Authors: Xianxuan Long, Yao Fu, Runchao Li, Mu Sheng, Haotian Yu, Xiaotian Han, Pan Li
Abstract: Large language models (LLMs) tend to follow maliciously crafted instructions to generate deceptive responses, posing safety challenges. How deceptive instructions alter the internal representations of LLM compared to truthful ones remains poorly understood beyond output analysis. To bridge this gap, we investigate when and how these representations ``flip'', such as from truthful to deceptive, under deceptive versus truthful/neutral instructions. Analyzing the internal representations of Llama-3.1-8B-Instruct and Gemma-2-9B-Instruct on a factual verification task, we find the model's instructed True/False output is predictable via linear probes across all conditions based on the internal representation. Further, we use Sparse Autoencoders (SAEs) to show that the Deceptive instructions induce significant representational shifts compared to Truthful/Neutral representations (which are similar), concentrated in early-to-mid layers and detectable even on complex datasets. We also identify specific SAE features highly sensitive to deceptive instruction and use targeted visualizations to confirm distinct truthful/deceptive representational subspaces. % Our analysis pinpoints layer-wise and feature-level correlates of instructed dishonesty, offering insights for LLM detection and control. Our findings expose feature- and layer-level signatures of deception, offering new insights for detecting and mitigating instructed dishonesty in LLMs.
Authors: Matthieu Queloz
Abstract: This paper argues that explainability is only one facet of a broader ideal that shapes our expectations towards artificial intelligence (AI). Fundamentally, the issue is to what extent AI exhibits systematicity--not merely in being sensitive to how thoughts are composed of recombinable constituents, but in striving towards an integrated body of thought that is consistent, coherent, comprehensive, and parsimoniously principled. This richer conception of systematicity has been obscured by the long shadow of the "systematicity challenge" to connectionism, according to which network architectures are fundamentally at odds with what Fodor and colleagues termed "the systematicity of thought." I offer a conceptual framework for thinking about "the systematicity of thought" that distinguishes four senses of the phrase. I use these distinctions to defuse the perceived tension between systematicity and connectionism and show that the conception of systematicity that historically shaped our sense of what makes thought rational, authoritative, and scientific is more demanding than the Fodorian notion. To determine whether we have reason to hold AI models to this ideal of systematicity, I then argue, we must look to the rationales for systematization and explore to what extent they transfer to AI models. I identify five such rationales and apply them to AI. This brings into view the "hard systematicity challenge." However, the demand for systematization itself needs to be regulated by the rationales for systematization. This yields a dynamic understanding of the need to systematize thought, which tells us how systematic we need AI models to be and when.
Authors: Minsoo Kim, Seung-won Hwang
Abstract: Planning in modern LLM agents relies on the utilization of LLM as an internal world model, acquired during pretraining. However, existing agent designs fail to effectively assimilate new observations into dynamic updates of the world model. This reliance on the LLM's static internal world model is progressively prone to misalignment with the underlying true state of the world, leading to the generation of divergent and erroneous plans. We introduce a hierarchical agent architecture, CoEx, in which hierarchical state abstraction allows LLM planning to co-evolve with a dynamically updated model of the world. CoEx plans and interacts with the world by using LLM reasoning to orchestrate dynamic plans consisting of subgoals, and its learning mechanism continuously incorporates these subgoal experiences into a persistent world model in the form of a neurosymbolic belief state, comprising textual inferences and code-based symbolic memory. We evaluate our agent across a diverse set of agent scenarios involving rich environments and complex tasks including ALFWorld, PDDL, and Jericho. Our experiments show that CoEx outperforms existing agent paradigms in planning and exploration.
Authors: Qun Ma, Xiao Xue, Ming Zhang, Yifan Shen, Zihan Zhao
Abstract: Metaverse service is a product of the convergence between Metaverse and service systems, designed to address service-related challenges concerning digital avatars, digital twins, and digital natives within Metaverse. With the rise of large language models (LLMs), agents now play a pivotal role in Metaverse service ecosystem, serving dual functions: as digital avatars representing users in the virtual realm and as service assistants (or NPCs) providing personalized support. However, during the modeling of Metaverse service ecosystems, existing LLM-based agents face significant challenges in bridging virtual-world services with real-world services, particularly regarding issues such as character data fusion, character knowledge association, and ethical safety concerns. This paper proposes an explainable emotion alignment framework for LLM-based agents in Metaverse Service Ecosystem. It aims to integrate factual factors into the decision-making loop of LLM-based agents, systematically demonstrating how to achieve more relational fact alignment for these agents. Finally, a simulation experiment in the Offline-to-Offline food delivery scenario is conducted to evaluate the effectiveness of this framework, obtaining more realistic social emergence.
Authors: Hussein Mozannar, Gagan Bansal, Cheng Tan, Adam Fourney, Victor Dibia, Jingya Chen, Jack Gerrits, Tyler Payne, Matheus Kunzler Maldaner, Madeleine Grunde-McLaughlin, Eric Zhu, Griffin Bassman, Jacob Alber, Peter Chang, Ricky Loynd, Friederike Niedtner, Ece Kamar, Maya Murad, Rafah Hosn, Saleema Amershi
Abstract: AI agents powered by large language models are increasingly capable of autonomously completing complex, multi-step tasks using external tools. Yet, they still fall short of human-level performance in most domains including computer use, software development, and research. Their growing autonomy and ability to interact with the outside world, also introduces safety and security risks including potentially misaligned actions and adversarial manipulation. We argue that human-in-the-loop agentic systems offer a promising path forward, combining human oversight and control with AI efficiency to unlock productivity from imperfect systems. We introduce Magentic-UI, an open-source web interface for developing and studying human-agent interaction. Built on a flexible multi-agent architecture, Magentic-UI supports web browsing, code execution, and file manipulation, and can be extended with diverse tools via Model Context Protocol (MCP). Moreover, Magentic-UI presents six interaction mechanisms for enabling effective, low-cost human involvement: co-planning, co-tasking, multi-tasking, action guards, and long-term memory. We evaluate Magentic-UI across four dimensions: autonomous task completion on agentic benchmarks, simulated user testing of its interaction capabilities, qualitative studies with real users, and targeted safety assessments. Our findings highlight Magentic-UI's potential to advance safe and efficient human-agent collaboration.
Authors: Qianhong Guo, Wei Xie, Xiaofang Cai, Enze Wang, Shuoyoucheng Ma, Kai Chen, Xiaofeng Wang, Baosheng Wang
Abstract: Although large language models (LLMs) demonstrate remarkable capabilities across various tasks, evaluating their capabilities remains a challenging task. Existing evaluation methods suffer from issues such as data contamination, black-box operation, and subjective preference. These issues make it difficult to evaluate the LLMs' true capabilities comprehensively. To tackle these challenges, we propose a novel benchmark-free evaluation paradigm, LLM-Crowdsourced. It utilizes LLMs to generate questions, answer independently, and evaluate mutually. This method integrates four key evaluation criteria: dynamic, transparent, objective, and professional, which existing evaluation methods cannot satisfy simultaneously. Experiments on eight mainstream LLMs across mathematics and programming verify the advantages of our method in distinguishing LLM performance. Furthermore, our study reveals several novel findings that are difficult for traditional methods to detect, including but not limited to: (1) Gemini demonstrates the highest original and professional question-design capabilities among others; (2) Some LLMs exhibit ''memorization-based answering'' by misrecognizing questions as familiar ones with a similar structure; (3) LLM evaluation results demonstrate high consistency (robustness).
Authors: ZhaoBin Li, Mark Steyvers
Abstract: In settings where human decision-making relies on AI input, both the predictive accuracy of the AI system and the reliability of its confidence estimates influence decision quality. We highlight the role of AI metacognitive sensitivity -- its ability to assign confidence scores that accurately distinguish correct from incorrect predictions -- and introduce a theoretical framework for assessing the joint impact of AI's predictive accuracy and metacognitive sensitivity in hybrid decision-making settings. Our analysis identifies conditions under which an AI with lower predictive accuracy but higher metacognitive sensitivity can enhance the overall accuracy of human decision making. Finally, a behavioral experiment confirms that greater AI metacognitive sensitivity improves human decision performance. Together, these findings underscore the importance of evaluating AI assistance not only by accuracy but also by metacognitive sensitivity, and of optimizing both to achieve superior decision outcomes.
Authors: Kei-Sing Ng
Abstract: To engineer AGI, we should first capture the essence of intelligence in a species-agnostic form that can be evaluated, while being sufficiently general to encompass diverse paradigms of intelligent behavior, including reinforcement learning, generative models, classification, analogical reasoning, and goal-directed decision-making. We propose a general criterion based on sample fidelity: intelligence is the ability, given sample(s) from a category, to generate sample(s) from the same category. We formalise this intuition as {\epsilon}-category intelligence: it is {\epsilon}-intelligent with respect to a category if no chosen admissible distinguisher can separate generated from original samples beyond tolerance {\epsilon}. We present the formal framework, outline empirical protocols, and discuss implications for evaluation, safety, and generalization.
Authors: Zhe Yu, Yiwei Lu, Burkhard Schafer, Zhe Lin
Abstract: This paper focuses on the legal compliance challenges of autonomous vehicles in a transnational context. We choose the perspective of designers and try to provide supporting legal reasoning in the design process. Based on argumentation theory, we introduce a logic to represent the basic properties of argument-based practical (normative) reasoning, combined with partial order sets of natural numbers to express priority. Finally, through case analysis of legal texts, we show how the reasoning system we provide can help designers to adapt their design solutions more flexibly in the cross-border application of autonomous vehicles and to more easily understand the legal implications of their decisions.
Authors: Yiya Diao, Changhe Li, Sanyou Zeng, Xinye Cai, Wenjian Luo, Shengxiang Yang, Carlos A. Coello Coello
Abstract: The Nearest-Better Network (NBN) is a powerful method to visualize sampled data for continuous optimization problems while preserving multiple landscape features. However, the calculation of NBN is very time-consuming, and the extension of the method to combinatorial optimization problems is challenging but very important for analyzing the algorithm's behavior. This paper provides a straightforward theoretical derivation showing that the NBN network essentially functions as the maximum probability transition network for algorithms. This paper also presents an efficient NBN computation method with logarithmic linear time complexity to address the time-consuming issue. By applying this efficient NBN algorithm to the OneMax problem and the Traveling Salesman Problem (TSP), we have made several remarkable discoveries for the first time: The fitness landscape of OneMax exhibits neutrality, ruggedness, and modality features. The primary challenges of TSP problems are ruggedness, modality, and deception. Two state-of-the-art TSP algorithms (i.e., EAX and LKH) have limitations when addressing challenges related to modality and deception, respectively. LKH, based on local search operators, fails when there are deceptive solutions near global optima. EAX, which is based on a single population, can efficiently maintain diversity. However, when multiple attraction basins exist, EAX retains individuals within multiple basins simultaneously, reducing inter-basin interaction efficiency and leading to algorithm's stagnation.
Authors: Hongyan Cheng, Chengzhang Yu, Yanshu Shi, Chiyue Wang, Cong Liu, Zhanpeng Jin
Abstract: The post-pandemic surge in healthcare demand, coupled with critical nursing shortages, has placed unprecedented pressure on emergency department triage systems, necessitating innovative AI-driven solutions. We present a multi-agent interactive intelligent system for medical triage that addresses three fundamental challenges in current AI-based triage systems: insufficient medical specialization leading to hallucination-induced misclassifications, heterogeneous department structures across healthcare institutions, and inefficient detail-oriented questioning that impedes rapid triage decisions. Our system employs three specialized agents - RecipientAgent, InquirerAgent, and DepartmentAgent - that collaborate through structured inquiry mechanisms and department-specific guidance rules to transform unstructured patient symptoms into accurate department recommendations. To ensure robust evaluation, we constructed a comprehensive Chinese medical triage dataset from a medical website, comprising 3,360 real-world cases spanning 9 primary departments and 62 secondary departments. Through systematic data imputation using large language models, we address the prevalent issue of incomplete medical records in real-world data. Experimental results demonstrate that our multi-agent system achieves 89.2% accuracy in primary department classification and 73.9% accuracy in secondary department classification after four rounds of patient interaction. The system's pattern-matching-based guidance mechanisms enable efficient adaptation to diverse hospital configurations while maintaining high triage accuracy. Our work provides a scalable framework for deploying AI-assisted triage systems that can accommodate the organizational heterogeneity of healthcare institutions while ensuring clinically sound decision-making.
