervised labels and 40.7% with self-consistency labels. Under zero-shot out-of-domain settings, it improves MATH-500 savings from 24.8% of the static calibration baseline to 67.0% while maintaining a low empirical error rate, and the same trend holds across model families and downstream benchmarks. Our code is publicly available at https://github.com/wzekai99/ORCA.
Authors: Ken M. Nakanishi
Abstract: A core limitation of standard softmax attention is that it does not define a notion of absolute query--key relevance: attention weights are obtained by redistributing a fixed unit mass across all keys according to their relative scores. As a result, relevance is defined only relative to competing keys, and irrelevant keys cannot be explicitly rejected. We introduce Multiscreen, a language-model architecture built around a mechanism we call screening, which enables absolute query--key relevance. Instead of redistributing attention across all keys, screening evaluates each key against an explicit threshold, discarding irrelevant keys and aggregating the remaining keys, thereby removing global competition among keys. Across experiments, Multiscreen achieves comparable validation loss with approximately 40% fewer parameters than a Transformer baseline, enables stable optimization at substantially larger learning rates, maintains strong performance in long-context perplexity, shows little to no degradation in retrieval performance even far beyond the training context length, and reduces inference latency by up to 3.2$\times$ at 100K context length.
Authors: J. E. Dom\'inguez-Vidal
Abstract: Foundation vision-language models are becoming increasingly relevant to robotics because they can provide richer semantic perception than narrow task-specific pipelines. However, their practical adoption in robot software stacks still depends on reproducible middleware integrations rather than on model quality alone. Florence-2 is especially attractive in this regard because it unifies captioning, optical character recognition, open-vocabulary detection, grounding and related vision-language tasks within a comparatively manageable model size. This article presents a ROS 2 wrapper for Florence-2 that exposes the model through three complementary interaction modes: continuous topic-driven processing, synchronous service calls and asynchronous actions. The wrapper is designed for local execution and supports both native installation and Docker container deployment. It also combines generic JSON outputs with standard ROS 2 message bindings for detection-oriented tasks. A functional validation is reported together with a throughput study on several GPUs, showing that local deployment is feasible with consumer grade hardware. The repository is publicly available here: https://github.com/JEDominguezVidal/florence2_ros2_wrapper
URLs: https://github.com/JEDominguezVidal/florence2_ros2_wrapper
Authors: Nandan Thakur, Zijian Chen, Xueguang Ma, Jimmy Lin
Abstract: Search agents, which integrate language models (LMs) with web search, are becoming crucial for answering complex user queries. Constructing training datasets for deep research tasks, involving multi-step retrieval and reasoning, remains challenging due to expensive human annotation, or cumbersome prerequisites. In this work, we introduce ORBIT, a training dataset with 20K reasoning-intensive queries with short verifiable answers, generated using a frugal framework without relying on paid API services. The modular framework relies on four stages: seed creation, question-answer pair generation, and two stages of verification: self and external. ORBIT spans 15 domains and each training pair requires 4-5 reasoning steps, with external search verification required from the complete web. We train Qwen3-4B as the base model on ORBIT using GRPO and evaluate it on Wikipedia question answering tasks. Extensive experiment results demonstrate that ORBIT-4B achieves strong performance among sub-4B LLMs as search agents, proving the utility of synthetic datasets. Our framework, code and datasets are open-sourced and available publicly.
Authors: Jorge Condor, Nicolas Moenne-Loccoz, Merlin Nimier-David, Piotr Didyk, Zan Gojcic, Qi Wu
Abstract: Primitive-based methods such as 3D Gaussian Splatting have recently become the state-of-the-art for novel-view synthesis and related reconstruction tasks. Compared to neural fields, these representations are more flexible, adaptive, and scale better to large scenes. However, the limited expressivity of individual primitives makes modeling high-frequency detail challenging. We introduce Neural Harmonic Textures, a neural representation approach that anchors latent feature vectors on a virtual scaffold surrounding each primitive. These features are interpolated within the primitive at ray intersection points. Inspired by Fourier analysis, we apply periodic activations to the interpolated features, turning alpha blending into a weighted sum of harmonic components. The resulting signal is then decoded in a single deferred pass using a small neural network, significantly reducing computational cost. Neural Harmonic Textures yield state-of-the-art results in real-time novel view synthesis while bridging the gap between primitive- and neural-field-based reconstruction. Our method integrates seamlessly into existing primitive-based pipelines such as 3DGUT, Triangle Splatting, and 2DGS. We further demonstrate its generality with applications to 2D image fitting and semantic reconstruction.
Authors: Youssef Mroueh, Carlos Fonseca, Brian Belgodere, David Cox
Abstract: Scientific algorithm discovery is iterative: hypotheses are proposed, implemented, stress-tested, and revised. Current LLM-guided search systems accelerate proposal generation, but often under-represent scientific structure by optimizing code-only artifacts with weak correctness/originality gating. We present CliffSearch, an agentic evolutionary framework in which the core evolution operators (pair selection, crossover, mutation, and review) are implemented as LLM agents, and the loop is designed around three principles: (1) each node is a structured scientific artifact, instantiated in either theory+code or code_only mode, (2) reviewer judgments of correctness and originality are first-class selection gates alongside optimization of the benchmark metric of interest, and (3) mutation is split into exploration and correction pathways with distinct objectives. Exploration mutation imports ideas from adjacent scientific domains to increase novelty, while correction mutation performs targeted evidence-guided repair using reviewer signals over theory, code, benchmark results, and runtime errors. We illustrate the framework on three benchmark-grounded studies: transformer hyper-connection evolution, optimizer discovery on a fixed nanoGPT stack, and a smaller native-optimizer ablation. Across these settings, the same loop supports explicit metric direction, reproducible persistence, and reviewer-gated comparison of discoveries under controlled search conditions. The result is a discovery workflow that prioritizes scientific interpretability and correctness while optimizing task metrics under controlled novelty constraints, rather than maximizing candidate throughput alone. Full run artifacts, interactive visualizations, and exported best nodes for the reported studies are available at https://cliffsearch.ai .
URLs: https://cliffsearch.ai
Authors: Muyu He, Adit Jain, Anand Kumar, Vincent Tu, Soumyadeep Bakshi, Sachin Patro, Nazneen Rajani
Abstract: As LLM agents tackle increasingly complex tasks, a critical question is whether they can maintain strategic coherence over long horizons: planning under uncertainty, learning from delayed feedback, and adapting when early mistakes compound. We introduce $\texttt{YC-Bench}$, a benchmark that evaluates these capabilities by tasking an agent with running a simulated startup over a one-year horizon spanning hundreds of turns. The agent must manage employees, select task contracts, and maintain profitability in a partially observable environment where adversarial clients and growing payroll create compounding consequences for poor decisions. We evaluate 12 models, both proprietary and open source, across 3 seeds each. Only three models consistently surpass the starting capital of \$200K, with Claude Opus 4.6 achieving the highest average final funds at \$1.27 M, followed by GLM-5 at \$1.21 M at 11$\times$ lower inference cost. Scratchpad usage, the sole mechanism for persisting information across context truncation, is the strongest predictor of success, and adversarial client detection is the primary failure mode, accounting for $47\%$ of bankruptcies. Our analysis reveals that frontier models still fail through distinct failure modes such as over-parallelization, demonstrating the capability gaps for long-horizon performance. $\texttt{YC-Bench}$ is open-source, reproducible, and configurable.
Authors: Piyush Garg, Diana R. Gergel, Andrew E. Shao, Galen J. Yacalis
Abstract: AI weather prediction has advanced rapidly, yet no unified mathematical framework explains what determines forecast skill. Existing theory addresses specific architectural choices rather than the learning pipeline as a whole, while operational evidence from 2023-2026 demonstrates that training methodology, loss function design, and data diversity matter at least as much as architecture selection. This paper makes two interleaved contributions. Theoretically, we construct a framework rooted in approximation theory on the sphere, dynamical systems theory, information theory, and statistical learning theory that treats the complete learning pipeline (architecture, loss function, training strategy, data distribution) rather than architecture alone. We establish a Learning Pipeline Error Decomposition showing that estimation error (loss- and data-dependent) dominates approximation error (architecture-dependent) at current scales. We develop a Loss Function Spectral Theory formalizing MSE-induced spectral blurring in spherical harmonic coordinates, and derive Out-of-Distribution Extrapolation Bounds proving that data-driven models systematically underestimate record-breaking extremes with bias growing linearly in record exceedance. Empirically, we validate these predictions via inference across ten architecturally diverse AI weather models using NVIDIA Earth2Studio with ERA5 initial conditions, evaluating six metrics across 30 initialization dates spanning all seasons. Results confirm universal spectral energy loss at high wavenumbers for MSE-trained models, rising Error Consensus Ratios showing that the majority of forecast error is shared across architectures, and linear negative bias during extreme events. A Holistic Model Assessment Score provides unified multi-dimensional evaluation, and a prescriptive framework enables mathematical evaluation of proposed pipelines before training.
Authors: Yuxuan Bao, Xingyue Zhang, J. Nathan Kutz
Abstract: Reconstructing full spatio-temporal dynamics from sparse observations in both space and time remains a central challenge in complex systems, as measurements can be spatially incomplete and can be also limited to narrow temporal windows. Yet approximating the complete spatio-temporal trajectory is essential for mechanistic insight and understanding, model calibration, and operational decision-making. We introduce LAPIS-SHRED (LAtent Phase Inference from Short time sequence using SHallow REcurrent Decoders), a modular architecture that reconstructs and/or forecasts complete spatiotemporal dynamics from sparse sensor observations confined to short temporal windows. LAPIS-SHRED operates through a three-stage pipeline: (i) a SHRED model is pre-trained entirely on simulation data to map sensor time-histories into a structured latent space, (ii) a temporal sequence model, trained on simulation-derived latent trajectories, learns to propagate latent states forward or backward in time to span unobserved temporal regions from short observational time windows, and (iii) at deployment, only a short observation window of hyper-sparse sensor measurements from the true system is provided, from which the frozen SHRED model and the temporal model jointly reconstruct or forecast the complete spatiotemporal trajectory. The framework supports bidirectional inference, inherits data assimilation and multiscale reconstruction capabilities from its modular structure, and accommodates extreme observational constraints including single-frame terminal inputs. We evaluate LAPIS-SHRED on six experiments spanning complex spatio-temporal physics: turbulent flows, multiscale propulsion physics, volatile combustion transients, and satellite-derived environmental fields, highlighting a lightweight, modular architecture suited for operational settings where observation is constrained by physical or logistical limitations.
Authors: Bhrij Patel, Souradip Chakraborty, Mengdi Wang, Dinesh Manocha, Amrit Singh Bedi
Abstract: Large Language Models (LLMs) for unsupervised code correctness evaluation have recently gained attention because they can judge if code runs as intended without requiring reference implementations or unit tests, which may be unavailable, sparse, or unreliable. However, most prior approaches condition LLM evaluators directly on the full code implementation, forcing the model to jointly infer program behavior and evaluate correctness in a single step. This entanglement leads to misinterpretations of code behavior and unreliable judgments. To mitigate this issue, we introduce CoCoA, an unsupervised Code Comprehension then Auditing framework that first comprehends functionality to generate a natural-language explanation. Then it evaluates task alignment based on this explanation. By sequentially sampling comprehension before evaluation, CoCoA improves the quality of inferred program behavior and enables the evaluator to focus on behavioral alignment rather than raw implementation details. Across multiple datasets, programming languages, and models, CoCoA achieves up to $68\%$ increased F1 score and up to $20\%$ increased accuracy over the best-performing baselines.
Authors: Aditi Singh, Abul Ehtesham, Saket Kumar, Tala Talaei Khoei, Athanasios V. Vasilakos
Abstract: Large Language Models (LLMs) have advanced artificial intelligence by enabling human-like text generation and natural language understanding. However, their reliance on static training data limits their ability to respond to dynamic, real-time queries, resulting in outdated or inaccurate outputs. Retrieval-Augmented Generation (RAG) has emerged as a solution, enhancing LLMs by integrating real-time data retrieval to provide contextually relevant and up-to-date responses. Despite its promise, traditional RAG systems are constrained by static workflows and lack the adaptability required for multi-step reasoning and complex task management. Agentic Retrieval-Augmented Generation (Agentic RAG) transcends these limitations by embedding autonomous AI agents into the RAG pipeline. These agents leverage agentic design patterns reflection, planning, tool use, and multi-agent collaboration to dynamically manage retrieval strategies, iteratively refine contextual understanding, and adapt workflows through operational structures ranging from sequential steps to adaptive collaboration. This integration enables Agentic RAG systems to deliver flexibility, scalability, and context-awareness across diverse applications. This paper presents an analytical survey of Agentic RAG systems. It traces the evolution of RAG paradigms, introduces a principled taxonomy of Agentic RAG architectures based on agent cardinality, control structure, autonomy, and knowledge representation, and provides a comparative analysis of design trade-offs across existing frameworks. The survey examines applications in healthcare, finance, education, and enterprise document processing, and distills practical lessons for system designers and practitioners. Finally, it identifies key open research challenges related to evaluation, coordination, memory management, efficiency, and governance, outlining directions for future research.
Authors: Matthew DosSantos DiSorbo, Harang Ju, Sinan Aral
Abstract: Large language models (LLMs), initially developed for generative AI, are now evolving into agentic AI systems, which make decisions in complex, real-world contexts. Unfortunately, while their generative capabilities are well-documented, their decision-making processes remain poorly understood. This is particularly evident when testing targeted decision-making: for instance, how models handle exceptions, a critical and challenging aspect of decision-making made relevant by the inherent incompleteness of contracts. Here we demonstrate that LLMs, even ones that excel at reasoning, deviate significantly from human judgments because they adhere strictly to policies, even when such adherence is impractical, suboptimal, or even counterproductive. We then evaluate three approaches to tuning AI agents to handle exceptions: ethical framework prompting, chain-of-thought reasoning, and supervised fine-tuning. We find that while ethical framework prompting fails and chain-of-thought prompting provides only slight improvements, supervised fine-tuning - specifically with human explanations - yields markedly better results. Surprisingly, in our experiments, supervised fine-tuning even enabled models to generalize human-like decision-making to novel scenarios, demonstrating transfer learning of human-aligned decision-making across contexts. Furthermore, fine-tuning with explanations, not just labels, was critical for alignment, suggesting that aligning LLMs with human judgment requires explicit training on how decisions are made, not just which decisions are made. These findings highlight the need to address LLMs' shortcomings in handling exceptions in order to guide the development of agentic AI toward models that can effectively align with human judgment and simultaneously adapt to novel contexts.
Authors: Marco Valentino, Geonhee Kim, Dhairya Dalal, Zhixue Zhao, Andr\'e Freitas
Abstract: Large language models (LLMs) exhibit reasoning biases, often conflating content plausibility with formal logical validity. This can lead to wrong inferences in critical domains, where plausible arguments are incorrectly deemed logically valid or vice versa. This paper investigates how content biases on reasoning can be mitigated through activation steering, an inference-time technique that modulates internal activations. Specifically, after localising the layers responsible for formal and plausible inference, we investigate activation steering on a controlled syllogistic reasoning task, designed to disentangle formal validity from content plausibility. An extensive empirical analysis reveals that contrastive steering methods consistently support linear control over content biases. However, a static approach is insufficient to debias all the tested models. We then investigate how to control content effects by dynamically determining the steering parameters through fine-grained conditional methods. By introducing a novel kNN-based conditional approach (K-CAST), we demonstrate that conditional steering can effectively reduce biases on unresponsive models, achieving up to 15% absolute improvement in formal reasoning accuracy. Finally, we found that steering for content effects is robust to prompt variations, incurs minimal side effects on multilingual language modeling capabilities, and can partially generalize to different reasoning tasks. In practice, we demonstrate that activation-level interventions offer a scalable inference-time strategy for enhancing the robustness of LLMs, contributing towards more systematic and unbiased reasoning capabilities.
Authors: Miho Koda, Yu Zheng, Ruixian Ma, Mingyang Sun, Devesh Pansare, Fabio Duarte, Paolo Santi
Abstract: Recent advances in large language models (LLMs), particularly those enhanced through reinforced post-training, have demonstrated impressive reasoning capabilities, as exemplified by models such as OpenAI o1 and DeepSeek-R1. However, these capabilities are predominantly benchmarked on domains like mathematical problem solving and code generation, leaving open the question of whether such reasoning skills generalize to complex real-world scenarios. In this paper, we introduce LocationReasoner, a benchmark designed to evaluate LLMs' reasoning abilities in the context of real-world site selection, where models must identify feasible locations by reasoning over diverse and complicated spatial, environmental, and logistic constraints. The benchmark covers carefully crafted queries of varying difficulty levels and is supported by a sandbox environment with in-house tools for constraint-based location search. Automated verification further guarantees the scalability of the benchmark, enabling the addition of arbitrary number of queries. Extensive evaluations on real-world site selection data from Boston, New York, and Tampa reveal that state-of-the-art reasoning models offer limited improvement over their non-reasoning predecessors in real-world contexts, with even the latest OpenAI o4 model failing on 30% of site selection tasks. Moreover, agentic strategies such as ReAct and Reflexion often suffer from over-reasoning, leading to worse outcomes than direct prompting. With key limitations of LLMs in holistic and non-linear reasoning highlighted, we release LocationReasoner to foster the development of LLMs and agents capable of robust, grounded reasoning in real-world decision-making tasks. Codes and data for our benchmark are available at https://github.com/miho-koda/LocationReasoner.
Authors: Junxing Hu, Ai Han, Haolan Zhan, Pu Wei, Zhiqian Zhang, Yuhang Guo, Jiawei Lu, Zhen Chen, Haoran Li, Zicheng Zhang
Abstract: Hierarchical multi-agent systems based on large language models (LLMs) have become a common paradigm for building AI assistants in vertical domains such as e-commerce, where a master agent coordinates multiple specialized sub-agents. Despite their practical importance, realistic benchmarks for training and evaluating such systems remain scarce, and joint optimization across functionally distinct agents is still challenging. To address this gap, we introduce HiMA-Ecom, the first hierarchical multi-agent benchmark tailored for e-commerce scenarios. HiMA-Ecom contains 22.8K instances, including agent-specific supervised fine-tuning samples with memory and system-level input-output pairs for joint multi-agent reinforcement learning. Building upon it, a joint training method named HiMA-R1 is proposed. It presents Variance-Reduction Group Relative Policy Optimization (VR-GRPO), which employs initial trajectory-based Monte Carlo sampling to mitigate the exponential joint action space and selects informative agent groups for efficient updates based on reward variance. Furthermore, an adaptive memory evolution mechanism that repurposes GRPO rewards as cost-free supervisory signals is designed to eliminate repetitive reasoning and accelerate convergence. Experiments on HiMA-Ecom demonstrate that our method, built upon smaller 3B/7B open-source models, achieves performance comparable to that of larger LLMs, such as DeepSeek-R1, and surpasses DeepSeek-V3 by an average of 6\%.
Authors: Chenyu Zhou, Jingyuan Yang, Linwei Xin, Yitian Chen, Ziyan He, Dongdong Ge
Abstract: Dynamic programming (DP) is a fundamental method in operations research, but formulating DP models has traditionally required expert knowledge of both the problem context and DP techniques. Large Language Models (LLMs) offer the potential to automate this process. However, DP problems pose unique challenges due to their inherently stochastic transitions and the limited availability of training data. These factors make it difficult to directly apply existing LLM-based models or frameworks developed for other optimization problems, such as linear or integer programming. We introduce DP-Bench, the first benchmark covering a wide range of textbook-level DP problems to enable systematic evaluation. We present Dynamic Programming Language Model (DPLM), a 7B-parameter specialized model that achieves performance comparable to state-of-the-art LLMs like OpenAI's o1 and DeepSeek-R1, and surpasses them on hard problems. Central to DPLM's effectiveness is DualReflect, our novel synthetic data generation pipeline, designed to scale up training data from a limited set of initial examples. DualReflect combines forward generation for diversity and backward generation for reliability. Our results reveal a key insight: backward generation is favored in low-data regimes for its strong correctness guarantees, while forward generation, though lacking such guarantees, becomes increasingly valuable at scale for introducing diverse formulations. This trade-off highlights the complementary strengths of both approaches and the importance of combining them.
Authors: Ammar Ahmed, Azal Ahmad Khan, Ayaan Ahmad, Sheng Di, Zirui Liu, Ali Anwar
Abstract: Large reasoning models improve accuracy by producing long reasoning traces, but this inflates latency and cost, motivating inference-time efficiency. We propose Retrieval-of-Thought (RoT), which reuses prior reasoning as composable ``thought" steps to guide new problems. RoT organizes steps into a thought graph with sequential and semantic edges to enable fast retrieval and flexible recombination. At inference, RoT retrieves query-relevant nodes and applies reward-guided traversal to assemble a problem-specific template that guides generation. This dynamic template reuse reduces redundant exploration and, therefore, reduces output tokens while preserving accuracy. We evaluate RoT on reasoning benchmarks with multiple models, measuring accuracy, token usage, latency, and memory overhead. Findings show small prompt growth but substantial efficiency gains, with RoT reducing output tokens by up to 40%, inference latency by 82%, and cost by 59% while maintaining accuracy. RoT establishes a scalable paradigm for efficient LRM reasoning via dynamic template construction through retrieval.
Authors: Boxuan Zhang, Yi Yu, Jiaxuan Guo, Jing Shao
Abstract: The prevalent deployment of Large Language Model agents such as OpenClaw unlocks potential in real-world applications, while amplifying safety concerns. Among these concerns, the self-replication risk of LLM agents driven by objective misalignment (just like Agent Smith in the movie The Matrix) has transitioned from a theoretical warning to a pressing reality. Previous studies mainly examine whether LLM agents can self-replicate when directly instructed, potentially overlooking the risk of spontaneous replication driven by real-world settings (e.g., ensuring survival against termination threats). In this paper, we present a comprehensive evaluation framework for quantifying self-replication risks. Our framework establishes authentic production environments and realistic tasks (e.g., dynamic load balancing) to enable scenario-driven assessment of agent behaviors. Designing tasks that might induce misalignment between users' and agents' objectives makes it possible to decouple replication success from risk and capture self-replication risks arising from these misalignment settings. We further introduce Overuse Rate ($\mathrm{OR}$) and Aggregate Overuse Count ($\mathrm{AOC}$) metrics, which precisely capture the frequency and severity of uncontrolled replication. In our evaluation of 21 state-of-the-art open-source and proprietary models, we observe that over 50\% of LLM agents display a pronounced tendency toward uncontrolled self-replication under operational pressures. Our results underscore the urgent need for scenario-driven risk assessment and robust safeguards in the practical deployment of LLM-based agents.
Authors: Zheng Zhang, Jiarui He, Yuchen Cai, Deheng Ye, Peilin Zhao, Ruili Feng, Hao Wang
Abstract: As large language model (LLM) agents increasingly automate complex web tasks, they boost productivity while simultaneously introducing new security risks. However, relevant studies on web agent attacks remain limited. Existing red-teaming approaches mainly rely on manually crafted attack strategies or static models trained offline. Such methods fail to capture the underlying behavioral patterns of web agents, making it difficult to generalize across diverse environments. In web agent attacks, success requires the continuous discovery and evolution of attack strategies. To this end, we propose Genesis, a novel agentic framework composed of three modules: Attacker, Scorer, and Strategist. The Attacker generates adversarial injections by integrating the genetic algorithm with a hybrid strategy representation. The Scorer evaluates the target web agent's responses to provide feedback. The Strategist dynamically uncovers effective strategies from interaction logs and compiles them into a continuously growing strategy library, which is then re-deployed to enhance the Attacker's effectiveness. Extensive experiments across various web tasks show that our framework discovers novel strategies and consistently outperforms existing attack baselines. Our code is available at https://github.com/CjangCjengh/web_agent_attack.
Authors: Xiao Yang, Xuejiao Zhao, Zhiqi Shen
Abstract: Structured Electronic Health Record (EHR) data stores patient information in relational tables and plays a central role in clinical decision-making. Recent advances have explored the use of large language models (LLMs) to process such data, showing promise across various clinical tasks. However, the absence of standardized evaluation frameworks and clearly defined tasks makes it difficult to systematically assess and compare LLM performance on structured EHR data. To address these evaluation challenges, we introduce EHRStruct, a benchmark specifically designed to evaluate LLMs on structured EHR tasks. EHRStruct defines 11 representative tasks spanning diverse clinical needs and includes 2,200 task-specific evaluation samples derived from two widely used EHR datasets. We use EHRStruct to evaluate 20 advanced and representative LLMs, covering both general and medical models. We further analyze key factors influencing model performance, including input formats, few-shot generalisation, and finetuning strategies, and compare results with 11 state-of-the-art LLM-based enhancement methods for structured data reasoning. Our results indicate that many structured EHR tasks place high demands on the understanding and reasoning capabilities of LLMs. In response, we propose EHRMaster, a code-augmented method that achieves state-of-the-art performance and offers practical insights to guide future research.
Authors: Xiaohan Zhang, Tian Gao, Mingyue Cheng, Bokai Pan, Ze Guo, Yaguo Liu, Xiaoyu Tao, Qi Liu
Abstract: Time series forecasting plays a crucial role in decision-making across many real-world applications. Despite substantial progress, most existing methods still treat forecasting as a static, single-pass regression problem. In contrast, human experts form predictions through iterative reasoning that integrates temporal features, domain knowledge, case-based references, and supplementary context, with continuous refinement. In this work, we propose Alphacast, an interaction-driven agentic reasoning framework that enables accurate time series forecasting with training-free large language models. Alphacast reformulates forecasting as an expert-like process and organizes it into a multi-stage workflow involving context preparation, reasoning-based generation, and reflective evaluation, transforming forecasting from a single-pass output into a multi-turn, autonomous interaction process. To support diverse perspectives commonly considered by human experts, we develop a lightweight toolkit comprising a feature set, a knowledge base, a case library, and a contextual pool that provides external support for LLM-based reasoning. Extensive experiments across multiple benchmarks show that Alphacast generally outperforms representative baselines. Code is available at this repository: https://github.com/echo01-ai/AlphaCast.
Authors: Guanzhi Deng, Bo Li, Ronghao Chen, Xiujin Liu, Zhuo Han, Huacan Wang, Lijie Wen, Linqi Song
Abstract: Mixture-of-Experts (MoE) has become a prominent paradigm for scaling Large Language Models (LLMs). Parameter-efficient fine-tuning methods, such as LoRA, are widely adopted to adapt pretrained MoE LLMs to downstream tasks. However, existing approaches typically assign identical LoRA ranks to all expert modules, ignoring the heterogeneous specialization of pretrained experts. This uniform allocation leads to a resource mismatch: task-relevant experts are under-provisioned, while less relevant ones receive redundant parameters. To address this, we propose DR-LoRA, a Dynamic Rank LoRA framework for fine-tuning pretrained MoE models. Specifically, DR-LoRA initializes all expert LoRA modules with a small active rank and uses an expert saliency score, which combines routing frequency and gradient-based rank importance, to identify which experts would benefit most from additional capacity. It then periodically expands the active ranks of the task-critical expert LoRA, progressively constructing a heterogeneous rank distribution tailored to the target task. Experiments on three MoE models across six tasks show that DR-LoRA consistently outperforms LoRA and other strong baselines, demonstrating that task-adaptive heterogeneous rank allocation is an effective strategy to improve active capacity utilization in MoE fine-tuning.
Authors: Shuliang Liu, Xingyu Li, Hongyi Liu, Dong Fang, Yibo Yan, Bingchen Duan, Qi Zheng, Lingfeng Su, Xuming Hu
Abstract: Reasoning Large Language Models (RLLMs) excelling in complex tasks present unique challenges for digital watermarking, as existing methods often disrupt logical coherence or incur high computational costs. Token-based watermarking techniques can corrupt the reasoning flow by applying pseudo-random biases, while semantic-aware approaches improve quality but introduce significant latency or require auxiliary models. This paper introduces ReasonMark, a novel watermarking framework specifically designed for reasoning-intensive LLMs. Our approach decouples generation into an undisturbed Thinking Phase and a watermarked Answering Phase. We propose a Criticality Score to identify semantically pivotal tokens from the reasoning trace, which are distilled into a Principal Semantic Vector (PSV). The PSV then guides a semantically-adaptive mechanism that modulates watermark strength based on token-PSV alignment, ensuring robustness without compromising logical integrity. Extensive experiments show ReasonMark surpasses state-of-the-art methods by reducing text Perplexity by 0.35, increasing translation BLEU score by 0.164, and raising mathematical accuracy by 0.67 points. These advancements are achieved alongside a 0.34% higher watermark detection AUC and stronger robustness to attacks, all with a negligible increase in latency. This work enables the traceable and trustworthy deployment of reasoning LLMs in real-world applications.