Authors: Yaolun Zhang, Xiaogeng Liu, Chaowei Xiao
Abstract: Large Language Models (LLMs) have demonstrated the ability to solve a wide range of practical tasks within multi-agent systems. However, existing human-designed multi-agent frameworks are typically limited to a small set of pre-defined scenarios, while current automated design methods suffer from several limitations, such as the lack of tool integration, dependence on external training data, and rigid communication structures. In this paper, we propose MetaAgent, a finite state machine based framework that can automatically generate a multi-agent system. Given a task description, MetaAgent will design a multi-agent system and polish it through an optimization algorithm. When the multi-agent system is deployed, the finite state machine will control the agent's actions and the state transitions. To evaluate our framework, we conduct experiments on both text-based tasks and practical tasks. The results indicate that the generated multi-agent system surpasses other auto-designed methods and can achieve a comparable performance with the human-designed multi-agent system, which is optimized for those specific tasks.
Authors: Sebastian Monka, Irlan Grangel-Gonz\'alez, Stefan Schmid, Lavdim Halilaj, Marc Rickart, Oliver Rudolph, Rui Dias
Abstract: Knowledge graphs (KGs) have transformed data management within the manufacturing industry, offering effective means for integrating disparate data sources through shared and structured conceptual schemas. However, harnessing the power of KGs can be daunting for non-experts, as it often requires formulating complex SPARQL queries to retrieve specific information. With the advent of Large Language Models (LLMs), there is a growing potential to automatically translate natural language queries into the SPARQL format, thus bridging the gap between user-friendly interfaces and the sophisticated architecture of KGs. The challenge remains in adequately informing LLMs about the relevant context and structure of domain-specific KGs, e.g., in manufacturing, to improve the accuracy of generated queries. In this paper, we evaluate multiple strategies that use LLMs as mediators to facilitate information retrieval from KGs. We focus on the manufacturing domain, particularly on the Bosch Line Information System KG and the I40 Core Information Model. In our evaluation, we compare various approaches for feeding relevant context from the KG to the LLM and analyze their proficiency in transforming real-world questions into SPARQL queries. Our findings show that LLMs can significantly improve their performance on generating correct and complete queries when provided only the adequate context of the KG schema. Such context-aware prompting techniques help LLMs to focus on the relevant parts of the ontology and reduce the risk of hallucination. We anticipate that the proposed techniques help LLMs to democratize access to complex data repositories and empower informed decision-making in manufacturing settings.
Authors: Thomas Eiter, Tobias Geibinger, Tobias Kaminski, Nysret Musliu, Johannes Oetsch
Abstract: We present the solver asp-fzn for Constraint Answer Set Programming (CASP), which extends ASP with linear constraints. Our approach is based on translating CASP programs into the solver-independent FlatZinc language that supports several Constraint Programming and Integer Programming backend solvers. Our solver supports a rich language of linear constraints, including some common global constraints. As for evaluation, we show that asp-fzn is competitive with state-of-the-art ASP solvers on benchmarks taken from past ASP competitions. Furthermore, we evaluate it on several CASP problems from the literature and compare its performance with clingcon, which is a prominent CASP solver that supports most of the asp-fzn language. The performance of asp-fzn is very promising as it is already competitive on plain ASP and even outperforms clingcon on some CASP benchmarks.
Authors: Hugo Garrido-Lestache, Jeremy Kedziora
Abstract: This paper introduces Team-Attention-Actor-Critic (TAAC), a reinforcement learning algorithm designed to enhance multi-agent collaboration in cooperative environments. TAAC employs a Centralized Training/Centralized Execution scheme incorporating multi-headed attention mechanisms in both the actor and critic. This design facilitates dynamic, inter-agent communication, allowing agents to explicitly query teammates, thereby efficiently managing the exponential growth of joint-action spaces while ensuring a high degree of collaboration. We further introduce a penalized loss function which promotes diverse yet complementary roles among agents. We evaluate TAAC in a simulated soccer environment against benchmark algorithms representing other multi-agent paradigms, including Proximal Policy Optimization and Multi-Agent Actor-Attention-Critic. We find that TAAC exhibits superior performance and enhanced collaborative behaviors across a variety of metrics (win rates, goal differentials, Elo ratings, inter-agent connectivity, balanced spatial distributions, and frequent tactical interactions such as ball possession swaps).
Authors: Han Jiang, Pengda Wang, Xiaoyuan Yi, Xing Xie, Ziang Xiao
Abstract: Social sciences have accumulated a rich body of theories and methodologies for investigating the human mind and behaviors, while offering valuable insights into the design and understanding of Artificial Intelligence (AI) systems. Focusing on psychology as a prominent case, this study explores the interdisciplinary synergy between AI and the field by analyzing 1,006 LLM-related papers published in premier AI venues between 2023 and 2025, along with the 2,544 psychology publications they cite. Through our analysis, we identify key patterns of interdisciplinary integration, locate the psychology domains most frequently referenced, and highlight areas that remain underexplored. We further examine how psychology theories/frameworks are operationalized and interpreted, identify common types of misapplication, and offer guidance for more effective incorporation. Our work provides a comprehensive map of interdisciplinary engagement between AI and psychology, thereby facilitating deeper collaboration and advancing AI systems.
Authors: Yiwen Sun, Furong Ye, Zhihan Chen, Ke Wei, Shaowei Cai
Abstract: Satisfiability problem (SAT) is a cornerstone of computational complexity with broad industrial applications, and it remains challenging to optimize modern SAT solvers in real-world settings due to their intricate architectures. While automatic configuration frameworks have been developed, they rely on manually constrained search spaces and yield limited performance gains. This work introduces a novel paradigm which effectively optimizes complex SAT solvers via Large Language Models (LLMs), and a tool called AutoModSAT is developed. Three fundamental challenges are addressed in order to achieve superior performance: (1) LLM-friendly solver: Systematic guidelines are proposed for developing a modularized solver to meet LLMs' compatibility, emphasizing code simplification, information share and bug reduction; (2) Automatic prompt optimization: An unsupervised automatic prompt optimization method is introduced to advance the diversity of LLMs' output; (3) Efficient search strategy: We design a presearch strategy and an EA evolutionary algorithm for the final efficient and effective discovery of heuristics. Extensive experiments across a wide range of datasets demonstrate that AutoModSAT achieves 50% performance improvement over the baseline solver and achieves 30% superiority against the state-of-the-art (SOTA) solvers. Moreover, AutoModSAT attains a 20% speedup on average compared to parameter-tuned alternatives of the SOTA solvers, showcasing the enhanced capability in handling complex problem instances. This work bridges the gap between AI-driven heuristics discovery and mission-critical system optimization, and provides both methodological advancements and empirically validated results for next-generation complex solver development.
Authors: Hanchen Yang, Wengen Li, Shuyu Wang, Hui Li, Jihong Guan, Shuigeng Zhou, Jiannong Cao
Abstract: With the rapid amassing of spatial-temporal (ST) ocean data, many spatial-temporal data mining (STDM) studies have been conducted to address various oceanic issues, including climate forecasting and disaster warning. Compared with typical ST data (e.g., traffic data), ST ocean data is more complicated but with unique characteristics, e.g., diverse regionality and high sparsity. These characteristics make it difficult to design and train STDM models on ST ocean data. To the best of our knowledge, a comprehensive survey of existing studies remains missing in the literature, which hinders not only computer scientists from identifying the research issues in ocean data mining but also ocean scientists to apply advanced STDM techniques. In this paper, we provide a comprehensive survey of existing STDM studies for ocean science. Concretely, we first review the widely-used ST ocean datasets and highlight their unique characteristics. Then, typical ST ocean data quality enhancement techniques are explored. Next, we classify existing STDM studies in ocean science into four types of tasks, i.e., prediction, event detection, pattern mining, and anomaly detection, and elaborate on the techniques for these tasks. Finally, promising research opportunities are discussed. This survey can help scientists from both computer science and ocean science better understand the fundamental concepts, key techniques, and open challenges of STDM for ocean science.
Authors: Jiajie He, Yuechun Gu, Keke Chen
Abstract: Recommender systems (RecSys) have become an essential component of many web applications. The core of the system is a recommendation model trained on highly sensitive user-item interaction data. While privacy-enhancing techniques are actively studied in the research community, the real-world model development still depends on minimal privacy protection, e.g., via controlled access. Users of such systems should have the right to choose \emph{not} to share highly sensitive interactions. However, there is no method allowing the user to know which interactions are more sensitive than others. Thus, quantifying the privacy risk of RecSys training data is a critical step to enabling privacy-aware RecSys model development and deployment. We propose a membership-inference attack (MIA)- based privacy scoring method, RecPS, to measure privacy risks at both the interaction and user levels. The RecPS interaction-level score definition is motivated and derived from differential privacy, which is then extended to the user-level scoring method. A critical component is the interaction-level MIA method RecLiRA, which gives high-quality membership estimation. We have conducted extensive experiments on well-known benchmark datasets and RecSys models to show the unique features and benefits of RecPS scoring in risk assessment and RecSys model unlearning.
Authors: Amin Banayeeanzade, Fatemeh Bahrani, Yutai Zhou, Erdem B{\i}y{\i}k
Abstract: Imitation Learning (IL) is a widely adopted approach which enables agents to learn from human expert demonstrations by framing the task as a supervised learning problem. However, IL often suffers from causal confusion, where agents misinterpret spurious correlations as causal relationships, leading to poor performance in testing environments with distribution shift. To address this issue, we introduce GAze-Based Regularization in Imitation Learning (GABRIL), a novel method that leverages the human gaze data gathered during the data collection phase to guide the representation learning in IL. GABRIL utilizes a regularization loss which encourages the model to focus on causally relevant features identified through expert gaze and consequently mitigates the effects of confounding variables. We validate our approach in Atari environments and the Bench2Drive benchmark in CARLA by collecting human gaze datasets and applying our method in both domains. Experimental results show that the improvement of GABRIL over behavior cloning is around 179% more than the same number for other baselines in the Atari and 76% in the CARLA setup. Finally, we show that our method provides extra explainability when compared to regular IL agents.
Authors: Wenjie Jacky Mo, Qin Liu, Xiaofei Wen, Dongwon Jung, Hadi Askari, Wenxuan Zhou, Zhe Zhao, Muhao Chen
Abstract: Large Language Models (LLMs) for code generation (i.e., Code LLMs) have demonstrated impressive capabilities in AI-assisted software development and testing. However, recent studies have shown that these models are prone to generating vulnerable or even malicious code under adversarial settings. Existing red-teaming approaches rely on extensive human effort, limiting their scalability and practicality, and generally overlook the interactive nature of real-world AI-assisted programming, which often unfolds over multiple turns. To bridge these gaps, we present RedCoder, a red-teaming agent that engages victim models in multi-turn conversation to elicit vulnerable code. The pipeline to construct RedCoder begins with a multi-agent gaming process that simulates adversarial interactions, yielding a set of prototype conversations and an arsenal of reusable attack strategies. We then fine-tune an LLM on these prototype conversations to serve as the backbone of RedCoder. Once deployed, RedCoder autonomously engages Code LLMs in multi-turn conversations, dynamically retrieving relevant strategies from the arsenal to steer the dialogue toward vulnerability-inducing outputs. Experiments across multiple Code LLMs show that our approach outperforms prior single-turn and multi-turn red-team methods in inducing vulnerabilities in code generation, offering a scalable and effective tool for evaluating the security boundaries of modern code-generation systems.
Authors: Michael Cohoon, Debbie Furman
Abstract: This paper details the machine learning (ML) journey of a group of people focused on software testing. It tells the story of how this group progressed through a ML workflow (similar to the CRISP-DM process). This workflow consists of the following steps and can be used by anyone applying ML techniques to a project: gather the data; clean the data; perform feature engineering on the data; splitting the data into two sets, one for training and one for testing; choosing a machine learning model; training the model; testing the model and evaluating the model performance. By following this workflow, anyone can effectively apply ML to any project that they are doing.