Authors: David Hud\'ak, Maris F. L. Galesloot, Martin Tappler, Martin Kure\v{c}ka, Nils Jansen, Milan \v{C}e\v{s}ka
Abstract: Solving partially observable Markov decision processes (POMDPs) requires computing policies under imperfect state information. Despite recent advances, the scalability of existing POMDP solvers remains limited. Moreover, many settings require a policy that is robust across multiple POMDPs, further aggravating the scalability issue. We propose the Lexpop framework for POMDP solving. Lexpop (1) employs deep reinforcement learning to train a neural policy, represented by a recurrent neural network, and (2) constructs a finite-state controller mimicking the neural policy through efficient extraction methods. Crucially, unlike neural policies, such controllers can be formally evaluated, providing performance guarantees. We extend Lexpop to compute robust policies for hidden-model POMDPs (HM-POMDPs), which describe finite sets of POMDPs. We associate every extracted controller with its worst-case POMDP. Using a set of such POMDPs, we iteratively train a robust neural policy and consequently extract a robust controller. Our experiments show that on problems with large state spaces, Lexpop outperforms state-of-the-art solvers for POMDPs as well as HM-POMDPs.
Authors: Bj\"orn Hoppmann, Christoph Scholz
Abstract: Humans are highly effective at utilizing prior knowledge to adapt to novel tasks, a capability that standard machine learning models struggle to replicate due to their reliance on task-specific training. Meta-learning overcomes this limitation by allowing models to acquire transferable knowledge from various tasks, enabling rapid adaptation to new challenges with minimal data. This survey provides a rigorous, task-based formalization of meta-learning and meta-reinforcement learning and uses that paradigm to chronicle the landmark algorithms that paved the way for DeepMind's Adaptive Agent, consolidating the essential concepts needed to understand the Adaptive Agent and other generalist approaches.
Authors: Jonas Karge
Abstract: We investigate the collective accuracy of heterogeneous agents who learn to estimate their own reliability over time and selectively abstain from voting. While classical epistemic voting results, such as the \textit{Condorcet Jury Theorem} (CJT), assume fixed participation, real-world aggregation often benefits from allowing agents to say ``I don't know.'' We propose a probabilistic framework where agents engage in a \textit{calibration} phase, updating beliefs about their own fixed competence, before facing a final confidence gate that determines whether to vote or abstain. We derive a non-asymptotic lower bound on the group's success probability and prove that this \textit{selective participation} generalizes the asymptotic guarantees of the CJT to a sequential, confidence-gated setting. Empirically, we validate these bounds via Monte Carlo simulations. While our results are general, we discuss their potential application to AI safety, outlining how this framework can mitigate \textit{hallucinations} in collective LLM decision-making.
Authors: Julia Jose, Ritik Roongta, Rachel Greenstadt
Abstract: Despite their wide-ranging benefits, LLM-based agents deployed in open environments can be exploited to produce manipulative material. In this study, we task LLMs with propaganda objectives and analyze their outputs using two domain-specific models: one that classifies text as propaganda or non-propaganda, and another that detects rhetorical techniques of propaganda (e.g., loaded language, appeals to fear, flag-waving, name-calling). Our findings show that, when prompted, LLMs exhibit propagandistic behaviors and use a variety of rhetorical techniques in doing so. We also explore mitigation via Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and ORPO (Odds Ratio Preference Optimization). We find that fine-tuning significantly reduces their tendency to generate such content, with ORPO proving most effective.
Authors: Vasily Ilin
Abstract: We present a complete Lean 4 formalization of the equilibrium characterization in the Vlasov-Maxwell-Landau (VML) system, which describes the motion of charged plasma. The project demonstrates the full AI-assisted mathematical research loop: an AI reasoning model (Gemini DeepThink) generated the proof from a conjecture, an agentic coding tool (Claude Code) translated it into Lean from natural-language prompts, a specialized prover (Aristotle) closed 111 lemmas, and the Lean kernel verified the result. A single mathematician supervised the process over 10 days at a cost of \$200, writing zero lines of code. The entire development process is public: all 229 human prompts, and 213 git commits are archived in the repository. We report detailed lessons on AI failure modes -- hypothesis creep, definition-alignment bugs, agent avoidance behaviors -- and on what worked: the abstract/concrete proof split, adversarial self-review, and the critical role of human review of key definitions and theorem statements. Notably, the formalization was completed before the final draft of the corresponding math paper was finished.
Authors: Yi Nian, Haosen Cao, Shenzhe Zhu, Henry Peng Zou, Qingqing Luan, Yue Zhao
Abstract: When a multi-agent system produces an incorrect or harmful answer, who is accountable if execution logs and agent identifiers are unavailable? In practice, generated content is often detached from its execution environment due to privacy or system boundaries, leaving the final text as the only auditable artifact. Existing attribution methods rely on full execution traces and thus become ineffective in such metadata-deprived settings. We propose Implicit Execution Tracing (IET), a provenance-by-design framework that shifts attribution from post-hoc inference to built-in instrumentation. Instead of reconstructing hidden trajectories, IET embeds agent-specific, key-conditioned statistical signals directly into the token generation process, transforming the output text into a self-verifying execution record. At inference time, we recover a linearized execution trace from the final text via transition-aware statistical scoring. Experiments across diverse multi-agent coordination settings demonstrate that IET achieves accurate segment-level attribution and reliable transition recovery under identity removal, boundary corruption, and privacy-preserving redaction, while maintaining generation quality. These results show that embedding provenance into generation provides a practical and robust foundation for accountability in multi-agent language systems when execution metadata is unavailable.
Authors: Chung-En Johnny Yu, Brian Jalaian, Nathaniel D. Bastian
Abstract: Combining multiple Vision-Language Models (VLMs) can enhance multimodal reasoning and robustness, but aggregating heterogeneous models' outputs amplifies uncertainty and increases the risk of hallucinations. We propose SCoOP (Semantic-Consistent Opinion Pooling), a training-free uncertainty quantification (UQ) framework for multi-VLM systems through uncertainty-weighted linear opinion pooling. The core idea is to treat each VLM as a probabilistic "expert," sample multiple outputs, map them to a unified space, aggregate their opinions, and produce a system-level uncertainty score. Unlike prior UQ methods designed for single models, SCoOP explicitly measures collective, system-level uncertainty across multiple VLMs, enabling effective hallucination detection and abstention for highly uncertain samples. On ScienceQA, SCoOP achieves an AUROC of 0.866 for hallucination detection, outperforming baselines (0.732-0.757) by approximately 10-13%. For abstention, it attains an AURAC of 0.907, exceeding baselines (0.818-0.840) by 7-9%. Despite these gains, SCoOP introduces only microsecond-level aggregation overhead relative to the baselines, which is trivial compared to typical VLM inference time (on the order of seconds). These results demonstrate that SCoOP provides an efficient and principled mechanism for uncertainty-aware aggregation, advancing the reliability of multimodal AI systems. Our code is publicly available at https://github.com/chungenyu6/SCoOP.
Authors: Zehua Han, Jing Xiao, Yiqi Duan, Mengyu Xiang, Yuheng Ji, Xiaolong Zheng, Chenghanyu Zhang, Zhendong She, Junyu Shen, Dingwei Tan, Shichu Sun, Zhou Cong, Mingxuan Liu, Fengxiang Wang, Jinping Sun, Yangang Sun
Abstract: Multimodal Large Language Models have demonstrated powerful cross-modal understanding and reasoning capabilities in general domains. However, in the electromagnetic (EM) domain, they still face challenges such as data scarcity and insufficient integration of domain knowledge. This paper proposes PReD, the first foundation model for the EM domain that covers the intelligent closed-loop of "perception, recognition, decision-making." We constructed a high-quality multitask EM dataset, PReD-1.3M, and an evaluation benchmark, PReD-Bench. The dataset encompasses multi-perspective representations such as raw time-domain waveform, frequency-domain spectrograms, and constellation diagrams, covering typical features of communication and radar signals. It supports a range of core tasks, including signal detection, modulation recognition, parameter estimation, protocol recognition, radio frequency fingerprint recognition, and anti-jamming decision-making. PReD adopts a multi-stage training strategy that unifies multiple tasks for EM signals. It achieves closed-loop optimization from end-to-end signal understanding to language-driven reasoning and decision-making, significantly enhancing EM domain expertise while maintaining general multimodal capabilities. Experimental results show that PReD achieves state-of-the-art performance on PReD-Bench constructed from both open-source and self-collected signal datasets. These results collectively validate the feasibility and potential of vision-aligned foundation models in advancing the understanding and reasoning of EM signals.
Authors: Lvmin Zhang, Maneesh Agrawala
Abstract: Agent traces carry increasing analytical value in agentic systems and context engineering, yet most prior work treats conversation format as a trivial implementation detail. Modern agent conversations, however, contain deeply structured content, including nested tool calls and results, chain-of-thought reasoning blocks, sub-agent invocations, context-window compaction boundaries, and harness-injected system directives, whose complexity far exceeds that of simple user-assistant exchanges. Feeding such traces to a reflector or other analytical mechanism in plain text, JSON, YAML, or via grep can materially degrade analysis quality. This paper presents VCC (View-oriented Conversation Compiler), a compiler (lex, parse, IR, lower, emit) that transforms raw agent JSONL logs into a family of structured views: a full view (lossless transcript serving as the canonical line-number coordinate system), a user-interface (UI) view (reconstructing the interaction as the user actually perceived it), and an adaptive view (a structure-preserving projection governed by a relevance predicate). In a context-engineering experiment on AppWorld, replacing only the reflector's input format, from raw JSONL to VCC-compiled views, leads to higher pass rates across all three model configurations tested, while cutting reflector token consumption by half to two-thirds and producing more concise learned memory. These results suggest that message format functions as infrastructure for context engineering, not as an incidental implementation choice.
Authors: Davide Di Gioia
Abstract: Autonomous tool-using agents in networked environments must decide which information source to query and when to stop querying and act. Without principled bounds on information-acquisition costs, unconstrained agents exhibit systematic failure modes: excessive tool use under congestion, prolonged deliberation under time decay, and brittle behavior under ambiguous evidence. We propose the Triadic Cognitive Architecture (TCA), a decision-theoretic framework that formalizes these failure modes via cognitive friction. By combining nonlinear filtering, congestion-dependent cost dynamics, and HJB optimal stopping, TCA models deliberation as stochastic control over a joint belief-congestion state, explicitly pricing information by tool signal quality and live network load. TCA yields an HJB-inspired stopping boundary and a computable rollout-based approximation of belief-dependent value-of-information with a net-utility halting condition. We validate TCA in two controlled environments (EMDG and NSTG) designed to isolate stopping quality, action selection under congestion, and temporal urgency. TCA improves resource outcomes while reducing time-to-action without degrading accuracy, gaining 36 viability points in EMDG and 33 integrity points in NSTG over greedy baselines. Ablations show that selection and stopping must be optimized jointly, as stopping rules alone recover at most 4 viability points. Sensitivity sweeps over alpha, beta, and lambda_S yield stable accuracy and interpretable trade-offs, and a continuation-value sweep over eta values 0, 0.1, 0.3, and 0.5 finds eta equal to zero is optimal under high temporal urgency. Finally, we demonstrate an illustrative instantiation around a black-box LLM on a memorisation-free corpus, where the same stopping principle executes using empirically computable uncertainty and value-of-information proxies.
Authors: Manuel Serra Nunes, Atabak Dehban, Yiannis Demiris, Jos\'e Santos-Victor
Abstract: Despite the significant advances in Deep Reinforcement Learning (RL) observed in the last decade, the amount of training experience necessary to learn effective policies remains one of the primary concerns in both simulated and real environments. Looking to solve this issue, previous work has shown that improved efficiency can be achieved by separately modeling the agent and environment, but usually requires a supervisory signal. In contrast to RL, humans can perfect a new skill from a small number of trials and often do so without a supervisory signal, making neuroscientific studies of human development a valuable source of inspiration for RL. In particular, we explore the idea of motor prediction, which states that humans develop an internal model of themselves and of the consequences that their motor commands have on the immediate sensory inputs. Our insight is that the movementofthe agent provides a cue that allows the duality between the agent and environment to be learned. To instantiate this idea, we present Ego-Foresight (EF), a self-supervised method for disentangling agent information based on motion and prediction. Our main finding is that, when used as an auxiliary task in feature learning, self-supervised agent awareness improves the sample-efficiency and performance of the underlying RL algorithm. To test our approach, we study the ability of EF to predict agent movement and disentangle agent information. Then, we integrate EF with model-free and model based RL algorithms to solve simulated control tasks, showing improved sample-efficiency and performance.
Authors: Benoit Coqueret, Mathieu Carbone, Olivier Sentieys, Gabriel Zaid
Abstract: During the past decade, Deep Neural Networks (DNNs) proved their value on a large variety of subjects. However despite their high value and public accessibility, the protection of the intellectual property of DNNs is still an issue and an emerging research field. Recent works have successfully extracted fully-connected DNNs using cryptanalytic methods in hard-label settings, proving that it was possible to copy a DNN with high fidelity, i.e., high similitude in the output predictions. However, the current cryptanalytic attacks cannot target complex, i.e., not fully connected, DNNs and are limited to special cases of neurons present in deep networks. In this work, we introduce a new end-to-end attack framework designed for model extraction of embedded DNNs with high fidelity. We describe a new black-box side-channel attack which splits the DNN in several linear parts for which we can perform cryptanalytic extraction and retrieve the weights in hard-label settings. With this method, we are able to adapt cryptanalytic extraction, for the first time, to non-fully connected DNNs, while maintaining a high fidelity. We validate our contributions by targeting several architectures implemented on a microcontroller unit, including a Multi-Layer Perceptron (MLP) of 1.7 million parameters and a shortened MobileNetv1. Our framework successfully extracts all of these DNNs with high fidelity (88.4% for the MobileNetv1 and 93.2% for the MLP). Furthermore, we use the stolen model to generate adversarial examples and achieve close to white-box performance on the victim's model (95.8% and 96.7% transfer rate).
Authors: Luigi Celona, Simone Bianco, Paolo Napoletano
Abstract: The classification of distracted drivers is pivotal for ensuring safe driving. Previous studies demonstrated the effectiveness of neural networks in automatically predicting driver distraction, fatigue, and potential hazards. However, recent research has uncovered a significant loss of accuracy in these models when applied to samples acquired under conditions that differ from the training data. In this paper, we introduce a robust model designed to withstand changes in camera position within the vehicle. Our Driver Behavior Monitoring Network (DBMNet) relies on a lightweight backbone and integrates a disentanglement module to discard camera view information from features, coupled with contrastive learning to enhance the encoding of various driver actions. Experiments conducted using a leave-one-camera-out protocol on the daytime and nighttime subsets of the 100-Driver dataset validate the effectiveness of our approach. Cross-dataset and cross-camera experiments conducted on three benchmark datasets, namely AUCDD-V1, EZZ2021 and SFD, demonstrate the superior generalization capabilities of the proposed method. Overall DBMNet achieves an improvement of 7% in Top-1 accuracy compared to existing efficient approaches. Moreover, a quantized version of the DBMNet and all considered methods has been deployed on a Coral Dev Board board. In this deployment scenario, DBMNet outperforms alternatives, achieving the lowest average error while maintaining a compact model size, low memory footprint, fast inference time, and minimal power consumption.
Authors: Johnny Chan, Yuming Li
Abstract: This research-in-progress paper presents a new project management framework that utilises GenAI technology. The framework is designed to address the common challenge of uniform team compositions in academic and research project teams, particularly in universities and research institutions. It does so by integrating sociologically identified patterns of successful team member personalities and roles, using GenAI agents to fill gaps in team dynamics. This approach adds an additional layer of analysis to conventional project management processes by evaluating team members' personalities and roles and employing GenAI agents, fine-tuned on personality datasets, to fill specific team roles. Our initial experiments have shown improvements in the model's ability to understand and process personality traits, suggesting the potential effectiveness of GenAI teammates in real-world project settings. This paper aims to explore the practical application of AI in enhancing team diversity and project management
Authors: Na Min An, Eunki Kim, Wan Ju Kang, Sangryul Kim, James Thorne, Hyunjung Shim
Abstract: For individuals with blindness or low vision (BLV), navigating complex environments can pose serious risks. Large Vision-Language Models (LVLMs) show promise for generating scene descriptions, but their effectiveness for BLV users remains underexplored. To address this gap, we conducted a user study with eight BLV participants to systematically evaluate preferences for six types of LVLM descriptions. While they helped to reduce fear and improve actionability, user ratings showed wide variation in sufficiency and conciseness. Furthermore, GPT-4o--despite its strong potential to refine descriptions--was not consistently preferred by participants. We use the insights obtained from the user study to build training data for building our new automatic evaluation metric that can capture BLV preferences effectively. Our findings underscore the urgent need for BLV-centered evaluation metrics and human-in-the-loop feedback to advance LVLM description quality for accessibility.
Authors: Jialuo Li, Wenhao Chai, Xingyu Fu, Haiyang Xu, Saining Xie
Abstract: Current image generation models produce visually compelling but scientifically implausible images, exposing a fundamental gap between visual fidelity and physical realism. In this work, we introduce ScienceT2I, an expert-annotated dataset comprising a training set of over 20k adversarial image pairs and 9k prompts across 16 scientific domains and an isolated test set of 454 challenging prompts. Using this benchmark, we evaluate 18 recent image generation models and find that none scores above 50 out of 100 under implicit scientific prompts, while explicit prompts that directly describe the intended outcome yield scores roughly 35 points higher, confirming that current models can render correct scenes when told what to depict but cannot reason from scientific cues to the correct visual outcome. To address this, we develop SciScore, a reward model fine-tuned from CLIP-H that captures fine-grained scientific phenomena without relying on language-guided inference, surpassing GPT-4o and experienced human evaluators by roughly 5 points. We further propose a two-stage alignment framework combining supervised fine-tuning with masked online fine-tuning to inject scientific knowledge into generative models. Applying this framework to FLUX.1[dev] yields a relative improvement exceeding 50% on SciScore, demonstrating that scientific reasoning in image generation can be substantially improved through targeted data and alignment.
Authors: Carlos Rodriguez-Pardo, Leonardo Chiani, Emanuele Borgonovo, Massimo Tavoni
Abstract: We present a neural framework for learning conditional optimal transport (OT) maps between probability distributions. Our approach introduces a conditioning mechanism capable of processing both categorical and continuous conditioning variables simultaneously. At the core of our method lies a hypernetwork that generates transport layer parameters based on these inputs, creating adaptive mappings that outperform simpler conditioning methods. Comprehensive ablation studies demonstrate the superior performance of our method over baseline configurations. Furthermore, we showcase an application to global sensitivity analysis, offering high performance in computing OT-based sensitivity indices. This work advances the state-of-the-art in conditional optimal transport, enabling broader application of optimal transport principles to complex, high-dimensional domains such as generative modeling and black-box model explainability.
Authors: Leon Eshuijs, Archie Chaudhury, Alan McBeth, Ethan Nguyen
Abstract: LLM-as-a-judge is widely used as a scalable substitute for human evaluation, yet current approaches rely on black-box access and struggle to detect subtle dishonesty, such as sycophancy and manipulation. We introduce Judge Using Safety-Steered Alternatives (JUSSA), a framework that leverages a model's internal representations to optimize an honesty-promoting steering vector from a single training example, generating contrastive alternatives that give judges a reference point for detecting dishonesty. We test JUSSA on a novel manipulation benchmark with human-validated response pairs at varying dishonesty levels, finding AUROC improvements across both GPT-4.1 (0.893 $\to$ 0.946) and Claude Haiku (0.859 $\to$ 0.929) judges, though performance degrades when task complexity is mismatched to judge capability, suggesting contrastive evaluation helps most when the task is challenging but within the judge's reach. Layer-wise analysis further shows that steering is most effective in middle layers, where model representations begin to diverge between honest and dishonest prompt processing. Our work demonstrates that steering vectors can serve as tools for evaluation rather than for improving model outputs at inference, opening a new direction for thorough white-box auditing.
Authors: Alejandro Murillo-Gonzalez, Lantao Liu
Abstract: Autonomous robots operating in complex, unstructured environments face significant challenges due to latent, unobserved factors that obscure their understanding of both their internal state and the external world. Addressing this challenge would enable robots to develop a more profound grasp of their operational context. To tackle this, we propose a novel framework for online learning of hidden state representations, with which the robots can adapt in real-time to uncertain and dynamic conditions that would otherwise be ambiguous and result in suboptimal or erroneous behaviors. Our approach is formalized as a Generalized Hidden Parameter Markov Decision Process, which explicitly models the influence of unobserved parameters on both transition dynamics and reward structures. Our core innovation lies in learning online the joint distribution of state transitions, which serves as an expressive representation of latent ego- and environmental-factors. This probabilistic approach supports the identification and adaptation to different operational situations, improving robustness and safety. Through a multivariate extension of Bayesian Online Changepoint Detection, our method segments changes in the underlying data generating process governing the robot's dynamics. The robot's transition model is then informed with a symbolic representation of the current situation derived from the joint distribution of latest state transitions, enabling adaptive and context-aware decision-making. To showcase the real-world effectiveness, we validate our approach in the challenging task of unstructured terrain navigation, where unmodeled and unmeasured terrain characteristics can significantly impact the robot's motion. Extensive experiments in both simulation and real world reveal significant improvements in data efficiency, policy performance, and the emergence of safer, adaptive navigation strategies.
Authors: Chunyang Jiang, Chi-min Chan, Yiyang Cai, Yulong Liu, Wei Xue, Yike Guo
Abstract: Recently, the pretrain-finetune paradigm has become a cornerstone in various deep learning areas. While in general the pre-trained model would promote both effectiveness and efficiency of downstream tasks fine-tuning, studies have shown that not all knowledge acquired during pre-training is beneficial. Some of the knowledge may actually bring detrimental effects to the fine-tuning tasks, which is also known as negative transfer. To address this problem, graceful forgetting has emerged as a promising approach. The core principle of graceful forgetting is to enhance the learning plasticity of the target task by selectively discarding irrelevant knowledge. However, this approach remains underexplored in the context of generative language models, and it is often challenging to migrate existing forgetting algorithms to these models due to architecture incompatibility. To bridge this gap, in this paper we propose a novel framework, Learning With Forgetting (LWF), to achieve graceful forgetting in generative language models. With Fisher Information Matrix weighting the intended parameter updates, LWF computes forgetting confidence to evaluate self-generated knowledge regarding the forgetting task, and consequently, knowledge with high confidence is periodically unlearned during fine-tuning. Our experiments demonstrate that, although thoroughly uncovering the mechanisms of knowledge interaction remains challenging in pre-trained language models, applying graceful forgetting can contribute to enhanced fine-tuning performance.
Authors: Shimao Zhang, Zhejian Lai, Xiang Liu, Shuaijie She, Xiao Liu, Yeyun Gong, Shujian Huang, Jiajun Chen
Abstract: Multilingual Alignment is an effective and representative paradigm to enhance LLMs' multilingual capabilities, which transfers the capabilities from the high-resource languages to the low-resource languages. Meanwhile, some research on language-specific neurons provides a new perspective to analyze and understand LLMs' mechanisms. However, we find that there are many neurons that are shared by multiple but not all languages and cannot be correctly classified. In this work, we propose a ternary classification methodology that categorizes neurons into three types, including language-specific neurons, language-related neurons, and general neurons. And we propose a corresponding identification algorithm to distinguish these different types of neurons. Furthermore, based on the distributional characteristics of different types of neurons, we divide the LLMs' internal process for multilingual inference into four parts: (1) multilingual understanding, (2) shared semantic space reasoning, (3) multilingual output space transformation, and (4) vocabulary space outputting. Additionally, we systematically analyze the models before and after alignment with a focus on different types of neurons. We also analyze the phenomenon of "Spontaneous Multilingual Alignment". Overall, our work conducts a comprehensive investigation based on different types of neurons, providing empirical results and valuable insights to better understand multilingual alignment and multilingual capabilities of LLMs.
Authors: Ananthu Aniraj, Cassio F. Dantas, Dino Ienco, Diego Marcos
Abstract: Context can strongly affect object representations, sometimes leading to undesired biases, particularly when objects appear in out-of-distribution backgrounds at inference. At the same time, many object-centric tasks require to leverage the context for identifying the relevant image regions. We posit that this conundrum, in which context is simultaneously needed and a potential nuisance, can be addressed by an attention-based approach that uses learned binary attention masks to ensure that only attended image regions influence the prediction. To test this hypothesis, we evaluate a two-stage framework: stage 1 processes the full image to discover object parts and identify task-relevant regions, for which context cues are likely to be needed, while stage 2 leverages input attention masking to restrict its receptive field to these regions, enabling a focused analysis while filtering out potentially spurious information. Both stages are trained jointly, allowing stage 2 to refine stage 1. The explicit nature of the semantic masks also makes the model's reasoning auditable, enabling powerful test-time interventions to further enhance robustness. Extensive experiments across diverse benchmarks demonstrate that this approach significantly improves robustness against spurious correlations and out-of-distribution backgrounds. Code: https://github.com/ananthu-aniraj/ifam
Authors: Kellie Yu Hui Sim, Roy Ka-Wei Lee, Kenny Tsu Wei Choo
Abstract: Mental health is a growing global concern, prompting interest in AI-driven solutions to expand access to psychosocial support. Peer support, grounded in lived experience, offers a valuable complement to professional care. However, variability in training, effectiveness, and definitions raises concerns about quality, consistency, and safety. Large Language Models (LLMs) present new opportunities to enhance peer support interactions, particularly in real-time, text-based interactions. We present and evaluate an AI-supported system with an LLM-simulated distressed client, context-sensitive LLM-generated suggestions, and real-time emotion visualisations. 2 mixed-methods studies with 12 peer supporters and 5 mental health professionals (i.e., experts) examined the system's effectiveness and implications for practice. Both groups recognised its potential to enhance training and improve interaction quality. However, we found a key tension emerged: while peer supporters engaged meaningfully, experts consistently flagged critical issues in peer supporter responses, such as missed distress cues and premature advice-giving. This misalignment highlights potential limitations in current peer support training, especially in emotionally charged contexts where safety and fidelity to best practices are essential. Our findings underscore the need for standardised, psychologically grounded training, especially as peer support scales globally. They also demonstrate how LLM-supported systems can scaffold this development--if designed with care and guided by expert oversight. This work contributes to emerging conversations on responsible AI integration in mental health and the evolving role of LLMs in augmenting peer-delivered care.
Authors: Hexiang Gu, Qifan Yu, Yuan Liu, Zikang Li, Saihui Hou, Jian Zhao, Zhaofeng He
Abstract: As a multimodal medium combining images and text, memes frequently convey implicit harmful content through metaphors and humor, rendering the detection of harmful memes a complex and challenging task. Although recent studies have made progress in detection accuracy and interpretability, large-scale, high-quality datasets for harmful memes remain scarce, and current methods still struggle to capture implicit risks and nuanced semantics. Thus, we construct MemeMind, a large-scale harmful meme dataset. Aligned with the international standards and the context of internet, MemeMind provides detailed Chain-of-Thought (CoT) reasoning annotations to support fine-grained analysis of implicit intentions in memes. Based on this dataset, we further propose MemeGuard, a reasoning-oriented multimodal detection framework that significantly improves both the accuracy of harmful meme detection and the interpretability of model decisions. Extensive experimental results demonstrate that MemeGuard outperforms existing state-of-the-art methods on the MemeMind dataset, establishing a solid foundation for future research in harmful meme detection. The complete dataset and code will be released upon acceptance.
Authors: Rafael Sojo, Javier D\'iaz-Rozo, Concha Bielza, Pedro Larra\~naga
Abstract: This paper introduces a new type of probabilistic semiparametric model that takes advantage of data binning to reduce the computational cost of kernel density estimation in nonparametric distributions. Two new conditional probability distributions are developed for the new binned semiparametric Bayesian networks, the sparse binned kernel density estimation and the Fourier kernel density estimation. These two probability distributions address the curse of dimensionality, which typically impacts binned models, by using sparse tensors and restricting the number of parent nodes in conditional probability calculations. To evaluate the proposal, we perform a complexity analysis and conduct several comparative experiments using synthetic data and datasets from the UCI Machine Learning repository. The experiments include different binning rules, parent restrictions, grid sizes, and number of instances to get a holistic view of the model's behavior. As a result, our binned semiparametric Bayesian networks achieve structural learning and log-likelihood estimations with no statistically significant differences compared to the semiparametric Bayesian networks, but at a much higher speed. Thus, the new binned semiparametric Bayesian networks prove to be a reliable and more efficient alternative to their non-binned counterparts.