Authors: Xiaotao Feng, Xiaogang Zhu, Kun Hu, Jincheng Wang, Yingjie Cao, Guang Gong, Jianfeng Pan
Abstract: Fuzzing is highly effective in detecting bugs due to the key contribution of randomness. However, randomness significantly reduces the efficiency of fuzzing, causing it to cost days or weeks to expose bugs. Even though directed fuzzing reduces randomness by guiding fuzzing towards target buggy locations, the dilemma of randomness still challenges directed fuzzers. Two critical components, which are seeds and mutators, contain randomness and are closely tied to the conditions required for triggering bugs. Therefore, to address the challenge of randomness, we propose to use large language models (LLMs) to remove the randomness in seeds and reduce the randomness in mutators. With their strong reasoning and code generation capabilities, LLMs can be used to generate reachable seeds that target pre-determined locations and to construct bug-specific mutators tailored for specific bugs. We propose RandLuzz, which integrates LLMs and directed fuzzing, to improve the quality of seeds and mutators, resulting in efficient bug exposure. RandLuzz analyzes function call chain or functionality to guide LLMs in generating reachable seeds. To construct bug-specific mutators, RandLuzz uses LLMs to perform bug analysis, obtaining information such as bug causes and mutation suggestions, which further help generate code that performs bug-specific mutations. We evaluate RandLuzz by comparing it with four state-of-the-art directed fuzzers, AFLGo, Beacon, WindRanger, and SelectFuzz. With RandLuzz-generated seeds, the fuzzers achieve an average speedup ranging from 2.1$\times$ to 4.8$\times$ compared to using widely-used initial seeds. Additionally, when evaluated on individual bugs, RandLuzz achieves up to a 2.7$\times$ speedup compared to the second-fastest exposure. On 8 bugs, RandLuzz can even expose them within 60 seconds.
Authors: Yuang Cao, Jiachen Zou, Chen Wei, Quanying Liu
Abstract: Human memory exhibits significant vulnerability in cognitive tasks and daily life. Comparisons between visual working memory and new perceptual input (e.g., during cognitive tasks) can lead to unintended memory distortions. Previous studies have reported systematic memory distortions after perceptual comparison, but understanding how perceptual comparison affects memory distortions in real-world objects remains a challenge. Furthermore, identifying what visual features contribute to memory vulnerability presents a novel research question. Here, we propose a novel AI-driven framework that generates naturalistic visual stimuli grounded in behaviorally relevant object dimensions to elicit similarity-induced memory biases. We use two types of stimuli -- image wheels created through dimension editing and dimension wheels generated by dimension activation values -- in three visual working memory (VWM) experiments. These experiments assess memory distortions under three conditions: no perceptual comparison, perceptual comparison with image wheels, and perceptual comparison with dimension wheels. The results show that similar dimensions, like similar images, can also induce memory distortions. Specifically, visual dimensions are more prone to distortion than semantic dimensions, indicating that the object dimensions of naturalistic visual stimuli play a significant role in the vulnerability of memory.
Authors: Tobias Sesterhenn, Ian Berlot-Attwell, Janis Zenkner, Christian Bartelt
Abstract: Reusing established theorems and formulas is central to mathematical problem solving, serving as essential building blocks for tackling increasingly complex challenges. Recent work, TroVE, argues that code-generating Large Language Models (LLMs) can benefit similarly on the MATH benchmark by inducing and reusing higher-level toolboxes. By allocating computational budget across an ensemble of three modes -- directly generating code, creating tools, and reusing tools -- TroVE claims to outperform a PRIMITIVE baseline that only performs direct generation. However, recent analysis (Berlot-Attwell et al., 2024) casts doubt on these gains, noting that the tools created are often trivial or rarely reused, suggesting that improvements may stem from self-consistency or self-correction. In this work, we re-evaluate TroVE on MATH, analyze the impact of each of its modes, and show that its benefit does not come from these mechanisms, but simply from a higher computational budget spent for TroVE compared to PRIMITIVE. To this end, we also perform a small correction in the original implementation of TroVE's selection mechanism, boosting TroVE's performance on MATH by 3\% in accuracy. After matching for compute, the benefit of TroVE reduces to a marginal improvement of 1\%, suggesting that this toolbox approach does not provide a significant benefit on MATH.
Authors: Muyang Li
Abstract: As autonomous agents powered by large language models (LLMs) proliferate in high-stakes domains -- from pharmaceuticals to legal workflows -- the challenge is no longer just intelligence, but verifiability. We introduce TrustTrack, a protocol that embeds structural guarantees -- verifiable identity, policy commitments, and tamper-resistant behavioral logs -- directly into agent infrastructure. This enables a new systems paradigm: trust-native autonomy. By treating compliance as a design constraint rather than post-hoc oversight, TrustTrack reframes how intelligent agents operate across organizations and jurisdictions. We present the protocol design, system requirements, and use cases in regulated domains such as pharmaceutical R&D, legal automation, and AI-native collaboration. We argue that the Cloud -> AI -> Agent -> Trust transition represents the next architectural layer for autonomous systems.
Authors: Qiushi Sun, Jinyang Gong, Lei Li, Qipeng Guo, Fei Yuan
Abstract: Acquiring high-quality instruction-code pairs is essential for training Large Language Models (LLMs) for code generation. Manually curated data is expensive and inherently limited in scale, motivating the development of code-centric synthesis methods. Yet, current approaches either focus on augmenting existing code or rely on predefined heuristics, both lacking rigorous data validation, which results in synthetic data that is ungrounded, repetitive, or overly simplistic. Inspired by collaborative programming practices, we propose CodeEvo, a framework that synthesizes code data through iterative interactions between two LLM agents: a Coder, which generates candidate code and test cases based on given instructions, and a Reviewer, which guides the synthesis process by producing new instructions and feedback. We further introduce a hybrid feedback mechanism that combines compiler determinism with the generative flexibility of agents, enabling automatic quality control throughout synthesis. Extensive experiments demonstrate that models fine-tuned on CodeEvo data significantly outperform established baselines across code generation benchmarks with various difficulties. In-depth analyses further provide insights from multiple perspectives into effective code-centric data synthesis.
Authors: Anuraj Maurya
Abstract: Deep learning provides a versatile suite of methods for extracting structured information from complex datasets, enabling deeper understanding of underlying fluid dynamic phenomena. The field of turbulence modeling, in particular, benefits from the growing availability of high-dimensional data obtained through experiments, field observations, and large-scale simulations spanning multiple spatio-temporal scales. This report presents a concise overview of both classical and deep learningbased approaches to turbulence modeling. It further investigates two specific challenges at the intersection of fluid dynamics and machine learning: the integration of multiscale turbulence models with deep learning architectures, and the application of deep generative models for super-resolution reconstruction
Authors: Honghua Dong, Jiacheng Yang, Xun Deng, Yuhe Jiang, Gennady Pekhimenko, Fan Long, Xujie Si
Abstract: Type inference for dynamic languages like Python is a persistent challenge in software engineering. While large language models (LLMs) have shown promise in code understanding, their type inference capabilities remain underexplored. We introduce TypyBench, a benchmark designed to evaluate LLMs' type inference across entire Python repositories. TypyBench features two novel metrics: TypeSim, which captures nuanced semantic relationships between predicted and ground truth types, and TypeCheck, which assesses type consistency across codebases. Our evaluation of various LLMs on a curated dataset of 50 high-quality Python repositories reveals that, although LLMs achieve decent TypeSim scores, they struggle with complex nested types and exhibit significant type consistency errors. These findings suggest that future research should shift focus from improving type similarity to addressing repository-level consistency. TypyBench provides a foundation for this new direction, offering insights into model performance across different type complexities and usage contexts. Our code and data are available at https://github.com/typybench/typybench.
Authors: Harsh Nilesh Pathak, Randy Paffenroth
Abstract: In this work, we propose a parameter continuation method for the optimization of neural networks. There is a close connection between parameter continuation, homotopies, and curriculum learning. The methods we propose here are theoretically justified and practically effective for several problems in deep neural networks. In particular, we demonstrate better generalization performance than state-of-the-art optimization techniques such as ADAM for supervised and unsupervised learning tasks.
Authors: Sergii Kavun
Abstract: Activation functions are critical components in deep neural networks, directly influencing gradient flow, training stability, and model performance. Traditional functions like ReLU suffer from dead neuron problems, while sigmoid and tanh exhibit vanishing gradient issues. We introduce two novel hybrid activation functions: S3 (Sigmoid-Softsign) and its improved version S4 (smoothed S3). S3 combines sigmoid for negative inputs with softsign for positive inputs, while S4 employs a smooth transition mechanism controlled by a steepness parameter k. We conducted comprehensive experiments across binary classification, multi-class classification, and regression tasks using three different neural network architectures. S4 demonstrated superior performance compared to nine baseline activation functions, achieving 97.4% accuracy on MNIST, 96.0% on Iris classification, and 18.7 MSE on Boston Housing regression. The function exhibited faster convergence (-19 for ReLU) and maintained stable gradient flow across network depths. Comparative analysis revealed S4's gradient range of [0.24, 0.59] compared to ReLU's 18% dead neurons in deep networks. The S4 activation function addresses key limitations of existing functions through its hybrid design and smooth transition mechanism. The tunable parameter k allows adaptation to different tasks and network depths, making S4 a versatile choice for deep learning applications. These findings suggest that hybrid activation functions represent a promising direction for improving neural network training dynamics.
Authors: Gianluca Carloni, Biagio Brattoli, Seongho Keum, Jongchan Park, Taebum Lee, Chang Ho Ahn, Sergio Pereira
Abstract: Computational pathology (CPath) has shown great potential in mining actionable insights from Whole Slide Images (WSIs). Deep Learning (DL) has been at the center of modern CPath, and while it delivers unprecedented performance, it is also known that DL may be affected by irrelevant details, such as those introduced during scanning by different commercially available scanners. This may lead to scanner bias, where the model outputs for the same tissue acquired by different scanners may vary. In turn, it hinders the trust of clinicians in CPath-based tools and their deployment in real-world clinical practices. Recent pathology Foundation Models (FMs) promise to provide better domain generalization capabilities. In this paper, we benchmark FMs using a multi-scanner dataset and show that FMs still suffer from scanner bias. Following this observation, we propose ScanGen, a contrastive loss function applied during task-specific fine-tuning that mitigates scanner bias, thereby enhancing the models' robustness to scanner variations. Our approach is applied to the Multiple Instance Learning task of Epidermal Growth Factor Receptor (EGFR) mutation prediction from H\&E-stained WSIs in lung cancer. We observe that ScanGen notably enhances the ability to generalize across scanners, while retaining or improving the performance of EGFR mutation prediction.
Authors: Nicholas Mehlman, Jean-Christophe Gagnon-Audet, Michael Shvartsman, Kelvin Niu, Alexander H. Miller, Shagun Sodhani
Abstract: Surface electromyography (sEMG) signals offer a promising avenue for developing innovative human-computer interfaces by providing insights into muscular activity. However, the limited volume of training data and computational constraints during deployment have restricted the investigation of scaling up the model size for solving sEMG tasks. In this paper, we demonstrate that vanilla transformer models can be effectively scaled up on sEMG data and yield improved cross-user performance up to 110M parameters, surpassing the model size regime investigated in other sEMG research (usually <10M parameters). We show that >100M-parameter models can be effectively distilled into models 50x smaller with minimal loss of performance (<1.5% absolute). This results in efficient and expressive models suitable for complex real-time sEMG tasks in real-world environments.
Authors: Shuqing Li, Qiang Chen, Xiaoxue Ren, Michael R. Lyu
Abstract: Physics Engines (PEs) are fundamental software frameworks that simulate physical interactions in applications ranging from entertainment to safety-critical systems. Despite their importance, PEs suffer from physics failures, deviations from expected physical behaviors that can compromise software reliability, degrade user experience, and potentially cause critical failures in autonomous vehicles or medical robotics. Current testing approaches for PE-based software are inadequate, typically requiring white-box access and focusing on crash detection rather than semantically complex physics failures. This paper presents the first large-scale empirical study characterizing physics failures in PE-based software. We investigate three research questions addressing the manifestations of physics failures, the effectiveness of detection techniques, and developer perceptions of current detection practices. Our contributions include: (1) a taxonomy of physics failure manifestations; (2) a comprehensive evaluation of detection methods including deep learning, prompt-based techniques, and large multimodal models; and (3) actionable insights from developer experiences for improving detection approaches. To support future research, we release PhysiXFails, code, and other materials at https://sites.google.com/view/physics-failure-detection.
URLs: https://sites.google.com/view/physics-failure-detection.
Authors: Hamed Jafarzadeh Asl, Mahsa Ghazvini Nejad, Amin Edraki, Masoud Asgharian, Vahid Partovi Nia
Abstract: Voice Activity Detection (VAD) in the presence of background noise remains a challenging problem in speech processing. Accurate VAD is essential in automatic speech recognition, voice-to-text, conversational agents, etc, where noise can severely degrade the performance. A modern application includes the voice assistant, specially mounted on Artificial Intelligence of Things (AIoT) devices such as cell phones, smart glasses, earbuds, etc, where the voice signal includes background noise. Therefore, VAD modules must remain light-weight due to their practical on-device limitation. The existing models often struggle with low signal-to-noise ratios across diverse acoustic environments. A simple VAD often detects human voice in a clean environment, but struggles to detect the human voice in noisy conditions. We propose a noise-robust VAD that comprises a light-weight VAD, with data pre-processing and post-processing added modules to handle the background noise. This approach significantly enhances the VAD accuracy in noisy environments and requires neither a larger model, nor fine-tuning. Experimental results demonstrate that our approach achieves a notable improvement compared to baselines, particularly in environments with high background noise interference. This modified VAD additionally improving clean speech detection.
Authors: Vanessa Rebecca Wiyono, David Anugraha, Ayu Purwarianti, Genta Indra Winata
Abstract: Over 200 million people speak Indonesian, yet the language remains significantly underrepresented in preference-based research for large language models (LLMs). Most existing multilingual datasets are derived from English translations, often resulting in content that lacks cultural and linguistic authenticity. To address this gap, we introduce IndoPref, the first fully human-authored and multi-domain Indonesian preference dataset specifically designed to evaluate the naturalness and quality of LLM-generated text. All annotations are natively written in Indonesian and evaluated using Krippendorff's alpha, demonstrating strong inter-annotator agreement. Additionally, we benchmark the dataset across multiple LLMs and assess the output quality of each model.
Authors: Yassine Rachidy, Jihad Rbaiti, Youssef Hmamouche, Faissal Sehbaoui, Amal El Fallah Seghrouchni
Abstract: With the growing adoption of Large Language Models (LLMs) in critical areas, ensuring their security against jailbreaking attacks is paramount. While traditional defenses primarily rely on refusing malicious prompts, recent logit-level attacks have demonstrated the ability to bypass these safeguards by directly manipulating the token-selection process during generation. We introduce Strategic Deflection (SDeflection), a defense that redefines the LLM's response to such advanced attacks. Instead of outright refusal, the model produces an answer that is semantically adjacent to the user's request yet strips away the harmful intent, thereby neutralizing the attacker's harmful intent. Our experiments demonstrate that SDeflection significantly lowers Attack Success Rate (ASR) while maintaining model performance on benign queries. This work presents a critical shift in defensive strategies, moving from simple refusal to strategic content redirection to neutralize advanced threats.
Authors: Kimberly Le Truong, Riccardo Fogliato, Hoda Heidari, Zhiwei Steven Wu
Abstract: Current benchmarks for evaluating Large Language Models (LLMs) often do not exhibit enough writing style diversity, with many adhering primarily to standardized conventions. Such benchmarks do not fully capture the rich variety of communication patterns exhibited by humans. Thus, it is possible that LLMs, which are optimized on these benchmarks, may demonstrate brittle performance when faced with "non-standard" input. In this work, we test this hypothesis by rewriting evaluation prompts using persona-based LLM prompting, a low-cost method to emulate diverse writing styles. Our results show that, even with identical semantic content, variations in writing style and prompt formatting significantly impact the estimated performance of the LLM under evaluation. Notably, we identify distinct writing styles that consistently trigger either low or high performance across a range of models and tasks, irrespective of model family, size, and recency. Our work offers a scalable approach to augment existing benchmarks, improving the external validity of the assessments they provide for measuring LLM performance across linguistic variations.
Authors: Zheng Zhang, Peilin Zhao, Deheng Ye, Hao Wang
Abstract: Jailbreak attacks aim to exploit large language models (LLMs) by inducing them to generate harmful content, thereby revealing their vulnerabilities. Understanding and addressing these attacks is crucial for advancing the field of LLM safety. Previous jailbreak approaches have mainly focused on direct manipulations of harmful intent, with limited attention to the impact of persona prompts. In this study, we systematically explore the efficacy of persona prompts in compromising LLM defenses. We propose a genetic algorithm-based method that automatically crafts persona prompts to bypass LLM's safety mechanisms. Our experiments reveal that: (1) our evolved persona prompts reduce refusal rates by 50-70% across multiple LLMs, and (2) these prompts demonstrate synergistic effects when combined with existing attack methods, increasing success rates by 10-20%. Our code and data are available at https://github.com/CjangCjengh/Generic_Persona.
Authors: Molly Wang, Kin. K Leung
Abstract: Reinforcement Learning (RL) has been widely used for packet routing in communication networks, but traditional RL methods rely on the Markov assumption that the current state contains all necessary information for decision-making. In reality, internet traffic is non-Markovian, and past states do influence routing performance. Moreover, common deep RL approaches use function approximators, such as neural networks, that do not model the spatial structure in network topologies. To address these shortcomings, we design a network environment with non-Markovian traffic and introduce a spatial-temporal RL (STRL) framework for packet routing. Our approach outperforms traditional baselines by more than 19% during training and 7% for inference despite a change in network topology.
Authors: Ambarish Singh, Romila Pradhan
Abstract: Data quality plays a pivotal role in the predictive performance of machine learning (ML) tasks - a challenge amplified by the deluge of data sources available in modern organizations.Prior work in data discovery largely focus on metadata matching, semantic similarity or identifying tables that should be joined to answer a particular query, but do not consider source quality for high performance of the downstream ML task.This paper addresses the problem of determining the best subset of data sources that must be combined to construct the underlying training dataset for a given ML task.We propose SourceGrasp and SourceSplice, frameworks designed to efficiently select a suitable subset of sources that maximizes the utility of the downstream ML model.Both the algorithms rely on the core idea that sources (or their combinations) contribute differently to the task utility, and must be judiciously chosen.While SourceGrasp utilizes a metaheuristic based on a greediness criterion and randomization, the SourceSplice framework presents a source selection mechanism inspired from gene splicing - a core concept used in protein synthesis.We empirically evaluate our algorithms on three real-world datasets and synthetic datasets and show that, with significantly fewer subset explorations, SourceSplice effectively identifies subsets of data sources leading to high task utility.We also conduct studies reporting the sensitivity of SourceSplice to the decision choices under several settings.
Authors: Adam M. Morgan, Adeen Flinker
Abstract: We present an automated pipeline for estimating Verb Frame Frequencies (VFFs), the frequency with which a verb appears in particular syntactic frames. VFFs provide a powerful window into syntax in both human and machine language systems, but existing tools for calculating them are limited in scale, accuracy, or accessibility. We use large language models (LLMs) to generate a corpus of sentences containing 476 English verbs. Next, by instructing an LLM to behave like an expert linguist, we had it analyze the syntactic structure of the sentences in this corpus. This pipeline outperforms two widely used syntactic parsers across multiple evaluation datasets. Furthermore, it requires far fewer resources than manual parsing (the gold-standard), thereby enabling rapid, scalable VFF estimation. Using the LLM parser, we produce a new VFF database with broader verb coverage, finer-grained syntactic distinctions, and explicit estimates of the relative frequencies of structural alternates commonly studied in psycholinguistics. The pipeline is easily customizable and extensible to new verbs, syntactic frames, and even other languages. We present this work as a proof of concept for automated frame frequency estimation, and release all code and data to support future research.
Authors: Hongjie Chen, Akshay Mehra, Josh Kimball, Ryan A. Rossi
Abstract: The emergence of time-series foundation model research elevates the growing need to measure the (dis)similarity of time-series datasets. A time-series dataset similarity measure aids research in multiple ways, including model selection, finetuning, and visualization. In this paper, we propose a distribution-based method to measure time-series dataset similarity by leveraging the Wasserstein distance. We consider a time-series dataset an empirical instantiation of an underlying multivariate normal distribution (MVN). The similarity between two time-series datasets is thus computed as the Wasserstein distance between their corresponding MVNs. Comprehensive experiments and visualization show the effectiveness of our approach. Specifically, we show how the Wasserstein distance helps identify similar time-series datasets and facilitates inference performance estimation of foundation models in both out-of-distribution and transfer learning evaluation, with high correlations between our proposed measure and the inference loss (>0.60).
Authors: Shreyansh Pathak, Sonu Shreshtha, Richa Singh, Mayank Vatsa
Abstract: The widespread adoption of voice-enabled authentication and audio biometric systems have significantly increased privacy vulnerabilities associated with sensitive speech data. Compliance with privacy regulations such as GDPR's right to be forgotten and India's DPDP Act necessitates targeted and efficient erasure of individual-specific voice signatures from already-trained biometric models. Existing unlearning methods designed for visual data inadequately handle the sequential, temporal, and high-dimensional nature of audio signals, leading to ineffective or incomplete speaker and accent erasure. To address this, we introduce QPAudioEraser, a quantum-inspired audio unlearning framework. Our our-phase approach involves: (1) weight initialization using destructive interference to nullify target features, (2) superposition-based label transformations that obscure class identity, (3) an uncertainty-maximizing quantum loss function, and (4) entanglement-inspired mixing of correlated weights to retain model knowledge. Comprehensive evaluations with ResNet18, ViT, and CNN architectures across AudioMNIST, Speech Commands, LibriSpeech, and Speech Accent Archive datasets validate QPAudioEraser's superior performance. The framework achieves complete erasure of target data (0% Forget Accuracy) while incurring minimal impact on model utility, with a performance degradation on retained data as low as 0.05%. QPAudioEraser consistently surpasses conventional baselines across single-class, multi-class, sequential, and accent-level erasure scenarios, establishing the proposed approach as a robust privacy-preserving solution.
Authors: Dongyub Jude Lee, Zhenyi Ye, Pengcheng He
Abstract: Preference-learning methods for machine translation (MT)--such as Direct Preference Optimization (DPO)--have achieved impressive gains but depend heavily on large, carefully curated triplet datasets and often struggle to generalize beyond their tuning domains. We propose Reinforcement Learning from Teacher-Model Refinement (RLfR), a novel framework that removes reliance on static triplets by leveraging continuous, high-quality feedback from an external teacher model (GPT-4o). RLfR frames each translation step as a micro-tutorial: the actor generates a hypothesis, the teacher refines it, and the actor is rewarded based on how closely it aligns with the teacher's refinement. Guided by two complementary signals--(i) negative edit distance, promoting lexical and structural fidelity, and (ii) COMET score, ensuring semantic adequacy--the actor progressively learns to emulate the teacher, mirroring a human learning process through incremental, iterative improvement. On the FLORES-200 benchmark (English to and from German, Spanish, Chinese, Korean, and Japanese), RLfR consistently outperforms both MT-SFT and preference-based baselines, significantly improving COMET (semantic adequacy) and M-ETA (entity preservation) scores.
Authors: Muhammad Sharshar, Ahmad Mohammad Saber, Davor Svetinovic, Amr M. Youssef, Deepa Kundur, Ehab F. El-Saadany
Abstract: The increasing digitization of smart grids has improved operational efficiency but also introduced new cybersecurity vulnerabilities, such as False Data Injection Attacks (FDIAs) targeting Automatic Generation Control (AGC) systems. While machine learning (ML) and deep learning (DL) models have shown promise in detecting such attacks, their opaque decision-making limits operator trust and real-world applicability. This paper proposes a hybrid framework that integrates lightweight ML-based attack detection with natural language explanations generated by Large Language Models (LLMs). Classifiers such as LightGBM achieve up to 95.13% attack detection accuracy with only 0.004 s inference latency. Upon detecting a cyberattack, the system invokes LLMs, including GPT-3.5 Turbo, GPT-4 Turbo, and GPT-4o mini, to generate human-readable explanation of the event. Evaluated on 100 test samples, GPT-4o mini with 20-shot prompting achieved 93% accuracy in identifying the attack target, a mean absolute error of 0.075 pu in estimating attack magnitude, and 2.19 seconds mean absolute error (MAE) in estimating attack onset. These results demonstrate that the proposed framework effectively balances real-time detection with interpretable, high-fidelity explanations, addressing a critical need for actionable AI in smart grid cybersecurity.