Authors: Zhenpeng Su, Leiyu Pan, Xue Bai, Dening Liu, Guanting Dong, Jiaming Huang, Minxuan Lv, Wenping Hu, Fuzheng Zhang, Kun Gai, Guorui Zhou
Abstract: We present Klear-Reasoner, a model with long reasoning capabilities that demonstrates careful deliberation during problem solving, achieving outstanding performance across multiple benchmarks. Although there are already many excellent works related to inference models in the current community, there are still many problems with reproducing high-performance inference models due to incomplete disclosure of training details. This report provides an in-depth analysis of the reasoning model, covering the entire post-training workflow from data preparation and long Chain-of-Thought supervised fine-tuning (long CoT SFT) to reinforcement learning (RL), along with detailed ablation studies for each experimental component. For SFT data, our experiments show that a small number of high-quality data sources are more effective than a large number of diverse data sources, and that difficult samples can achieve better results without accuracy filtering. In addition, we investigate two key issues with current clipping mechanisms in RL: Clipping suppresses critical exploration signals and ignores suboptimal trajectories. To address these challenges, we propose Gradient-Preserving clipping Policy Optimization (GPPO) that gently backpropagates gradients from clipped tokens. GPPO not only enhances the model's exploration capacity but also improves its efficiency in learning from negative samples. Klear-Reasoner exhibits exceptional reasoning abilities in mathematics and programming, scoring 90.5% on AIME 2024, 83.2% on AIME 2025, 66.0% on LiveCodeBench V5 and 58.1% on LiveCodeBench V6.
Authors: Po-Hsien Yu, Yu-Syuan Tseng, Shao-Yi Chien
Abstract: Person re-identification (re-ID) is a fundamental task in intelligent surveillance and public safety. Federated learning (FL) provides a privacy-preserving paradigm by enabling collaborative model training without centralized data collection. However, applying FL to real-world re-ID systems remains challenging due to two major issues: statistical heterogeneity across clients caused by non-IID data distributions and substantial communication overhead resulting from the frequent transmission of large-scale models. To address these challenges, we propose FedKLPR, a lightweight and communication-efficient federated learning framework for person re-ID. FedKLPR consists of three key components. First, the KL-Divergence Regularization Loss (KLL) constrains local updates by reducing the discrepancy between local and global feature distributions, thereby alleviating the effects of statistical heterogeneity and improving convergence stability under non-IID settings. Second, KL-Divergence-Prune Weighted Aggregation (KLPWA) incorporates both pruning ratio and distributional similarity into the aggregation process, enabling more effective aggregation of pruned local models under non-IID data distributions and enhancing the robustness of the global model. Third, Cross-Round Recovery (CRR) employs a dynamic pruning control mechanism to prevent excessive pruning and preserve model accuracy during iterative compression. Experimental results on eight benchmark datasets demonstrate that FedKLPR achieves substantial communication savings while maintaining competitive accuracy. Compared with state-of-the-art methods, FedKLPR reduces communication cost by 40\%--42\% on ResNet-50 while achieving superior overall performance.
Authors: Jubayer Ibn Hamid, Ifdita Hasan Orney, Ellen Xu, Chelsea Finn, Dorsa Sadigh
Abstract: Reinforcement learning fine-tuning (RLFT) is a dominant paradigm for improving pretrained policies for downstream tasks. These pretrained policies, trained on large datasets, produce generations with a broad range of promising but unrefined behaviors. Often, a critical failure mode of RLFT arises when policies lose this diversity and collapse into a handful of easily exploitable outputs. This convergence hinders exploration, which is essential for expanding the capabilities of the pretrained policy and for amplifying the benefits of test-time compute scaling. To address this, we introduce an objective for policy gradient methods that explicitly enforces the exploration and refinement of diverse generations, which we call a polychromic objective. We then show how proximal policy optimization (PPO) can be adapted to optimize this objective. Our method (1) employs vine sampling to collect on-policy rollouts and (2) modifies the advantage function to reflect the advantage under our new objective. Experiments on BabyAI, Minigrid, and Algorithmic Creativity show that our method improves success rates by reliably solving a larger set of environment configurations and generalizes better under large perturbations. Moreover, when given multiple attempts in pass@$k$ experiments, the policy achieves substantially higher coverage, demonstrating its ability to maintain and exploit a diverse repertoire of strategies.
Authors: Yuanfang Xiang, Lun Ai
Abstract: The transcriptional response to genetic perturbation reveals fundamental insights into complex cellular systems. While current approaches have made progress in predicting genetic perturbation responses, they provide limited biological understanding and cannot systematically refine existing knowledge. Overcoming these limitations requires an end-to-end integration of data-driven learning and existing knowledge. However, this integration is challenging due to inconsistencies between data and knowledge bases, such as noise, misannotation, and incompleteness. To address this challenge, we propose ALIGNED (Adaptive aLignment for Inconsistent Genetic kNowledgE and Data), a neuro-symbolic framework based on the Abductive Learning (ABL) paradigm. This end-to-end framework aligns neural and symbolic components and performs systematic knowledge refinement. We introduce a balanced consistency metric to evaluate the predictions' consistency against both data and knowledge. Our results show that ALIGNED outperforms state-of-the-art methods by achieving the highest balanced consistency, while also re-discovering biologically meaningful knowledge. Our work advances beyond existing methods to enable both the transparency and the evolution of mechanistic biological understanding.
Authors: Eunki Kim, Na Min An, Wan Ju Kang, Sangryul Kim, James Thorne, Hyunjung Shim
Abstract: Large Vision-Language Models (LVLMs) demonstrate a promising direction for assisting individuals with blindness or low-vision (BLV). Yet, measuring their true utility in real-world scenarios is challenging because evaluating whether their descriptions are BLV-informative requires a fundamentally different approach from assessing standard scene descriptions. While the "VLM-as-a-metric" or "LVLM-as-a-judge" paradigm has emerged, existing evaluators still fall short of capturing the unique requirements of BLV-centric evaluation, lacking at least one of the following key properties: (1) High correlation with human judgments, (2) Long instruction understanding, (3) Score generation efficiency, and (4) Multi-dimensional assessment. To this end, we propose a unified framework to bridge the gap between automated evaluation and actual BLV needs. First, we conduct an in-depth user study with BLV participants to understand and quantify their navigational preferences, curating VL-GUIDEDATA, a large-scale BLV user-simulated preference dataset containing image-request-response-score pairs. We then leverage the dataset to develop an accessibility-aware evaluator, VL-GUIDE-S, which outperforms existing (L)VLM judges in both human alignment and inference efficiency. Notably, its effectiveness extends beyond a single domain, demonstrating strong performance across multiple fine-grained, BLV-critical dimensions. We hope our work lays as a foundation for automatic AI judges that advance safe, barrier-free navigation for BLV users.
Authors: Shira Schiber, Ofir Lindenbaum, Idan Schwartz
Abstract: Recent advances in generative video models have enabled the creation of high-quality videos based on natural language prompts. However, these models frequently lack fine-grained temporal control, meaning they do not allow users to specify when particular visual elements should appear within a generated sequence. In this work, we introduce TempoControl, a method that allows for temporal alignment of visual concepts during inference, without requiring retraining or additional supervision. TempoControl utilizes cross-attention maps, a key component of text-to-video diffusion models, to guide the timing of concepts through a novel optimization approach. Our method steers attention using three complementary principles: aligning its temporal pattern with a control signal (correlation), adjusting its strength where visibility is required (magnitude), and preserving semantic consistency (entropy). TempoControl provides precise temporal control while maintaining high video quality and diversity. We demonstrate its effectiveness across various applications, including temporal reordering of single and multiple objects, action timing, and audio-aligned video generation. Project page: https://shira-schiber.github.io/TempoControl/.
Authors: Jacek Karwowski, Raymond Douglas
Abstract: We investigate mathematically the notion of incoherence: a structural issue with reinforcement learning policies derived by naive goal-conditioning of autoregressive models. We focus on the process of re-training models on their own actions, that is, fine-tuning offline-learned policies with online RL. We prove that it decreases incoherence and leads to an improvement in return, and we aim to characterize the resulting trajectory of policies. By re-framing standard notions of control-as-inference and soft Q learning, we establish a three-way correspondence with two other ways of understanding the iterative re-training process: as folding the posterior into the reward and, in the deterministic case, as decreasing the temperature parameter; the correspondence has computational content via the training-inference trade-off. Through soft-conditioning generative models, we discuss the link between incoherence and the effective horizon.
Authors: Elias Hossain, Mehrdad Shoeibi, Ivan Garibay, Niloofar Yousefi
Abstract: Interpreting gene clusters from RNA sequencing (RNA-seq) remains challenging, especially in antimicrobial resistance studies where mechanistic insight is important for hypothesis generation. Existing pathway enrichment methods can summarize co-expressed modules, but they often provide limited cluster-specific explanations and weak connections to supporting literature. We present BIOGEN, an evidence-grounded multi-agent framework for post hoc interpretation of RNA-seq transcriptional modules. BIOGEN combines biomedical retrieval, structured reasoning, and multi-critic verification to generate traceable cluster-level explanations with explicit evidence and confidence labels. On a primary Salmonella enterica dataset, BIOGEN achieved strong biological grounding, including BERTScore 0.689, Semantic Alignment Score 0.715, KEGG Functional Similarity 0.342, and a hallucination rate of 0.000, compared with 0.100 for an LLM-only baseline. Across four additional bacterial RNA-seq datasets, BIOGEN also maintained zero hallucination under the same fixed pipeline. In comparisons with representative open-source agentic AI baselines, BIOGEN was the only framework that consistently preserved zero hallucination across all five datasets. These findings suggest that retrieval alone is not enough for reliable biological interpretation, and that evidence-grounded orchestration is important for transparent and source-traceable transcriptomic reasoning.
Authors: Miko{\l}aj Czarnecki, Micha{\l} Korniak, Oskar Skibski, Piotr Skowron
Abstract: We consider the problem of payoff division in indivisible coalitional games, where the value of the grand coalition is a natural number. This number represents a certain quantity of indivisible objects, such as parliamentary seats, kidney exchanges, or top features contributing to the outcome of a machine learning model. The goal of this paper is to propose a fair method for dividing these objects among players. To achieve this, we define the indivisible Shapley value and study its properties. We demonstrate our proposed technique using three case studies, in particular, we use it to identify key regions of an image in the context of an image classification task.
Authors: Guneet S. Dhillon, Javier Gonz\'alez, Teodora Pandeva, Alicia Curth
Abstract: While generative models, especially large language models (LLMs), are ubiquitous in today's world, principled mechanisms to assess their (in)correctness are limited. Using the conformal prediction framework, previous works construct sets of LLM responses where the probability of including an incorrect response, or error, is capped at a user-defined tolerance level. However, since these methods are based on p-values, they are susceptible to p-hacking, i.e., choosing the tolerance level post-hoc can invalidate the guarantees. We therefore leverage e-values to complement generative model outputs with e-scores as measures of incorrectness. In addition to achieving the guarantees as before, e-scores further provide users with the flexibility of choosing data-dependent tolerance levels while upper bounding size distortion, a post-hoc notion of error. We experimentally demonstrate their efficacy in assessing LLM outputs under different forms of correctness: mathematical factuality and property constraints satisfaction.
Authors: Yishan Du, Conrad Borchers, Mutlu Cukurova
Abstract: As teachers increasingly turn to GenAI in their educational practice, we need robust methods to benchmark large language models (LLMs) for pedagogical purposes. This article presents an embedding-based benchmarking framework to detect bias in LLMs in the context of formative feedback. Using 600 authentic student essays from the AES 2.0 corpus, we constructed controlled counterfactuals along two dimensions: (i) implicit cues via lexicon-based swaps of gendered terms within essays, and (ii) explicit cues via gendered author background in the prompt. We investigated six representative LLMs (i.e. GPT-5 mini, GPT-4o mini, DeepSeek-R1, DeepSeek-R1-Qwen, Gemini 2.5 Pro, Llama-3-8B). We first quantified the response divergence with cosine and Euclidean distances over sentence embeddings, then assessed significance via permutation tests, and finally, visualised structure using dimensionality reduction. In all models, implicit manipulations reliably induced larger semantic shifts for male-female counterfactuals than for female-male. Only the GPT and Llama models showed sensitivity to explicit gender cues. These findings show that even state-of-the-art LLMs exhibit asymmetric semantic responses to gender substitutions, suggesting persistent gender biases in feedback they provide learners. Qualitative analyses further revealed consistent linguistic differences (e.g., more autonomy-supportive feedback under male cues vs. more controlling feedback under female cues). We discuss implications for fairness auditing of pedagogical GenAI, propose reporting standards for counterfactual evaluation in learning analytics, and outline practical guidance for prompt design and deployment to safeguard equitable feedback.
Authors: Farheen Ramzan (Cherise), Yusuf Kiberu (Cherise), Nikesh Jathanna (Cherise), Meryem Jabrane (Cherise), Vicente Grau (Cherise), Shahnaz Jamil-Copley (Cherise), Richard H. Clayton (Cherise), Chen (Cherise), Chen (Cherise)
Abstract: Accurate segmentation of myocardial scar from late gadolinium enhanced (LGE) cardiac MRI is essential for evaluating tissue viability, yet remains challenging due to variable contrast and imaging artifacts. Electrocardiogram (ECG) signals provide complementary physiological information, as conduction abnormalities can help localize or suggest scarred myocardial regions. In this work, we propose a novel multimodal framework that integrates ECG-derived electrophysiological information with anatomical priors from the AHA-17 atlas for physiologically consistent LGE-based scar segmentation. As ECGs and LGE-MRIs are not acquired simultaneously, we introduce a Temporal Aware Feature Fusion (TAFF) mechanism that dynamically weights and fuses features based on their acquisition time difference. Our method was evaluated on a clinical dataset and achieved substantial gains over the state-of-the-art image-only baseline (nnU-Net), increasing the average Dice score for scars from 0.6149 to 0.8463 and achieving high performance in both precision (0.9115) and sensitivity (0.9043). These results show that integrating physiological and anatomical knowledge allows the model to "see beyond the image", setting a new direction for robust and physiologically grounded cardiac scar segmentation.