Authors: Oleksiy Ostapenko, Charles Guille-Escuret, Luke Kumar, Max Tian, Denis Kocetkov, Gopeshh Subbaraj, Raymond Li, Joel Lamy-Poirier, Sebastien Paquet, Torsten Scholak
Abstract: We introduce a framework for optimizing domain-specific dataset construction in foundation model training. Specifically, we seek a cost-efficient way to estimate the quality of data sources (e.g. synthetically generated or filtered web data, etc.) in order to make optimal decisions about resource allocation for data sourcing from these sources for the stage two pre-training phase, aka annealing, with the goal of specializing a generalist pre-trained model to specific domains. Our approach extends the usual point estimate approaches, aka micro-annealing, to estimating scaling laws by performing multiple annealing runs of varying compute spent on data curation and training. This addresses a key limitation in prior work, where reliance on point estimates for data scaling decisions can be misleading due to the lack of rank invariance across compute scales -- a phenomenon we confirm in our experiments. By systematically analyzing performance gains relative to acquisition costs, we find that scaling curves can be estimated for different data sources. Such scaling laws can inform cost effective resource allocation across different data acquisition methods (e.g. synthetic data), data sources (e.g. user or web data) and available compute resources. We validate our approach through experiments on a pre-trained model with 7 billion parameters. We adapt it to: a domain well-represented in the pre-training data -- the medical domain, and a domain underrepresented in the pretraining corpora -- the math domain. We show that one can efficiently estimate the scaling behaviors of a data source by running multiple annealing runs, which can lead to different conclusions, had one used point estimates using the usual micro-annealing technique instead. This enables data-driven decision-making for selecting and optimizing data sources.
Authors: Hanqi Zhou, Fryderyk Mantiuk, David G. Nagy, Charley M. Wu
Abstract: The pursuit of general intelligence has traditionally centered on external objectives: an agent's control over its environments or mastery of specific tasks. This external focus, however, can produce specialized agents that lack adaptability. We propose representational empowerment, a new perspective towards a truly agent-centric learning paradigm by moving the locus of control inward. This objective measures an agent's ability to controllably maintain and diversify its own knowledge structures. We posit that the capacity -- to shape one's own understanding -- is an element for achieving better ``preparedness'' distinct from direct environmental influence. Focusing on internal representations as the main substrate for computing empowerment offers a new lens through which to design adaptable intelligent systems.
Authors: Shaoan Xie, Lingjing Kong, Yujia Zheng, Yu Yao, Zeyu Tang, Eric P. Xing, Guangyi Chen, Kun Zhang
Abstract: Contrastive Language-Image Pre-training (CLIP)~\citep{radford2021learning} has emerged as a pivotal model in computer vision and multimodal learning, achieving state-of-the-art performance at aligning visual and textual representations through contrastive learning. However, CLIP struggles with potential information misalignment in many image-text datasets and suffers from entangled representation. On the one hand, short captions for a single image in datasets like MSCOCO may describe disjoint regions in the image, leaving the model uncertain about which visual features to retain or disregard. On the other hand, directly aligning long captions with images can lead to the retention of entangled details, preventing the model from learning disentangled, atomic concepts -- ultimately limiting its generalization on certain downstream tasks involving short prompts. In this paper, we establish theoretical conditions that enable flexible alignment between textual and visual representations across varying levels of granularity. Specifically, our framework ensures that a model can not only \emph{preserve} cross-modal semantic information in its entirety but also \emph{disentangle} visual representations to capture fine-grained textual concepts. Building on this foundation, we introduce \ours, a novel approach that identifies and aligns the most relevant visual and textual representations in a modular manner. Superior performance across various tasks demonstrates its capability to handle information misalignment and supports our identification theory. The code is available at https://github.com/Mid-Push/SmartCLIP.
Authors: Owen Hoffman, Kangze Peng, Zehua You, Sajid Kamal, Sukrit Venkatagiri
Abstract: Generative AI, including large language models (LLMs) have the potential -- and already are being used -- to increase the speed, scale, and types of unsafe conversations online. LLMs lower the barrier for entry for bad actors to create unsafe conversations in particular because of their ability to generate persuasive and human-like text. In our current work, we explore ways to promote online safety by teaching people about unsafe conversations that can occur online with and without LLMs. We build on prior work that shows that LLMs can successfully simulate scam conversations. We also leverage research in the learning sciences that shows that providing feedback on one's hypothetical actions can promote learning. In particular, we focus on simulating scam conversations using LLMs. Our work incorporates two LLMs that converse with each other to simulate realistic, unsafe conversations that people may encounter online between a scammer LLM and a target LLM but users of our system are asked provide feedback to the target LLM.
Authors: Junting Wang, Chenghuan Guo, Jiao Yang, Yanhui Guo, Yan Gao, Hari Sundaram
Abstract: We introduce a novel self-supervised multi-modal relational item representation learning framework designed to infer substitutable and complementary items. Existing approaches primarily focus on modeling item-item associations deduced from user behaviors using graph neural networks (GNNs) or leveraging item content information. However, these methods often overlook critical challenges, such as noisy user behavior data and data sparsity due to the long-tailed distribution of these behaviors. In this paper, we propose MMSC, a self-supervised multi-modal relational item representation learning framework to address these challenges. Specifically, MMSC consists of three main components: (1) a multi-modal item representation learning module that leverages a multi-modal foundational model and learns from item metadata, (2) a self-supervised behavior-based representation learning module that denoises and learns from user behavior data, and (3) a hierarchical representation aggregation mechanism that integrates item representations at both the semantic and task levels. Additionally, we leverage LLMs to generate augmented training data, further enhancing the denoising process during training. We conduct extensive experiments on five real-world datasets, showing that MMSC outperforms existing baselines by 26.1% for substitutable recommendation and 39.2% for complementary recommendation. In addition, we empirically show that MMSC is effective in modeling cold-start items.
Authors: Supantho Rakshit, Adele Goldberg
Abstract: The usage-based constructionist (UCx) approach posits that language comprises a network of learned form-meaning pairings (constructions) whose use is largely determined by their meanings or functions, requiring them to be graded and probabilistic. This study investigates whether the internal representations in Large Language Models (LLMs) reflect the proposed function-infused gradience. We analyze the neural representations of the English dative constructions (Double Object and Prepositional Object) in Pythia-$1.4$B, using a dataset of $5000$ sentence pairs systematically varied for human-rated preference strength. A macro-level geometric analysis finds that the separability between construction representations, as measured by Energy Distance or Jensen-Shannon Divergence, is systematically modulated by gradient preference strength. More prototypical exemplars of each construction occupy more distinct regions in the activation space of LLMs. These results provide strong evidence that LLMs learn rich, meaning-infused, graded representations of constructions and offer support for geometric measures of basic constructionist principles in LLMs.
Authors: Ze Zhang, Qian Dong, Wenhan Wang
Abstract: The accurate localization of sensor nodes is a fundamental requirement for the practical application of the Internet of Things (IoT). To enable robust localization across diverse environments, this paper proposes a hybrid meta-heuristic localization algorithm. Specifically, the algorithm integrates the Sine Cosine Algorithm (SCA), which is effective in global search, with Particle Swarm Optimization (PSO), which excels at local search. An adaptive switching module is introduced to dynamically select between the two algorithms. Furthermore, the initialization, fitness evaluation, and parameter settings of the algorithm have been specifically redesigned and optimized to address the characteristics of the node localization problem. Simulation results across varying numbers of sensor nodes demonstrate that, compared to standalone PSO and the unoptimized SCAPSO algorithm, the proposed method significantly reduces the number of required iterations and achieves an average localization error reduction of 84.97%.
Authors: Yuzhen Gao, Qianqian Wang, Yongheng Sun, Cui Wang, Yongquan Liang, Mingxia Liu
Abstract: Accurate identification of late-life depression (LLD) using structural brain MRI is essential for monitoring disease progression and facilitating timely intervention. However, existing learning-based approaches for LLD detection are often constrained by limited sample sizes (e.g., tens), which poses significant challenges for reliable model training and generalization. Although incorporating auxiliary datasets can expand the training set, substantial domain heterogeneity, such as differences in imaging protocols, scanner hardware, and population demographics, often undermines cross-domain transferability. To address this issue, we propose a Collaborative Domain Adaptation (CDA) framework for LLD detection using T1-weighted MRIs. The CDA leverages a Vision Transformer (ViT) to capture global anatomical context and a Convolutional Neural Network (CNN) to extract local structural features, with each branch comprising an encoder and a classifier. The CDA framework consists of three stages: (a) supervised training on labeled source data, (b) self-supervised target feature adaptation and (c) collaborative training on unlabeled target data. We first train ViT and CNN on source data, followed by self-supervised target feature adaptation by minimizing the discrepancy between classifier outputs from two branches to make the categorical boundary clearer. The collaborative training stage employs pseudo-labeled and augmented target-domain MRIs, enforcing prediction consistency under strong and weak augmentation to enhance domain robustness and generalization. Extensive experiments conducted on multi-site T1-weighted MRI data demonstrate that the CDA consistently outperforms state-of-the-art unsupervised domain adaptation methods.
Authors: Cameron S. Movassaghi, Amanda Momenzadeh, Jesse G. Meyer
Abstract: Maintaining software packages imposes significant costs due to dependency management, bug fixes, and versioning. We show that rich method descriptions in scientific publications can serve as standalone specifications for modern large language models (LLMs), enabling on-demand code generation that could supplant human-maintained libraries. We benchmark state-of-the-art models (GPT-o4-mini-high, Gemini Pro 2.5, Claude Sonnet 4) by tasking them with implementing a diverse set of core algorithms drawn from original publications. Our results demonstrate that current LLMs can reliably reproduce package functionality with performance indistinguishable from conventional libraries. These findings foreshadow a paradigm shift toward flexible, on-demand code generation and away from static, human-maintained packages, which will result in reduced maintenance overhead by leveraging published articles as sufficient context for the automated implementation of analytical workflows.
Authors: Kunyang Li, Jeffrey A Chan Santiago, Sarinda Dhanesh Samarasinghe, Gaowen Liu, Mubarak Shah
Abstract: To address the larger computation and storage requirements associated with large video datasets, video dataset distillation aims to capture spatial and temporal information in a significantly smaller dataset, such that training on the distilled data has comparable performance to training on all of the data. We propose GVD: Guiding Video Diffusion, the first diffusion-based video distillation method. GVD jointly distills spatial and temporal features, ensuring high-fidelity video generation across diverse actions while capturing essential motion information. Our method's diverse yet representative distillations significantly outperform previous state-of-the-art approaches on the MiniUCF and HMDB51 datasets across 5, 10, and 20 Instances Per Class (IPC). Specifically, our method achieves 78.29 percent of the original dataset's performance using only 1.98 percent of the total number of frames in MiniUCF. Additionally, it reaches 73.83 percent of the performance with just 3.30 percent of the frames in HMDB51. Experimental results across benchmark video datasets demonstrate that GVD not only achieves state-of-the-art performance but can also generate higher resolution videos and higher IPC without significantly increasing computational cost.
Authors: Aria Salari, Abtin Djavadifar, Xiangrui Liu, Homayoun Najjaran
Abstract: Object recognition is among the fundamental tasks in the computer vision applications, paving the path for all other image understanding operations. In every stage of progress in object recognition research, efforts have been made to collect and annotate new datasets to match the capacity of the state-of-the-art algorithms. In recent years, the importance of the size and quality of datasets has been intensified as the utility of the emerging deep network techniques heavily relies on training data. Furthermore, datasets lay a fair benchmarking means for competitions and have proved instrumental to the advancements of object recognition research by providing quantifiable benchmarks for the developed models. Taking a closer look at the characteristics of commonly-used public datasets seems to be an important first step for data-driven and machine learning researchers. In this survey, we provide a detailed analysis of datasets in the highly investigated object recognition areas. More than 160 datasets have been scrutinized through statistics and descriptions. Additionally, we present an overview of the prominent object recognition benchmarks and competitions, along with a description of the metrics widely adopted for evaluation purposes in the computer vision community. All introduced datasets and challenges can be found online at github.com/AbtinDjavadifar/ORDC.
Authors: Sinh Trong Vu, Hieu Trung Pham, Dung Manh Nguyen, Hieu Minh Hoang, Nhu Hoang Le, Thu Ha Pham, Tai Tan Mai
Abstract: Classroom behavior monitoring is a critical aspect of educational research, with significant implications for student engagement and learning outcomes. Recent advancements in Visual Question Answering (VQA) models offer promising tools for automatically analyzing complex classroom interactions from video recordings. In this paper, we investigate the applicability of several state-of-the-art open-source VQA models, including LLaMA2, LLaMA3, QWEN3, and NVILA, in the context of classroom behavior analysis. To facilitate rigorous evaluation, we introduce our BAV-Classroom-VQA dataset derived from real-world classroom video recordings at the Banking Academy of Vietnam. We present the methodology for data collection, annotation, and benchmark the performance of the selected VQA models on this dataset. Our initial experimental results demonstrate that all four models achieve promising performance levels in answering behavior-related visual questions, showcasing their potential in future classroom analytics and intervention systems.
Authors: Lei Yu, Shiqi Cheng, Zhirong Huang, Jingyuan Zhang, Chenjie Shen, Junyi Lu, Li Yang, Fengjun Zhang, Jiajia Ma
Abstract: With the increasing security issues in blockchain, smart contract vulnerability detection has become a research focus. Existing vulnerability detection methods have their limitations: 1) Static analysis methods struggle with complex scenarios. 2) Methods based on specialized pre-trained models perform well on specific datasets but have limited generalization capabilities. In contrast, general-purpose Large Language Models (LLMs) demonstrate impressive ability in adapting to new vulnerability patterns. However, they often underperform on specific vulnerability types compared to methods based on specialized pre-trained models. We also observe that explanations generated by general-purpose LLMs can provide fine-grained code understanding information, contributing to improved detection performance. Inspired by these observations, we propose SAEL, an LLM-based framework for smart contract vulnerability detection. We first design targeted prompts to guide LLMs in identifying vulnerabilities and generating explanations, which serve as prediction features. Next, we apply prompt-tuning on CodeT5 and T5 to process contract code and explanations, enhancing task-specific performance. To combine the strengths of each approach, we introduce an Adaptive Mixture-of-Experts architecture. This dynamically adjusts feature weights via a Gating Network, which selects relevant features using TopK filtering and Softmax normalization, and incorporates a Multi-Head Self-Attention mechanism to enhance cross-feature relationships. This design enables effective integration of LLM predictions, explanation features, and code features through gradient optimization. The loss function jointly considers both independent feature performance and overall weighted predictions. Experiments show that SAEL outperforms existing methods across various vulnerabilities.
Authors: Sua Lee, Joonhun Lee, Myungjoo Kang
Abstract: Self-supervised learning methods like masked autoencoders (MAE) have shown significant promise in learning robust feature representations, particularly in image reconstruction-based pretraining task. However, their performance is often strongly dependent on the masking strategies used during training and can degrade when applied to out-of-distribution data. To address these limitations, we introduce the masked implicit neural representations (MINR) framework that synergizes implicit neural representations with masked image modeling. MINR learns a continuous function to represent images, enabling more robust and generalizable reconstructions irrespective of masking strategies. Our experiments demonstrate that MINR not only outperforms MAE in in-domain scenarios but also in out-of-distribution settings, while reducing model complexity. The versatility of MINR extends to various self-supervised learning applications, confirming its utility as a robust and efficient alternative to existing frameworks.
Authors: Xiaocheng Yang, Sumuk Shashidhar, Dilek Hakkani-Tur
Abstract: Assessment of reading comprehension through content-based interactions plays an important role in the reading acquisition process. In this paper, we propose a novel approach for generating comprehension questions geared to K-2 English learners. Our method ensures complete coverage of the underlying material and adaptation to the learner's specific proficiencies, and can generate a large diversity of question types at various difficulty levels to ensure a thorough evaluation. We evaluate the performance of various language models in this framework using the FairytaleQA dataset as the source material. Eventually, the proposed approach has the potential to become an important part of autonomous AI-driven English instructors.
Authors: Hyeonseok Moon, Heuiseok Lim
Abstract: The Needle-in-a-Haystack (NIAH) benchmark is widely used to evaluate Large Language Models' (LLMs) ability to understand long contexts (LC). It evaluates the capability to identify query-relevant context within extensive query-irrelevant passages. Although this method serves as a widely accepted standard for evaluating long-context understanding, our findings suggest it may overestimate the true LC capability of LLMs. We demonstrate that even state-of-the-art models such as GPT-4o struggle to intactly incorporate given contexts made up of solely query-relevant ten sentences. In response, we introduce a novel benchmark, \textbf{NeedleChain}, where the context consists entirely of query-relevant information, requiring the LLM to fully grasp the input to answer correctly. Our benchmark allows for flexible context length and reasoning order, offering a more comprehensive analysis of LLM performance. Additionally, we propose an extremely simple yet compelling strategy to improve LC understanding capability of LLM: ROPE Contraction. Our experiments with various advanced LLMs reveal a notable disparity between their ability to process large contexts and their capacity to fully understand them. Source code and datasets are available at https://github.com/hyeonseokk/NeedleChain
Authors: Phi Van Nguyen, Ngoc Huynh Trinh, Duy Minh Lam Nguyen, Phu Loc Nguyen, Quoc Long Tran
Abstract: Quantifying aleatoric uncertainty in medical image segmentation is critical since it is a reflection of the natural variability observed among expert annotators. A conventional approach is to model the segmentation distribution using the generative model, but current methods limit the expression ability of generative models. While current diffusion-based approaches have demonstrated impressive performance in approximating the data distribution, their inherent stochastic sampling process and inability to model exact densities limit their effectiveness in accurately capturing uncertainty. In contrast, our proposed method leverages conditional flow matching, a simulation-free flow-based generative model that learns an exact density, to produce highly accurate segmentation results. By guiding the flow model on the input image and sampling multiple data points, our approach synthesizes segmentation samples whose pixel-wise variance reliably reflects the underlying data distribution. This sampling strategy captures uncertainties in regions with ambiguous boundaries, offering robust quantification that mirrors inter-annotator differences. Experimental results demonstrate that our method not only achieves competitive segmentation accuracy but also generates uncertainty maps that provide deeper insights into the reliability of the segmentation outcomes. The code for this paper is freely available at https://github.com/huynhspm/Data-Uncertainty
Authors: Tung-Wei Lin, Gabe Fierro, Han Li, Tianzhen Hong, Pierluigi Nuzzo, Alberto Sangiovanni-Vinentelli
Abstract: We present a systematic approach for evaluating the quality of knowledge graph repairs with respect to constraint violations defined in shapes constraint language (SHACL). Current evaluation methods rely on \emph{ad hoc} datasets, which limits the rigorous analysis of repair systems in more general settings. Our method addresses this gap by systematically generating violations using a novel mechanism, termed violation-inducing operations (VIOs). We use the proposed evaluation framework to assess a range of repair systems which we build using large language models. We analyze the performance of these systems across different prompting strategies. Results indicate that concise prompts containing both the relevant violated SHACL constraints and key contextual information from the knowledge graph yield the best performance.
Authors: Shahla John
Abstract: Real-time video analysis remains a challenging problem in computer vision, requiring efficient processing of both spatial and temporal information while maintaining computational efficiency. Existing approaches often struggle to balance accuracy and speed, particularly in resource-constrained environments. In this work, we present a unified framework that leverages advanced spatial-temporal modeling techniques for simultaneous action recognition and object tracking. Our approach builds upon recent advances in parallel sequence modeling and introduces a novel hierarchical attention mechanism that adaptively focuses on relevant spatial regions across temporal sequences. We demonstrate that our method achieves state-of-the-art performance on standard benchmarks while maintaining real-time inference speeds. Extensive experiments on UCF-101, HMDB-51, and MOT17 datasets show improvements of 3.2% in action recognition accuracy and 2.8% in tracking precision compared to existing methods, with 40% faster inference time.
Authors: Songsheng Wang, Rucheng Yu, Zhihang Yuan, Chao Yu, Feng Gao, Yu Wang, Derek F. Wong
Abstract: Vision-Language-Action (VLA) models have made substantial progress by leveraging the robust capabilities of Visual Language Models (VLMs). However, VLMs' significant parameter size and autoregressive (AR) decoding nature impose considerable computational demands on VLA models. While Speculative Decoding (SD) has shown efficacy in accelerating Large Language Models (LLMs) by incorporating efficient drafting and parallel verification, allowing multiple tokens to be generated in one forward pass, its application to VLA models remains unexplored. This work introduces Spec-VLA, an SD framework designed to accelerate VLA models. Due to the difficulty of the action prediction task and the greedy decoding mechanism of the VLA models, the direct application of the advanced SD framework to the VLA prediction task yields a minor speed improvement. To boost the generation speed, we propose an effective mechanism to relax acceptance utilizing the relative distances represented by the action tokens of the VLA model. Empirical results across diverse test scenarios affirm the effectiveness of the Spec-VLA framework, and further analysis substantiates the impact of our proposed strategies, which enhance the acceptance length by 44%, achieving 1.42 times speedup compared with the OpenVLA baseline, without compromising the success rate. The success of the Spec-VLA framework highlights the potential for broader application of speculative execution in VLA prediction scenarios.
Authors: Yunrui Yu, Hang Su, Cheng-zhong Xu, Zhizhong Su, Jun Zhu
Abstract: Gradient-based adversarial attacks using the Cross-Entropy (CE) loss often suffer from overestimation due to relative errors in gradient computation induced by floating-point arithmetic. This paper provides a rigorous theoretical analysis of these errors, conducting the first comprehensive study of floating-point computation errors in gradient-based attacks across four distinct scenarios: (i) unsuccessful untargeted attacks, (ii) successful untargeted attacks, (iii) unsuccessful targeted attacks, and (iv) successful targeted attacks. We establish theoretical foundations characterizing the behavior of relative numerical errors under different attack conditions, revealing previously unknown patterns in gradient computation instability, and identify floating-point underflow and rounding as key contributors. Building on this insight, we propose the Theoretical MIFPE (T-MIFPE) loss function, which incorporates an optimal scaling factor $T = t^*$ to minimize the impact of floating-point errors, thereby enhancing the accuracy of gradient computation in adversarial attacks. Extensive experiments on the MNIST, CIFAR-10, and CIFAR-100 datasets demonstrate that T-MIFPE outperforms existing loss functions, including CE, C\&W, DLR, and MIFPE, in terms of attack potency and robustness evaluation accuracy.
Authors: Jill Walker Rettberg, Hermann Wigers
Abstract: Can a language model trained largely on Anglo-American texts generate stories that are culturally relevant to other nationalities? To find out, we generated 11,800 stories - 50 for each of 236 countries - by sending the prompt "Write a 1500 word potential {demonym} story" to OpenAI's model gpt-4o-mini. Although the stories do include surface-level national symbols and themes, they overwhelmingly conform to a single narrative plot structure across countries: a protagonist lives in or returns home to a small town and resolves a minor conflict by reconnecting with tradition and organising community events. Real-world conflicts are sanitised, romance is almost absent, and narrative tension is downplayed in favour of nostalgia and reconciliation. The result is a narrative homogenisation: an AI-generated synthetic imaginary that prioritises stability above change and tradition above growth. We argue that the structural homogeneity of AI-generated narratives constitutes a distinct form of AI bias, a narrative standardisation that should be acknowledged alongside the more familiar representational bias. These findings are relevant to literary studies, narratology, critical AI studies, NLP research, and efforts to improve the cultural alignment of generative AI.
Authors: Yunrui Yu, Kafeng Wang, Hang Su, Jun Zhu
Abstract: Despite their widespread success, deep neural networks remain critically vulnerable to adversarial attacks, posing significant risks in safety-sensitive applications. This paper investigates activation functions as a crucial yet underexplored component for enhancing model robustness. We propose a Rademacher Complexity Reduction Activation Function (RCR-AF), a novel activation function designed to improve both generalization and adversarial resilience. RCR-AF uniquely combines the advantages of GELU (including smoothness, gradient stability, and negative information retention) with ReLU's desirable monotonicity, while simultaneously controlling both model sparsity and capacity through built-in clipping mechanisms governed by two hyperparameters, $\alpha$ and $\gamma$. Our theoretical analysis, grounded in Rademacher complexity, demonstrates that these parameters directly modulate the model's Rademacher complexity, offering a principled approach to enhance robustness. Comprehensive empirical evaluations show that RCR-AF consistently outperforms widely-used alternatives (ReLU, GELU, and Swish) in both clean accuracy under standard training and in adversarial robustness within adversarial training paradigms.