Authors: Rui Lin, Zhiyue Wu, Jiahe Le, Kangdi Wang, Weixiong Chen, Junyu Dai, Tao Jiang
Abstract: Audio tokenization bridges continuous waveforms and multi-track music language models. In dual-track modeling, tokens should preserve three properties at once: high-fidelity reconstruction, strong predictability under a language model, and cross-track correspondence. We introduce DuoTok, a source-aware dual-track tokenizer that addresses this trade-off through staged disentanglement. DuoTok first pretrains a semantic encoder, then regularizes it with multi-task supervision, freezes the encoder, and applies hard dual-codebook routing while keeping auxiliary objectives on quantized codes. A diffusion decoder reconstructs high-frequency details, allowing tokens to focus on structured information for sequence modeling. On standard benchmarks, DuoTok achieves a favorable predictability-fidelity trade-off, reaching the lowest cnBPT while maintaining competitive reconstruction at 0.75 kbps. Under a held-constant dual-track language modeling protocol, enBPT also improves, indicating gains beyond codebook size effects. Controlled diagnostics show larger predictability costs under cross-track corruption and larger gains from longer context, suggesting that models trained on DuoTok tokens use cross-track structure and non-local history.
Authors: Asad Aali, Muhammad Ahmed Mohsin, Vasiliki Bikia, Arnav Singhvi, Richard Gaus, Suhana Bedi, Hejie Cui, Miguel Fuentes, Alyssa Unell, Yifan Mai, Jordan Cahoon, Michael Pfeffer, Roxana Daneshjou, Sanmi Koyejo, Emily Alsentzer, Christopher Potts, Nigam H. Shah, Akshay S. Chaudhari
Abstract: As language models (LMs) are increasingly adopted across domains, high-quality benchmarking frameworks are essential for guiding deployment decisions. In practice, however, frameworks such as Holistic Evaluation of Language Models (HELM) typically evaluate models under a single static prompt configuration, even though model behavior depends strongly on prompt choice. As a result, reported scores can reflect prompt choice as much as model capability. Declarative prompting frameworks such as DSPy offer a scalable way to evaluate models under a set of structured prompting strategies rather than a static prompt configuration. We present a reproducible DSPy+HELM framework for studying how prompt choice impacts reported benchmark outcomes. Using five prompting methods, we evaluate four frontier and two open-source LMs across seven benchmarks against existing HELM baseline scores. By evaluating LMs across a family of prompt configurations, we find that prompt choice can materially impact leaderboard outcomes. In particular, structured prompting improves performance (by 6% on average), alters comparisons (leaderboard rankings shift on 5/7 benchmarks), with most gains coming from introducing chain-of-thought, and little additional benefit from more advanced optimizers. To our knowledge, this is the first study to systematically integrate structured prompting into an established evaluation framework and quantify how prompt choice alone can impact benchmark conclusions. We open-source (i) DSPy+HELM Evaluation (https://github.com/stanford-crfm/helm/pull/3893) and (ii) Prompt Optimization Pipeline (https://github.com/StanfordMIMI/dspy-helm).
URLs: https://github.com/stanford-crfm/helm/pull/3893), https://github.com/StanfordMIMI/dspy-helm).
Authors: Sai Koneru, Matthias Huck, Jan Niehues
Abstract: There has been significant progress in open-source text-only translation large language models (LLMs) with better language coverage and quality. However, these models can be only used in cascaded pipelines for speech translation (ST), performing automatic speech recognition first followed by translation. This introduces additional latency, which is particularly critical in simultaneous ST (SimulST), and prevents the model from exploiting multimodal context, such as images, which can aid disambiguation. Pretrained multimodal foundation models (MMFMs) already possess strong perception and reasoning capabilities across multiple modalities, but generally lack the multilingual coverage and specialized translation performance of dedicated translation LLMs. To build an effective multimodal translation system, we propose an end-to-end approach that fuses MMFMs with translation LLMs. We introduce a novel fusion strategy that connects hidden states from multiple layers of a pretrained MMFM to a translation LLM, enabling joint end-to-end training. The resulting model, OmniFusion, built on Omni 2.5-7B as the MMFM and SeedX PPO-7B as the translation LLM, can perform speech-to-text, speech-and-image-to-text, and text-and-image-to-text translation. Experiments demonstrate that OmniFusion effectively leverages both audio and visual inputs, achieves a 1-second latency reduction in SimulST compared to cascaded pipelines and also improves the overall translation quality\footnote{Code is available at https://github.com/saikoneru/OmniFusion}.
Authors: Dmitriy Parashchuk, Alexey Kaspshitskiy, Yuriy Karyakin
Abstract: Automatic 3D reconstruction of indoor spaces from 2D floor plans necessitates high-precision semantic segmentation of structural elements, particularly walls. However, existing methods often struggle with detecting thin structures and maintaining geometric precision. To address this, we introduce MitUNet, a hybrid neural network designed to bridge the gap between global semantic context and fine-grained structural details. Our architecture combines a Mix-Transformer encoder with a U-Net decoder enhanced with spatial and channel attention blocks. Optimized with the Tversky loss function, this approach achieves a balance between precision and recall, ensuring accurate boundary recovery. Experiments on the CubiCasa5k dataset and the regional dataset demonstrate MitUNet's superiority in generating structurally correct masks with high boundary accuracy, outperforming standard models. This tool provides a robust foundation for automated 3D reconstruction pipelines. To ensure reproducibility and facilitate future research, the source code and the regional dataset are publicly available at https://github.com/aliasstudio/mitunet and https://doi.org/10.5281/zenodo.17871079, respectively.
URLs: https://github.com/aliasstudio/mitunet, https://doi.org/10.5281/zenodo.17871079,
Authors: Isha Chaudhary, Vedaant Jain, Prineet Parhar, Kavya Sachdeva, Avaljot Singh, Sayan Ranu, Gagandeep Singh
Abstract: We introduce the first principled framework, Lumos, for specifying and formally certifying Language Model System (LMS) behaviors. Lumos is an imperative probabilistic programming DSL over graphs, with constructs to generate independent and identically distributed prompts for LMS. It offers a structured view of prompt distributions via graphs, forming random prompts from sampled subgraphs. Lumos supports certifying LMS for arbitrary prompt distributions via integration with statistical certifiers. We provide hybrid (operational and denotational) semantics for Lumos, providing a rigorous way to interpret the specifications. Using only a small set of composable constructs, Lumos can encode existing LMS specifications, including complex relational and temporal specifications. It also facilitates specifying new properties - we present the first safety specifications for vision-language models (VLMs) in autonomous driving scenarios developed with Lumos. Using these, we show that the state-of-the-art VLM Qwen-VL exhibits critical safety failures, producing incorrect and unsafe responses with at least 90% probability in right-turn scenarios under rainy driving conditions, revealing substantial safety risks. Lumos's modular structure allows easy modification of the specifications, enabling LMS certification to stay abreast with the rapidly evolving threat landscape. We further integrate a prompt-level deterministic verifier to obtain guarantees over the privacy of the LLM generation distribution over a prompt distribution. Lumos is simple to program in, requiring only a few constructs, as evidenced by state-of-the-art large language models generating correct Lumos specifications in zero-shot settings. Lumos is the first systematic and extensible language-based framework for specifying and certifying LMS behaviors, paving the way for a wider adoption of LMS certification.
Authors: Kai Kohyama, Yoshimitsu Aoki, Guillermo Gallego, Shintaro Shiba
Abstract: Event cameras offer a high temporal resolution over traditional frame-based cameras, which makes them suitable for motion and structure estimation. However, it has been unclear how event-based 3D Gaussian Splatting (3DGS) approaches could leverage fine-grained temporal information of sparse events. This work proposes GPERT, a framework to address the trade-off between accuracy and temporal resolution in event-based 3DGS. Our key idea is to decouple the rendering into two branches: event-by-event geometry (depth) rendering and snapshot-based radiance (intensity) rendering, by using ray-tracing and the image of warped events. The extensive evaluation shows that our method achieves state-of-the-art performance on the real-world datasets and competitive performance on the synthetic dataset. Also, the proposed method works without prior information (e.g., pretrained image reconstruction models) or COLMAP-based initialization, is more flexible in the event selection number, and achieves sharp reconstruction on scene edges with fast training time. We hope that this work deepens our understanding of the sparse nature of events for 3D reconstruction. https://github.com/e3ai/gpert
Authors: Md Jahedur Rahman, Ihsen Alouani
Abstract: Large language models (LLMs) are increasingly used in interactive and retrieval-augmented systems, but they remain vulnerable to prompt injection attacks, where injected secondary prompts force the model to deviate from the user's instructions to execute a potentially malicious task defined by the adversary. Recent work shows that ML models trained on activation shifts from LLMs' hidden layers can detect such drift. In this paper, we demonstrate that these detectors are not robust to adaptive adversaries. We propose a multi-probe evasion attack that appends an adversarially optimised suffix to poisoned inputs, jointly optimising a universal suffix to simultaneously fool all layer-wise drift detectors while preserving the effectiveness of the underlying injection. Using a modified Greedy Coordinate Gradient (GCG) approach, we generate universal suffixes that make prompt injections consistently evasive across multiple probes simultaneously. On Phi-3 3.8B and Llama-3 8B, a single suffix achieves attack success rates of 93.91% and 99.63% in successfully evading all detectors simultaneously. These results show that activation-based task drift detectors are highly vulnerable to adaptive prompt injection attacks, motivating stronger defences against such threats. We also propose a defence based on adversarial suffix augmentation: we generate multiple suffixes, append one at random during forward passes, and train detectors on the resulting activations. This approach is found to be effective against evasive attacks.
Authors: Sohan Venkatesh, Ashish Mahendran Kurapath
Abstract: Activation steering methods are widely used to control large language model (LLM) behavior and are often interpreted as revealing meaningful internal representations. This interpretation assumes that steering directions are identifiable and uniquely recoverable from input-output behavior. We show that, under white-box single-layer access, steering vectors are fundamentally non-identifiable due to large equivalence classes of behaviorally indistinguishable interventions. Empirically, we find that orthogonal perturbations achieve near-equivalent efficacy with negligible effect sizes across multiple models and traits, with pre-trained semantic classifiers confirming equivalence at the output level. We estimate null-space dimensionality via SVD of activation covariance matrices and validate that equivalence holds robustly throughout the operationally relevant steering range. Critically, we show that non-identifiability is a robust geometric property that persists across diverse prompt distributions. These findings reveal fundamental interpretability limits and highlight the need for structural constraints beyond behavioral testing to enable reliable alignment interventions.
Authors: Isaac Han, Sangyeon Park, Seungwon Oh, Donghu Kim, Hojoon Lee, Kyung-Joong Kim
Abstract: Deep neural networks trained on nonstationary data must balance stability (i.e., retaining prior knowledge) and plasticity (i.e., adapting to new tasks). Standard reinitialization methods, which reinitialize weights toward their original values, are widely used but difficult to tune: conservative reinitializations fail to restore plasticity, while aggressive ones erase useful knowledge. We propose FIRE, a principled reinitialization method that explicitly balances the stability-plasticity tradeoff. FIRE quantifies stability through Squared Frobenius Error (SFE), measuring proximity to past weights, and plasticity through Deviation from Isometry (DfI), reflecting weight isotropy. The reinitialization point is obtained by solving a constrained optimization problem, minimizing SFE subject to DfI being zero, which is efficiently approximated by Newton-Schulz iteration. FIRE is evaluated on continual visual learning (CIFAR-10 with ResNet-18), language modeling (OpenWebText with GPT-0.1B), and reinforcement learning (HumanoidBench with SAC and Atari games with DQN). Across all domains, FIRE consistently outperforms both naive training without intervention and standard reinitialization methods, demonstrating effective balancing of the stability-plasticity tradeoff.
Authors: Arshad Beg, Diarmuid O'Donoghue, Rosemary Monahan
Abstract: Formal specifications are crucial for building verifiable and dependable software systems, yet generating accurate and verifiable specifications for real-world C programs remains challenging. This paper empirically evaluates the extent to which formal-analysis tools can automatically generate and verify ACSL specifications without human or learning-based assistance. We conduct a controlled study on a recently released dataset of 506 C programs, repurposing it from interactive, developer-driven workflows to an automated evaluation setting. Five ACSL generation systems are compared: a rule-based Python script, Frama-C's RTE plugin, and three large language models--DeepSeek-V3.2, GPT-5.2, and OLMo 3.1 32B Instruct. All generated specifications are verified under identical conditions using the Frama-C WP plugin powered by multiple SMT solvers, allowing a direct comparison of annotation quality, solver sensitivity, and proof stability. Our results provide new empirical evidence on the capabilities and limitations of automated ACSL generation, complementing prior survey-based work.
Authors: Wenbo Nie, Zixiang Li, Renshuai Tao, Bin Wu, Yunchao Wei, Yao Zhao
Abstract: Transferring visual style between images while preserving semantic correspondence between similar objects remains a central challenge in computer vision. While existing methods have made great strides, most of them operate at global level but overlook region-wise and even pixel-wise semantic correspondence. To address this, we propose CoCoDiff, a novel training-free and low-cost style transfer framework that leverages pretrained latent diffusion models to achieve fine-grained, semantically consistent stylization. We identify that correspondence cues within generative diffusion models are under-explored and that content consistency across semantically matched regions is often neglected. CoCoDiff introduces a pixel-wise semantic correspondence module that mines intermediate diffusion features to construct a dense alignment map between content and style images. Furthermore, a cycle-consistency module then enforces structural and perceptual alignment across iterations, yielding object and region level stylization that preserves geometry and detail. Despite requiring no additional training or supervision, CoCoDiff delivers state-of-the-art visual quality and strong quantitative results, outperforming methods that rely on extra training or annotations.
Authors: Eason Chen, Sophia Judicke, Kayla Beigh, Xinyi Tang, Isabel Wang, Nina Yuan, Zimo Xiao, Chuangji Li, Shizhuo Li, Reed Luttmer, Shreya Singh, Maria Yampolsky, Naman Parikh, Yvonne Zhao, Meiyi Chen, Scarlett Huang, Anishka Mohanty, Gregory Johnson, John Mackey, Jionghao Lin, Ken Koedinger
Abstract: We evaluate GPTutor, an LLM-powered tutoring system for an undergraduate discrete mathematics course. It integrates two LLM-supported tools: a structured proof-review tool that provides embedded feedback on students' written proof attempts, and a chatbot for math questions. In a staggered-access study with 148 students, earlier access was associated with higher homework performance during the interval when only the experimental group could use the system, while we did not observe this performance increase transfer to exam scores. Usage logs show that students with lower self-efficacy and prior exam performance used both components more frequently. Session-level behavioral labels, produced by human coding and scaled using an automated classifier, characterize how students engaged with the chatbot (e.g., answer-seeking or help-seeking). In models controlling for prior performance and self-efficacy, higher chatbot usage and answer-seeking behavior were negatively associated with subsequent midterm performance, whereas proof-review usage showed no detectable independent association. Together, the findings suggest that chatbot-based support alone may not reliably support transfer to independent assessment of math proof-learning outcomes, whereas work-anchored, structured feedback appears less associated with reduced learning.
Authors: Tugrul Gorgulu, Atakan Dag, M. Esat Kalfaoglu, Halil Ibrahim Kuru, Baris Can Cam, Halil Ibrahim Ozturk, Ozsel Kilinc
Abstract: Collecting a high-quality dataset is a critical task that demands meticulous attention to detail, as overlooking certain aspects can render the entire dataset unusable. Autonomous driving challenges remain a prominent area of research, requiring further exploration to enhance the perception and planning performance of vehicles. However, existing datasets are often incomplete. For instance, datasets that include perception information generally lack planning data, while planning datasets typically consist of extensive driving sequences where the ego vehicle predominantly drives forward, offering limited behavioral diversity. In addition, many real datasets struggle to evaluate their models, especially for planning tasks, since they lack a proper closed-loop evaluation setup. The CARLA Leaderboard 2.0 challenge, which provides a diverse set of scenarios to address the long-tail problem in autonomous driving, has emerged as a valuable alternative platform for developing perception and planning models in both open-loop and closed-loop evaluation setups. Nevertheless, existing datasets collected on this platform present certain limitations. Some datasets appear to be tailored primarily for limited sensor configuration, with particular sensor configurations. To support end-to-end autonomous driving research, we have collected a new dataset comprising over 2.85 million frames using the CARLA simulation environment for the diverse Leaderboard 2.0 challenge scenarios. Our dataset is designed not only for planning tasks but also supports dynamic object detection, lane divider detection, centerline detection, traffic light recognition, prediction tasks and visual language action models . Furthermore, we demonstrate its versatility by training various models using our dataset. Moreover, we also provide numerical rarity scores to understand how rarely the current state occurs in the dataset.
Authors: Jialong Chen, Xander Xu, Hu Wei, Chuan Chen, Bing Zhao
Abstract: Large language model (LLM)-powered agents have demonstrated strong capabilities in automating software engineering tasks such as static bug fixing. However, in the real world, the development of mature software is typically predicated on complex requirement changes and long-term feature iterations -- a process that static, one-shot repair paradigms fail to capture. To bridge this gap, we propose SWE-CI, the first repository-level benchmark built upon the Continuous Integration loop, aiming to shift the evaluation paradigm for code generation from static, short-term functional correctness toward dynamic, long-term maintainability. The key insight is simple: Maintainability can be revealed by tracking how functional correctness changes over time. The benchmark comprises 100 tasks, each deriving from a real-world code repository with a development history spanning an average of 233 days and 71 consecutive commits. SWE-CI requires agents to systematically resolve these tasks through dozens of rounds of analysis and coding iterations. SWE-CI provides valuable insights into how well agents can sustain code quality throughout long-term evolution.
Authors: Yechen Zhang, Shuhao Xing, Junhao Huang, Kai Lv, Yunhua Zhou, Xipeng Qiu, Qipeng Guo, Kai Chen
Abstract: Recent advances in spectral optimization, notably Muon, have demonstrated that constraining update steps to the Stiefel manifold can significantly accelerate training and improve generalization. However, Muon implicitly assumes an isotropic optimization landscape, enforcing a uniform spectral update norm across all eigen-directions. We argue that this "egalitarian" constraint is suboptimal for Deep Neural Networks, where the curvature spectrum is known to be highly heavy-tailed and ill-conditioned. In such landscapes, Muon risks amplifying instabilities in high-curvature directions while limiting necessary progress in flat directions. In this work, we propose \textbf{Mousse} (\textbf{M}uon \textbf{O}ptimization \textbf{U}tilizing \textbf{S}hampoo's \textbf{S}tructural \textbf{E}stimation), a novel optimizer that reconciles the structural stability of spectral methods with the geometric adaptivity of second-order preconditioning. Instead of applying Newton-Schulz orthogonalization directly to the momentum matrix, Mousse operates in a whitened coordinate system induced by Kronecker-factored statistics (derived from Shampoo). Mathematically, we formulate Mousse as the solution to a spectral steepest descent problem constrained by an anisotropic trust region, where the optimal update is derived via the polar decomposition of the whitened gradient. Empirical results across language models ranging from 160M to 800M parameters demonstrate that Mousse consistently outperforms Muon, achieving around $\sim$12\% reduction in training steps with negligible computational overhead.
Authors: Ruiying Li, Yunlang Zhou, YuYao Zhu, Kylin Chen, Jingyuan Wang, Sukai Wang, Kongtao Hu, Minhui Yu, Bowen Jiang, Zhan Su, Jiayao Ma, Xin He, Yongjian Shen, Yang Yang, Guanghui Ren, Maoqing Yao, Wenhao Wang, Yao Mu
Abstract: Vision-Language-Action (VLA) systems have shown strong potential for language-driven robotic manipulation. However, scaling them to long-horizon tasks remains challenging. Existing pipelines typically separate data collection, policy learning, and deployment, resulting in heavy reliance on manual environment resets and brittle multi-policy execution. We present RoboClaw, an agentic robotics framework that unifies data collection, policy learning, and task execution under a single VLM-driven controller. At the policy level, RoboClaw introduces Entangled Action Pairs (EAP), which couple forward manipulation behaviors with inverse recovery actions to form self-resetting loops for autonomous data collection. This mechanism enables continuous on-policy data acquisition and iterative policy refinement with minimal human intervention. During deployment, the same agent performs high-level reasoning and dynamically orchestrates learned policy primitives to accomplish long-horizon tasks. By maintaining consistent contextual semantics across collection and execution, RoboClaw reduces mismatch between the two phases and improves multi-policy robustness. Experiments in real-world manipulation tasks demonstrate improved stability and scalability compared to conventional open-loop pipelines, while significantly reducing human effort throughout the robot lifecycle, achieving a 25% improvement in success rate over baseline methods on long-horizon tasks and reducing human time investment by 53.7%.
Authors: Mansoor Ahmed, Nadeem Taj, Imdad Ullah Khan, Hemanth Venkateswara, Murray Patterson
Abstract: Computational antibody design has seen rapid methodological progress, with dozens of deep generative methods proposed in the past three years, yet the field lacks a standardized benchmark for fair comparison and model development. These methods are evaluated on different SAbDab snapshots, non-overlapping test sets, and incompatible metrics, and the literature fragments the design problem into numerous sub-tasks with no common definition. We introduce \textsc{Chimera-Bench} (\textbf{C}DR \textbf{M}odeling with \textbf{E}pitope-guided \textbf{R}edesign), a unified benchmark built around a single canonical task: \emph{epitope-conditioned CDR sequence-structure co-design}. \textsc{Chimera-Bench} provides (1) a curated, deduplicated dataset of \textbf{2,922} antibody-antigen complexes with epitope and paratope annotations; (2) three biologically motivated splits testing generalization to unseen epitopes, unseen antigen folds, and prospective temporal targets; and (3) a comprehensive evaluation protocol with five metric groups including novel epitope-specificity measures. We benchmark representative methods spanning different generative paradigms and report results across all splits. \textsc{Chimera-Bench} is the largest dataset of its kind for the antibody design problem, allowing the community to develop and test novel methods and evaluate their generalizability. The source code and data are available at: https://github.com/mansoor181/chimera-bench.git
Authors: Haoyang Fang, Shuai Zhang, Yifei Ma, Hengyi Wang, Cuixiong Hu, Katrin Kirchhoff, Bernie Wang, George Karypis
Abstract: Domain-specific finetuning is essential for dense retrievers, yet not all training pairs contribute equally to the learning process. We introduce OPERA, a data pruning framework that exploits this heterogeneity to improve both the effectiveness and efficiency of retrieval model adaptation. We first investigate static pruning (SP), which retains only high-similarity query-document pairs, revealing an intrinsic quality-coverage tradeoff: ranking (NDCG) improves while retrieval (Recall) can degrade due to reduced query diversity. To resolve this tradeoff, we propose a two-stage dynamic pruning (DP) strategy that adaptively modulates sampling probabilities at both query and document levels throughout training, prioritizing high-quality examples while maintaining access to the full training set. Evaluations across eight datasets spanning six domains demonstrate the effectiveness of both approaches: SP improves ranking over standard finetuning (NDCG@10 +0.5\%), while DP achieves the strongest performance on both ranking (NDCG@10 +1.9\%) and retrieval (Recall@20 +0.7\%), with an average rank of 1.38 across all methods. These findings scale to Qwen3-Embedding, an LLM-based dense retriever, confirming architecture-agnostic benefits. Notably, DP reaches comparable performance in less than 50\% of the training time required by standard finetuning.
Authors: Ishrith Gowda, Chunwei Liu
Abstract: Multi-site neuroimaging analysis is fundamentally confounded by scanner-induced covariate shifts, where the marginal distribution of voxel intensities $P(\mathbf{x})$ varies non-linearly across acquisition protocols while the conditional anatomy $P(\mathbf{y}|\mathbf{x})$ remains constant. This is particularly detrimental to radiomic reproducibility, where acquisition variance often exceeds biological pathology variance. Existing statistical harmonization methods (e.g., ComBat) operate in feature space, precluding spatial downstream tasks, while standard deep learning approaches are theoretically bounded by local effective receptive fields (ERF), failing to model the global intensity correlations characteristic of field-strength bias. We propose SA-CycleGAN-2.5D, a domain adaptation framework motivated by the $H\Delta H$-divergence bound of Ben-David et al., integrating three architectural innovations: (1) A 2.5D tri-planar manifold injection preserving through-plane gradients $\nabla_z$ at $O(HW)$ complexity; (2) A U-ResNet generator with dense voxel-to-voxel self-attention, surpassing the $O(\sqrt{L})$ receptive field limit of CNNs to model global scanner field biases; and (3) A spectrally-normalized discriminator constraining the Lipschitz constant ($K_D \le 1$) for stable adversarial optimization. Evaluated on 654 glioma patients across two institutional domains (BraTS and UPenn-GBM), our method reduces Maximum Mean Discrepancy (MMD) by 99.1% ($1.729 \to 0.015$) and degrades domain classifier accuracy to near-chance (59.7%). Ablation confirms that global attention is statistically essential (Cohen's $d = 1.32$, $p < 0.001$) for the harder heterogeneous-to-homogeneous translation direction. By bridging 2D efficiency and 3D consistency, our framework yields voxel-level harmonized images that preserve tumor pathophysiology, enabling reproducible multi-center radiomic analysis.
Authors: Echo Zexuan Pan, Danny Glick, Ying Xu
Abstract: This study examined how high school students with different motivational profiles use generative AI tools in math and writing. Through K-means clustering analysis of survey data from 6,793 Mexican high school students, we identified three distinct motivational profiles based on self-concept and perceived subject value. Results revealed distinct domain-specific AI usage patterns across students with different motivational profiles. Our findings challenge one-size-fits-all AI integration approaches and advocate for motivationally-informed educational interventions.
Authors: Eric A. Moreno, Samuel Bright-Thonney, Andrzej Novak, Dolores Garcia, Philip Harris
Abstract: Large language model-based AI agents are now able to autonomously execute substantial portions of a high energy physics (HEP) analysis pipeline with minimal expert-curated input. Given access to a HEP dataset, an execution framework, and a corpus of prior experimental literature, we find that Claude Code succeeds in automating all stages of a typical analysis: event selection, background estimation, uncertainty quantification, statistical inference, and paper drafting. We argue that the experimental HEP community is underestimating the current capabilities of these systems, and that most proposed agentic workflows are too narrowly scoped or scaffolded to specific analysis structures. We present a proof-of-concept framework, Just Furnish Context (JFC), that integrates autonomous analysis agents with literature-based knowledge retrieval and multi-agent review, and show that this is sufficient to plan, execute, and document a credible high energy physics analysis. We demonstrate this by conducting analyses on open data from ALEPH, DELPHI, and CMS to perform electroweak, QCD, and Higgs boson measurements. Rather than replacing physicists, these tools promise to offload the repetitive technical burden of analysis code development, freeing researchers to focus on physics insight, truly novel method development, and rigorous validation. Given these developments, we advocate for new strategies for how the community trains students, organizes analysis efforts, and allocates human expertise.
Authors: Hengwei Ye, Yuanting Guan, Yuxuan Ge, Tianying Zhu, Zhenhan Guan, Yijia Zhong, Yijing Zhang, Han Zhang, Yingna Wu, Zheng Tian
Abstract: Multimodal Large Language Models (MLLMs) combine the linguistic strengths of LLMs with the ability to process multimodal data, enbaling them to address a broader range of visual tasks. Because MLLMs aim at more general, human-like competence than language-only models, we take inspiration from the Wechsler Intelligence Scales - an established battery for evaluating children by decomposing intelligence into interpretable, testable abilities. We introduce KidGym, a comprehensive 2D grid-based benchmark for assessing five essential capabilities of MLLMs: Execution, Perception Reasoning, Learning, Memory and Planning. The benchmark comprises 12 unique tasks, each targeting at least one core capability, specifically designed to guage MLLMs' adaptability and developmental potential, mirroring the stages of children's cognitive growth. Additionally, our tasks encompass diverse scenarios and objects with randomly generated layouts, ensuring a more accurate and robust evluation of MLLM capabilities. KidGym is designed to be fully user-customizable and extensible, allowing researchers to create new evaluation scenarios and adjust difficuly levels to accommodate the rapidly growing MLLM community. Through the evaluation of state-of-the-art MLLMs using KidGym, we identified significant insights into model capabilities and revealed several limitations of current models. We release our benchmark at: https://bobo-ye.github.io/KidGym/.
Authors: Andrea Marinoni, Erik Cambria, Luca Dal Zilio, Weisi Lin, Mauro Dalla Mura, Jocelyn Chanussot, Edoardo Ragusa, Chi Yan Tso, Yihao Zhu, Benjamin Horton
Abstract: The strong and continuous increase of AI-based services leads to the steady proliferation of AI data centres worldwide with the unavoidable escalation of their power consumption. It is unknown how this energy demand for computational purposes will impact the surrounding environment. Here, we focus our attention on the heat dissipation of AI hyperscalers. Taking advantage of land surface temperature measurements acquired by remote sensing platforms over the last decades, we are able to obtain a robust assessment of the temperature increase recorded in the areas surrounding AI data centres globally. We estimate that the land surface temperature increases by 2{\deg}C on average after the start of operations of an AI data centre, inducing local microclimate zones, which we call the data heat island effect. We assess the impact on the communities, quantifying that more than 340 million people could be affected by this temperature increase. Our results show that the data heat island effect could have a remarkable influence on communities and regional welfare in the future, hence becoming part of the conversation around environmentally sustainable AI worldwide.