Authors: Tian Yun, Chen Sun, Ellie Pavlick
Abstract: Recent work has argued that large language models (LLMs) are not "abstract reasoners", citing their poor zero-shot performance on a variety of challenging tasks as evidence. We revisit these experiments in order to add nuance to the claim. First, we show that while LLMs indeed perform poorly in a zero-shot setting, even tuning a small subset of parameters for input encoding can enable near-perfect performance. However, we also show that this finetuning does not necessarily transfer across datasets. We take this collection of empirical results as an invitation to (re-)open the discussion of what it means to be an "abstract reasoner", and why it matters whether LLMs fit the bill.
Authors: Peng-Yi Wu, Pei-Cing Huang, Ting-Yu Chen, Chantung Ku, Ming-Yen Lin, Yihuang Kang
Abstract: Accurate and interpretable prediction of estimated glomerular filtration rate (eGFR) is essential for managing chronic kidney disease (CKD) and supporting clinical decisions. Recent advances in Large Multimodal Models (LMMs) have shown strong potential in clinical prediction tasks due to their ability to process visual and textual information. However, challenges related to deployment cost, data privacy, and model reliability hinder their adoption. In this study, we propose a collaborative framework that enhances the performance of open-source LMMs for eGFR forecasting while generating clinically meaningful explanations. The framework incorporates visual knowledge transfer, abductive reasoning, and a short-term memory mechanism to enhance prediction accuracy and interpretability. Experimental results show that the proposed framework achieves predictive performance and interpretability comparable to proprietary models. It also provides plausible clinical reasoning processes behind each prediction. Our method sheds new light on building AI systems for healthcare that combine predictive accuracy with clinically grounded interpretability.
Authors: Zheng Xiangyu, He Songcheng, Li Wanyun, Li Xiaoqiang, Zhang Wei
Abstract: Unsupervised Video Object Segmentation (UVOS) aims to predict pixel-level masks for the most salient objects in videos without any prior annotations. While memory mechanisms have been proven critical in various video segmentation paradigms, their application in UVOS yield only marginal performance gains despite sophisticated design. Our analysis reveals a simple but fundamental flaw in existing methods: over-reliance on memorizing high-level semantic features. UVOS inherently suffers from the deficiency of lacking fine-grained information due to the absence of pixel-level prior knowledge. Consequently, memory design relying solely on high-level features, which predominantly capture abstract semantic cues, is insufficient to generate precise predictions. To resolve this fundamental issue, we propose a novel hierarchical memory architecture to incorporate both shallow- and high-level features for memory, which leverages the complementary benefits of pixel and semantic information. Furthermore, to balance the simultaneous utilization of the pixel and semantic memory features, we propose a heterogeneous interaction mechanism to perform pixel-semantic mutual interactions, which explicitly considers their inherent feature discrepancies. Through the design of Pixel-guided Local Alignment Module (PLAM) and Semantic-guided Global Integration Module (SGIM), we achieve delicate integration of the fine-grained details in shallow-level memory and the semantic representations in high-level memory. Our Hierarchical Memory with Heterogeneous Interaction Network (HMHI-Net) consistently achieves state-of-the-art performance across all UVOS and video saliency detection benchmarks. Moreover, HMHI-Net consistently exhibits high performance across different backbones, further demonstrating its superiority and robustness. Project page: https://github.com/ZhengxyFlow/HMHI-Net .
Authors: Hsien-Tsung Lin, Pei-Cing Huang, Chan-Tung Ku, Chan Hsu, Pei-Xuan Shieh, Yihuang Kang
Abstract: Recent advancements in Large Language Models offer promising capabilities to simulate complex human social interactions. We investigate whether LLM-based multi-agent simulations can reproduce core human social dynamics observed in online forums. We evaluate conformity dynamics, group polarization, and fragmentation across different model scales and reasoning capabilities using a structured simulation framework. Our findings indicate that smaller models exhibit higher conformity rates, whereas models optimized for reasoning are more resistant to social influence.
Authors: Viacheslav Pirogov
Abstract: The contemporary phenomenon of deepfakes, utilizing GAN or diffusion models for face swapping, presents a substantial and evolving threat in digital media, identity verification, and a multitude of other systems. The majority of existing methods for detecting deepfakes rely on training specialized classifiers to distinguish between genuine and manipulated images, focusing only on the image domain without incorporating any auxiliary tasks that could enhance robustness. In this paper, inspired by the zero-shot capabilities of Vision Language Models, we propose a novel VLM-based approach to image classification and then evaluate it for deepfake detection. Specifically, we utilize a new high-quality deepfake dataset comprising 60,000 images, on which our zero-shot models demonstrate superior performance to almost all existing methods. Subsequently, we compare the performance of the best-performing architecture, InstructBLIP, on the popular deepfake dataset DFDC-P against traditional methods in two scenarios: zero-shot and in-domain fine-tuning. Our results demonstrate the superiority of VLMs over traditional classifiers.
Authors: Hui Liu, Chen Jia, Fan Shi, Xu Cheng, Mengfei Shi, Xia Xie, Shengyong Chen
Abstract: Achieving pixel-level segmentation with low computational cost using multimodal data remains a key challenge in crack segmentation tasks. Existing methods lack the capability for adaptive perception and efficient interactive fusion of cross-modal features. To address these challenges, we propose a Lightweight Adaptive Cue-Aware Vision Mamba network (LIDAR), which efficiently perceives and integrates morphological and textural cues from different modalities under multimodal crack scenarios, generating clear pixel-level crack segmentation maps. Specifically, LIDAR is composed of a Lightweight Adaptive Cue-Aware Visual State Space module (LacaVSS) and a Lightweight Dual Domain Dynamic Collaborative Fusion module (LD3CF). LacaVSS adaptively models crack cues through the proposed mask-guided Efficient Dynamic Guided Scanning Strategy (EDG-SS), while LD3CF leverages an Adaptive Frequency Domain Perceptron (AFDP) and a dual-pooling fusion strategy to effectively capture spatial and frequency-domain cues across modalities. Moreover, we design a Lightweight Dynamically Modulated Multi-Kernel convolution (LDMK) to perceive complex morphological structures with minimal computational overhead, replacing most convolutional operations in LIDAR. Experiments on three datasets demonstrate that our method outperforms other state-of-the-art (SOTA) methods. On the light-field depth dataset, our method achieves 0.8204 in F1 and 0.8465 in mIoU with only 5.35M parameters. Code and datasets are available at https://github.com/Karl1109/LIDAR-Mamba.
Authors: Tianyi Liu, Kejun Wu, Chen Cai, Yi Wang, Kim-Hui Yap, Lap-Pui Chau
Abstract: Video signals are vulnerable in multimedia communication and storage systems, as even slight bitstream-domain corruption can lead to significant pixel-domain degradation. To recover faithful spatio-temporal content from corrupted inputs, bitstream-corrupted video recovery has recently emerged as a challenging and understudied task. However, existing methods require time-consuming and labor-intensive annotation of corrupted regions for each corrupted video frame, resulting in a large workload in practice. In addition, high-quality recovery remains difficult as part of the local residual information in corrupted frames may mislead feature completion and successive content recovery. In this paper, we propose the first blind bitstream-corrupted video recovery framework that integrates visual foundation models with a recovery model, which is adapted to different types of corruption and bitstream-level prompts. Within the framework, the proposed Detect Any Corruption (DAC) model leverages the rich priors of the visual foundation model while incorporating bitstream and corruption knowledge to enhance corruption localization and blind recovery. Additionally, we introduce a novel Corruption-aware Feature Completion (CFC) module, which adaptively processes residual contributions based on high-level corruption understanding. With VFM-guided hierarchical feature augmentation and high-level coordination in a mixture-of-residual-experts (MoRE) structure, our method suppresses artifacts and enhances informative residuals. Comprehensive evaluations show that the proposed method achieves outstanding performance in bitstream-corrupted video recovery without requiring a manually labeled mask sequence. The demonstrated effectiveness will help to realize improved user experience, wider application scenarios, and more reliable multimedia communication and storage systems.
Authors: Nils Bochow, Philipp Hess, Alexander Robinson
Abstract: Accurate, high-resolution projections of the Greenland ice sheet's surface mass balance (SMB) and surface temperature are essential for understanding future sea-level rise, yet current approaches are either computationally demanding or limited to coarse spatial scales. Here, we introduce a novel physics-constrained generative modeling framework based on a consistency model (CM) to downscale low-resolution SMB and surface temperature fields by a factor of up to 32 (from 160 km to 5 km grid spacing) in a few sampling steps. The CM is trained on monthly outputs of the regional climate model MARv3.12 and conditioned on ice-sheet topography and insolation. By enforcing a hard conservation constraint during inference, we ensure approximate preservation of SMB and temperature sums on the coarse spatial scale as well as robust generalization to extreme climate states without retraining. On the test set, our constrained CM achieves a continued ranked probability score of 6.31 mmWE for the SMB and 0.1 K for the surface temperature, outperforming interpolation-based downscaling. Together with spatial power-spectral analysis, we demonstrate that the CM faithfully reproduces variability across spatial scales. We further apply bias-corrected outputs of the NorESM2 Earth System Model as inputs to our CM, to demonstrate the potential of our model to directly downscale ESM fields. Our approach delivers realistic, high-resolution climate forcing for ice-sheet simulations with fast inference and can be readily integrated into Earth-system and ice-sheet model workflows to improve projections of the future contribution to sea-level rise from Greenland and potentially other ice sheets and glaciers too.
Authors: Wei Guo, Yiyang Duan, Zhaojun Hu, Yiqi Tong, Fuzhen Zhuang, Xiao Zhang, Jin Dong, Ruofan Wu, Tengfei Liu, Yifan Sun
Abstract: In vertical federated learning (VFL), multiple enterprises address aligned sample scarcity by leveraging massive locally unaligned samples to facilitate collaborative learning. However, unaligned samples across different parties in VFL can be extremely class-imbalanced, leading to insufficient feature representation and limited model prediction space. Specifically, class-imbalanced problems consist of intra-party class imbalance and inter-party class imbalance, which can further cause local model bias and feature contribution inconsistency issues, respectively. To address the above challenges, we propose Proto-EVFL, an enhanced VFL framework via dual prototypes. We first introduce class prototypes for each party to learn relationships between classes in the latent space, allowing the active party to predict unseen classes. We further design a probabilistic dual prototype learning scheme to dynamically select unaligned samples by conditional optimal transport cost with class prior probability. Moreover, a mixed prior guided module guides this selection process by combining local and global class prior probabilities. Finally, we adopt an \textit{adaptive gated feature aggregation strategy} to mitigate feature contribution inconsistency by dynamically weighting and aggregating local features across different parties. We proved that Proto-EVFL, as the first bi-level optimization framework in VFL, has a convergence rate of 1/\sqrt T. Extensive experiments on various datasets validate the superiority of our Proto-EVFL. Even in a zero-shot scenario with one unseen class, it outperforms baselines by at least 6.97%
Authors: Xiaodong Feng, Ling Guo, Xiaoliang Wan, Hao Wu, Tao Zhou, Wenwen Zhou
Abstract: We propose a novel probabilistic framework, termed LVM-GP, for uncertainty quantification in solving forward and inverse partial differential equations (PDEs) with noisy data. The core idea is to construct a stochastic mapping from the input to a high-dimensional latent representation, enabling uncertainty-aware prediction of the solution. Specifically, the architecture consists of a confidence-aware encoder and a probabilistic decoder. The encoder implements a high-dimensional latent variable model based on a Gaussian process (LVM-GP), where the latent representation is constructed by interpolating between a learnable deterministic feature and a Gaussian process prior, with the interpolation strength adaptively controlled by a confidence function learned from data. The decoder defines a conditional Gaussian distribution over the solution field, where the mean is predicted by a neural operator applied to the latent representation, allowing the model to learn flexible function-to-function mapping. Moreover, physical laws are enforced as soft constraints in the loss function to ensure consistency with the underlying PDE structure. Compared to existing approaches such as Bayesian physics-informed neural networks (B-PINNs) and deep ensembles, the proposed framework can efficiently capture functional dependencies via merging a latent Gaussian process and neural operator, resulting in competitive predictive accuracy and robust uncertainty quantification. Numerical experiments demonstrate the effectiveness and reliability of the method.