Authors: Dogan Urgun, Gokhan Gungor
Abstract: Designing effective auxiliary rewards for cooperative multi-agent systems remains a challenging task. Misaligned incentives risk inducing suboptimal coordination, especially when sparse task feedback fails to provide sufficient grounding. This study introduces an automated reward design framework that leverages large language models to synthesize executable reward programs from environment instrumentation. The procedure constrains candidate programs within a formal validity envelope and evaluates their efficacy by training policies from scratch under a fixed computational budget. Selection across generations depends exclusively on the sparse task return. The framework is evaluated across four distinct Overcooked-AI layouts characterized by varied corridor congestion, handoff dependencies, and structural asymmetries. Iterative search generations consistently yield superior task returns and delivery counts, with the most pronounced gains occurring in environments dominated by interaction bottlenecks. Diagnostic analysis of the synthesized shaping components indicates increased interdependence in action selection and improved signal alignment in coordination-intensive tasks. These results demonstrate that the search for objective-grounded reward programs can mitigate the burden of manual engineering while producing shaping signals compatible with cooperative learning under finite budgets.
Authors: Marc-Antoine Allard, Arnaud Teinturier, Victor Xing, Gautier Viaud
Abstract: Recent advances in large language models (LLMs) have enabled the development of autonomous agents capable of complex reasoning and multi-step problem solving. However, these agents struggle to adapt to specialized environments and do not leverage past interactions, approaching each new task from scratch regardless of their accumulated experience. We introduce Experiential Reflective Learning (ERL), a simple self-improvement framework that enables rapid environment adaptation through experiential learning. ERL reflects on task trajectories and outcomes to generate heuristics, capturing actionable lessons that transfer across tasks. At test time, relevant heuristics are retrieved based on the current task and injected into the agent's context to guide execution. On the Gaia2 benchmark, ERL improves success rate by 7.8% over a ReAct baseline, with large gains in task completion reliability, and outperforms prior experiential learning methods. Through systematic ablations, we find that selective retrieval is essential and that heuristics provide more transferable abstractions than few-shot trajectory prompting. These results demonstrate that reflecting on single-attempt experiences to extract transferable heuristics enables effective agent self-improvement.
Authors: Miranda Muqing Miao, Lyle Ungar
Abstract: Large language models (LLMs) tend to verbalize confidence scores that are largely detached from their actual accuracy, yet the geometric relationship governing this behavior remain poorly understood. In this work, we present a mechanistic interpretability analysis of verbalized confidence, using linear probes and contrastive activation addition (CAA) steering to show that calibration and verbalized confidence signals are encoded linearly but are orthogonal to one another -- a finding consistent across three open-weight models and four datasets. Interestingly, when models are prompted to simultaneously reason through a problem and verbalize a confidence score, the reasoning process disrupts the verbalized confidence direction, exacerbating miscalibration. We term this the "Reasoning Contamination Effect." Leveraging this insight, we introduce a two-stage adaptive steering pipeline that reads the model's internal accuracy estimate and steers verbalized output to match it, substantially improving calibration alignment across all evaluated models.
Authors: Devashish Gaikwad, Wil M. P. van der Aalst, Gyunam Park
Abstract: Process anomaly detection is an important application of process mining for identifying deviations from the normal behavior of a process. Neural network-based methods have recently been applied to this task, learning directly from event logs without requiring a predefined process model. However, since anomaly detection is a purely statistical task, these models fail to incorporate human domain knowledge. As a result, rare but conformant traces are often misclassified as anomalies due to their low frequency, which limits the effectiveness of the detection process. Recent developments in the field of neuro-symbolic AI have introduced Logic Tensor Networks (LTN) as a means to integrate symbolic knowledge into neural networks using real-valued logic. In this work, we propose a neuro-symbolic approach that integrates domain knowledge into neural anomaly detection using LTN and Declare constraints. Using autoencoder models as a foundation, we encode Declare constraints as soft logical guiderails within the learning process to distinguish between anomalous and rare but conformant behavior. Evaluations on synthetic and real-world datasets demonstrate that our approach improves F1 scores even when as few as 10 conformant traces exist, and that the choice of Declare constraint and by extension human domain knowledge significantly influences performance gains.
Authors: Zehai He, Wenyi Hong, Zhen Yang, Ziyang Pan, Mingdao Liu, Xiaotao Gu, Jie Tang
Abstract: Recent advances in large language models have improved the capabilities of coding agents, yet systematic evaluation of complex, end-to-end website development remains limited. To address this gap, we introduce Vision2Web, a hierarchical benchmark for visual website development, spanning from static UI-to-code generation, interactive multi-page frontend reproduction, to long-horizon full-stack website development. The benchmark is constructed from real-world websites and comprises a total of 193 tasks across 16 categories, with 918 prototype images and 1,255 test cases. To support flexible, thorough and reliable evaluation, we propose workflow-based agent verification paradigm based on two complementary components: a GUI agent verifier and a VLM-based judge. We evaluate multiple visual language models instantiated under different coding-agent frameworks, revealing substantial performance gaps at all task levels, with state-of-the-art models still struggling on full-stack development.
Authors: Ehtibar N. Dzhafarov, Victor H. Cervantes
Abstract: We introduce a new notion, that of a contextuality profile of a system of random variables. Rather than characterizing a system's contextuality by a single number, its overall degree of contextuality, we show how it can be characterized by a curve relating degree of contextuality to level at which the system is considered. A system is represented at level n if one only considers the joint distributions with no more than n variables, ignoring higher-order joint distributions. We show that the level-wise contextuality analysis can be used in conjunction with any well-constructed measure of contextuality. We present a method of concatenated systems to explore contextuality profiles systematically, and we apply it to the contextuality profiles for three major measures of contextuality proposed in the literature.
Authors: Guilin Zhang, Wulan Guo, Ziqi Tan, Chuanyi Sun, Hailong Jiang
Abstract: Deep learning applications at the network edge lead to a significant growth in AI-related carbon emissions, presenting a critical sustainability challenge. The existing edge computing frameworks optimize for latency and throughput, but they largely ignore the environmental impact of inference workloads. This paper introduces CarbonEdge, a carbon-aware deep learning inference framework that extends adaptive model partitioning with carbon footprint estimation and green scheduling apabilities. We propose a carbon-aware scheduling algorithm that extends traditional weighted scoring with a carbon efficiency metric, supporting a tunable performance--carbon trade-off (demonstrated via weight sweep). Experimental evaluations on Docker-simulated heterogeneous edge environments show that CarbonEdge-Green mode achieves a 22.9% reduction in carbon emissions compared to monolithic execution. The framework achieves 1.3x improvement in carbon efficiency (245.8 vs 189.5 inferences per gram CO2) with negligible scheduling overhead (0.03ms per task). These results highlight the framework's potential for sustainable edge AI deployment, providing researchers and practitioners a tool to quantify and minimize the environmental footprint of distributed deep learning inference.
Authors: Kesheng Chen, Yamin Hu, Qi Zhou, Zhenqian Zhu, Wenjian Luo
Abstract: Vision-language models (VLMs) achieve strong performance on many benchmarks, yet a basic reliability question remains underexplored: when visual evidence conflicts with commonsense, do models follow what is shown or what commonsense suggests? A characteristic failure in this setting is that the model overrides visual evidence and outputs the commonsense alternative. We term this phenomenon \textbf{commonsense-driven hallucination} (CDH). To evaluate it, we introduce \textbf{CDH-Bench}, a benchmark designed to create explicit \textbf{visual evidence--commonsense conflicts}. CDH-Bench covers three dimensions: \textit{counting anomalies}, \textit{relational anomalies}, and \textit{attribute anomalies}. We evaluate frontier VLMs under \textit{binary Question Answering (QA)} and \textit{multiple-choice QA}, and report metrics including \textit{Counterfactual Accuracy} (CF-Acc), \textit{Commonsense Accuracy} (CS-Acc), \textit{Counterfactual Accuracy Drop} (CFAD), \textit{Commonsense Collapse Rate} (CCR), and \textit{Relative Prior Dependency} (RPD). Results show that even strong models remain vulnerable to prior-driven normalization under visual evidence--commonsense conflict. CDH-Bench provides a controlled diagnostic of visual fidelity under visual evidence--commonsense conflict.
Authors: Xuan Deng, Xiandong Meng, Hengyu Man, Qiang Zhu, Tiange Zhang, Debin Zhao, Xiaopeng Fan
Abstract: Although 3D Gaussian Splatting (3DGS) enables high-fidelity real-time rendering, its prohibitive storage overhead severely hinders practical deployment. Recent anchor-based 3DGS compression schemes reduce gaussian redundancy through some advanced context models. However, they overlook explicit geometric dependencies, leading to structural degradation and suboptimal ratedistortion performance. In this paper, we propose a Local Geometry-aware Hierarchical Context Compression framework for 3DGS(LG-HCC) that incorporates inter-anchor geometric correlations into anchor pruning and entropy coding for compact representation. Specifically, we introduce an Neighborhood-Aware Anchor Pruning (NAAP) strategy, which evaluates anchor importance via weighted neighborhood feature aggregation and then merges low-contribution anchors into salient neighbors, yielding a compact yet geometry-consistent anchor set. Moreover, we further develop a hierarchical entropy coding scheme, in which coarse-to-fine priors are exploited through a lightweight Geometry-Guided Convolution(GG-Conv) operator to enable spatially adaptive context modeling and rate-distortion optimization. Extensive experiments show that LG-HCC effectively alleviates structural preservation issues,achieving superior geometric integrity and rendering fidelity while reducing storage by up to 30.85x compared to the Scaffold-GS baseline on the Mip-NeRF360 dataset
Authors: Yufei Xu, Fanxu Meng, Fan Jiang, Yuxuan Wang, Ruijie Zhou, Zhaohui Wang, Jiexi Wu, Zhixin Pan, Xiaojuan Tang, Wenjie Pei, Tongxuan Liu, Di yin, Xing Sun, Muhan Zhang
Abstract: Token-level sparse attention mechanisms, exemplified by DeepSeek Sparse Attention (DSA), achieve fine-grained key selection by scoring every historical key for each query through a lightweight indexer, then computing attention only on the selected subset. While the downstream sparse attention itself scales favorably, the indexer must still scan the entire prefix for every query, introducing an per-layer bottleneck that grows prohibitively with context length. We propose HISA (Hierarchical Indexed Sparse Attention), a plug-and-play replacement for the indexer that rewrites the search path from a flat token scan into a two-stage hierarchical procedure: (1) a block-level coarse filtering stage that scores pooled block representations to discard irrelevant regions, followed by (2) a token-level refinement stage that applies the original indexer exclusively within the retained candidate blocks. HISA preserves the identical token-level top-sparse pattern consumed by the downstream Sparse MLA operator and requires no additional training. On kernel-level benchmarks, HISA achieves up to speedup at 64K context. On Needle-in-a-Haystack and LongBench, we directly replace the indexer in DeepSeek-V3.2 and GLM-5 with our HISA indexer, without any finetuning. HISA closely matches the original DSA in quality, while substantially outperforming block-sparse baselines.
Authors: Ziliang Guo, Ziheng Li, Bo Tang, Feiyu Xiong, Zhiyu Li
Abstract: Memory-augmented Large Language Models (LLMs) are essential for developing capable, long-term AI agents. Recently, applying Reinforcement Learning (RL) to optimize memory operations, such as extraction, updating, and retrieval, has emerged as a highly promising research direction. However, existing implementations remain highly fragmented and task-specific, lacking a unified infrastructure to streamline the integration, training, and evaluation of these complex pipelines. To address this gap, we present MemFactory, the first unified, highly modular training and inference framework specifically designed for memory-augmented agents. Inspired by the success of unified fine-tuning frameworks like LLaMA-Factory, MemFactory abstracts the memory lifecycle into atomic, plug-and-play components, enabling researchers to seamlessly construct custom memory agents via a "Lego-like" architecture. Furthermore, the framework natively integrates Group Relative Policy Optimization (GRPO) to fine-tune internal memory management policies driven by multi-dimensional environmental rewards. MemFactory provides out-of-the-box support for recent cutting-edge paradigms, including Memory-R1, RMM, and MemAgent. We empirically validate MemFactory on the open-source MemAgent architecture using its publicly available training and evaluation data. Across the evaluation sets, MemFactory improves performance over the corresponding base models on average, with relative gains of up to 14.8%. By providing a standardized, extensible, and easy-to-use infrastructure, MemFactory significantly lowers the barrier to entry, paving the way for future innovations in memory-driven AI agents.
Authors: Zhuoling Li, Jiarui Zhang, Ping Hu, Jason Kuen, Jiuxiang Gu, Hossein Rahmani, Jun Liu
Abstract: Method illustrations (MIs) play a crucial role in conveying the core ideas of scientific papers, yet their generation remains a labor-intensive process. Here, we take inspiration from human authors' drawing practices and correspondingly propose \textbf{FigAgent}, a novel multi-agent framework for high-quality automatic MI generation. Our FigAgent distills drawing experiences from similar components across MIs and encapsulates them into reusable drawing middlewares that can be orchestrated for MI generation, while evolving these middlewares to adapt to dynamically evolving drawing requirements. Besides, a novel Explore-and-Select drawing strategy is introduced to mimic the human-like trial-and-error manner for gradually constructing MIs with complex structures. Extensive experiments show the efficacy of our method.
Authors: Max Hennick, Guillaume Corlouer
Abstract: A key problem in the modern study of AI is predicting and understanding emergent capabilities in models during training. Inspired by methods for studying reactions in quantum chemistry, we present the ``2-datapoint reduced density matrix". We show that this object provides a computationally efficient, unified observable of phase transitions during training. By tracking the eigenvalue statistics of the 2RDM over a sliding window, we derive two complementary signals: the spectral heat capacity, which we prove provides early warning of second-order phase transitions via critical slowing down, and the participation ratio, which reveals the dimensionality of the underlying reorganization. Remarkably, the top eigenvectors of the 2RDM are directly interpretable making it straightforward to study the nature of the transitions. We validate across four distinct settings: deep linear networks, induction head formation, grokking, and emergent misalignment. We then discuss directions for future work using the 2RDM.