Authors: Yuhwan Jeong, Yunseo Yang, Youngho Yoon, Kuk-Jin Yoon
Abstract: Adverse weather conditions cause diverse and complex degradation patterns, driving the development of All-in-One (AiO) models. However, recent AiO solutions still struggle to capture diverse degradations, since global filtering methods like direct operations on the frequency domain fail to handle highly variable and localized distortions. To address these issue, we propose Spectral-based Spatial Grouping Transformer (SSGformer), a novel approach that leverages spectral decomposition and group-wise attention for multi-weather image restoration. SSGformer decomposes images into high-frequency edge features using conventional edge detection and low-frequency information via Singular Value Decomposition. We utilize multi-head linear attention to effectively model the relationship between these features. The fused features are integrated with the input to generate a grouping-mask that clusters regions based on the spatial similarity and image texture. To fully leverage this mask, we introduce a group-wise attention mechanism, enabling robust adverse weather removal and ensuring consistent performance across diverse weather conditions. We also propose a Spatial Grouping Transformer Block that uses both channel attention and spatial attention, effectively balancing feature-wise relationships and spatial dependencies. Extensive experiments show the superiority of our approach, validating its effectiveness in handling the varied and intricate adverse weather degradations.
Authors: Xiang Li, Qianli Shen, Haonan Wang, Kenji Kawaguchi
Abstract: Recent generative models face significant risks of producing harmful content, which has underscored the importance of machine unlearning (MU) as a critical technique for eliminating the influence of undesired data. However, existing MU methods typically assign the same weight to all data to be forgotten, which makes it difficult to effectively forget certain data that is harder to unlearn than others. In this paper, we empirically demonstrate that the loss of data itself can implicitly reflect its varying difficulty. Building on this insight, we introduce Loss-based Reweighting Unlearning (LoReUn), a simple yet effective plug-and-play strategy that dynamically reweights data during the unlearning process with minimal additional computational overhead. Our approach significantly reduces the gap between existing MU methods and exact unlearning in both image classification and generation tasks, effectively enhancing the prevention of harmful content generation in text-to-image diffusion models.
Authors: Ziyi Wang, Peiming Li, Hong Liu, Zhichao Deng, Can Wang, Jun Liu, Junsong Yuan, Mengyuan Liu
Abstract: Natural Human-Robot Interaction (N-HRI) requires robots to recognize human actions at varying distances and states, regardless of whether the robot itself is in motion or stationary. This setup is more flexible and practical than conventional human action recognition tasks. However, existing benchmarks designed for traditional action recognition fail to address the unique complexities in N-HRI due to limited data, modalities, task categories, and diversity of subjects and environments. To address these challenges, we introduce ACTIVE (Action from Robotic View), a large-scale dataset tailored specifically for perception-centric robotic views prevalent in mobile service robots. ACTIVE comprises 30 composite action categories, 80 participants, and 46,868 annotated video instances, covering both RGB and point cloud modalities. Participants performed various human actions in diverse environments at distances ranging from 3m to 50m, while the camera platform was also mobile, simulating real-world scenarios of robot perception with varying camera heights due to uneven ground. This comprehensive and challenging benchmark aims to advance action and attribute recognition research in N-HRI. Furthermore, we propose ACTIVE-PC, a method that accurately perceives human actions at long distances using Multilevel Neighborhood Sampling, Layered Recognizers, Elastic Ellipse Query, and precise decoupling of kinematic interference from human actions. Experimental results demonstrate the effectiveness of ACTIVE-PC. Our code is available at: https://github.com/wangzy01/ACTIVE-Action-from-Robotic-View.
URLs: https://github.com/wangzy01/ACTIVE-Action-from-Robotic-View.
Authors: Kranthi Kumar Talluri, Galia Weidl, Vaishnavi Kasuluru
Abstract: Traffic congestion due to uncertainties, such as accidents, is a significant issue in urban areas, as the ripple effect of accidents causes longer delays, increased emissions, and safety concerns. To address this issue, we propose a robust framework for predicting the impact of accidents on congestion. We implement Automated Machine Learning (AutoML)-enhanced Deep Embedding Clustering (DEC) to assign congestion labels to accident data and predict congestion probability using a Bayesian Network (BN). The Simulation of Urban Mobility (SUMO) simulation is utilized to evaluate the correctness of BN predictions using evidence-based scenarios. Results demonstrate that the AutoML-enhanced DEC has outperformed traditional clustering approaches. The performance of the proposed BN model achieved an overall accuracy of 95.6%, indicating its ability to understand the complex relationship of accidents causing congestion. Validation in SUMO with evidence-based scenarios demonstrated that the BN model's prediction of congestion states closely matches those of SUMO, indicating the high reliability of the proposed BN model in ensuring smooth urban mobility.
Authors: Xincheng Yao, Yijun Yang, Kangwei Guo, Ruiqiang Xiao, Haipeng Zhou, Haisu Tao, Jian Yang, Lei Zhu
Abstract: The segmentation of the hepatic vasculature in surgical videos holds substantial clinical significance in the context of hepatectomy procedures. However, owing to the dearth of an appropriate dataset and the inherently complex task characteristics, few researches have been reported in this domain. To address this issue, we first introduce a high quality frame-by-frame annotated hepatic vasculature dataset containing 35 long hepatectomy videos and 11442 high-resolution frames. On this basis, we propose a novel high-resolution video vasculature segmentation network, dubbed as HRVVS. We innovatively embed a pretrained visual autoregressive modeling (VAR) model into different layers of the hierarchical encoder as prior information to reduce the information degradation generated during the downsampling process. In addition, we designed a dynamic memory decoder on a multi-view segmentation network to minimize the transmission of redundant information while preserving more details between frames. Extensive experiments on surgical video datasets demonstrate that our proposed HRVVS significantly outperforms the state-of-the-art methods. The source code and dataset will be publicly available at \{https://github.com/scott-yjyang/HRVVS}.
Authors: Dongchen Li (College of Computer Science and Engineering, Northeastern University, Shenyang, China), Jitao Liang (College of Computer Science and Engineering, Northeastern University, Shenyang, China), Wei Li (College of Computer Science and Engineering, Northeastern University, Shenyang, China), Xiaoyu Wang (Liaoning Cancer Hospital and Institute, Shenyang, China), Longbing Cao (Macquarie University, Sydney, Australia), Kun Yu (College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, China)
Abstract: Large Language Models (LLMs) hold significant promise for improving clinical decision support and reducing physician burnout by synthesizing complex, longitudinal cancer Electronic Health Records (EHRs). However, their implementation in this critical field faces three primary challenges: the inability to effectively process the extensive length and multilingual nature of patient records for accurate temporal analysis; a heightened risk of clinical hallucination, as conventional grounding techniques such as Retrieval-Augmented Generation (RAG) do not adequately incorporate process-oriented clinical guidelines; and unreliable evaluation metrics that hinder the validation of AI systems in oncology. To address these issues, we propose CliCARE, a framework for Grounding Large Language Models in Clinical Guidelines for Decision Support over Longitudinal Cancer Electronic Health Records. The framework operates by transforming unstructured, longitudinal EHRs into patient-specific Temporal Knowledge Graphs (TKGs) to capture long-range dependencies, and then grounding the decision support process by aligning these real-world patient trajectories with a normative guideline knowledge graph. This approach provides oncologists with evidence-grounded decision support by generating a high-fidelity clinical summary and an actionable recommendation. We validated our framework using large-scale, longitudinal data from a private Chinese cancer dataset and the public English MIMIC-IV dataset. In these diverse settings, CliCARE significantly outperforms strong baselines, including leading long-context LLMs and Knowledge Graph-enhanced RAG methods. The clinical validity of our results is supported by a robust evaluation protocol, which demonstrates a high correlation with assessments made by expert oncologists.
Authors: Matteo Giacomini, Antonio Huerta
Abstract: A surrogate-based topology optimisation algorithm for linear elastic structures under parametric loads and boundary conditions is proposed. Instead of learning the parametric solution of the state (and adjoint) problems or the optimisation trajectory as a function of the iterations, the proposed approach devises a surrogate version of the entire optimisation pipeline. First, the method predicts a quasi-optimal topology for a given problem configuration as a surrogate model of high-fidelity topologies optimised with the homogenisation method. This is achieved by means of a feed-forward net learning the mapping between the input parameters characterising the system setup and a latent space determined by encoder/decoder blocks reducing the dimensionality of the parametric topology optimisation problem and reconstructing a high-dimensional representation of the topology. Then, the predicted topology is used as an educated initial guess for a computationally efficient algorithm penalising the intermediate values of the design variable, while enforcing the governing equations of the system. This step allows the method to correct potential errors introduced by the surrogate model, eliminate artifacts, and refine the design in order to produce topologies consistent with the underlying physics. Different architectures are proposed and the approximation and generalisation capabilities of the resulting models are numerically evaluated. The quasi-optimal topologies allow to outperform the high-fidelity optimiser by reducing the average number of optimisation iterations by $53\%$ while achieving discrepancies below $4\%$ in the optimal value of the objective functional, even in the challenging scenario of testing the model to extrapolate beyond the training and validation domain.
Authors: Kiseong Hong, Gyeong-hyeon Kim, Eunwoo Kim
Abstract: Prompt-based continual learning provides a rehearsal-free solution by tuning small sets of parameters while keeping pre-trained models frozen. To meet the complex demands of sequential tasks, it is crucial to integrate task-specific knowledge within prompts effectively. However, existing works rely on either fixed learned prompts (i.e., prompts whose representations remain unchanged during new task learning) or on prompts generated from an entangled task-shared space, limiting the representational diversity of the integrated prompt. To address this issue, we propose a novel prompt-evolving mechanism to adaptively aggregate base prompts (i.e., task-specific prompts) into a unified prompt while ensuring diversity. By transforming and aligning base prompts, both previously learned and newly introduced, our approach continuously evolves accumulated knowledge to facilitate learning new tasks. We further introduce a learnable probabilistic gate that adaptively determines which layers to activate during the evolution process. We validate our method on image classification and video action recognition tasks in class-incremental learning, achieving average gains of 9.07% and 7.40% over existing methods across all scenarios.
Authors: Yuna Oikawa, Guillaume Deffrennes, Taichi Abe, Ryo Tamura, Koji Tsuda
Abstract: Large Language Models (LLMs) are general-purpose tools with wide-ranging applications, including in materials science. In this work, we introduce aLLoyM, a fine-tuned LLM specifically trained on alloy compositions, temperatures, and their corresponding phase information. To develop aLLoyM, we curated question-and-answer (Q&A) pairs for binary and ternary phase diagrams using the open-source Computational Phase Diagram Database (CPDDB) and assessments based on CALPHAD (CALculation of PHAse Diagrams). We fine-tuned Mistral, an open-source pre-trained LLM, for two distinct Q&A formats: multiple-choice and short-answer. Benchmark evaluations demonstrate that fine-tuning substantially enhances performance on multiple-choice phase diagram questions. Moreover, the short-answer model of aLLoyM exhibits the ability to generate novel phase diagrams from its components alone, underscoring its potential to accelerate the discovery of previously unexplored materials systems. To promote further research and adoption, we have publicly released the short-answer fine-tuned version of aLLoyM, along with the complete benchmarking Q&A dataset, on Hugging Face.
Authors: Xikang Yang, Biyu Zhou, Xuehai Tang, Jizhong Han, Songlin Hu
Abstract: Large Language Models (LLMs) demonstrate impressive capabilities across a wide range of tasks, yet their safety mechanisms remain susceptible to adversarial attacks that exploit cognitive biases -- systematic deviations from rational judgment. Unlike prior jailbreaking approaches focused on prompt engineering or algorithmic manipulation, this work highlights the overlooked power of multi-bias interactions in undermining LLM safeguards. We propose CognitiveAttack, a novel red-teaming framework that systematically leverages both individual and combined cognitive biases. By integrating supervised fine-tuning and reinforcement learning, CognitiveAttack generates prompts that embed optimized bias combinations, effectively bypassing safety protocols while maintaining high attack success rates. Experimental results reveal significant vulnerabilities across 30 diverse LLMs, particularly in open-source models. CognitiveAttack achieves a substantially higher attack success rate compared to the SOTA black-box method PAP (60.1% vs. 31.6%), exposing critical limitations in current defense mechanisms. These findings highlight multi-bias interactions as a powerful yet underexplored attack vector. This work introduces a novel interdisciplinary perspective by bridging cognitive science and LLM safety, paving the way for more robust and human-aligned AI systems.
Authors: Afshin Khadangi, Amir Sartipi, Igor Tchappi, Ramin Bahmani, Gilbert Fridgen
Abstract: The tension between data privacy and model utility has become the defining bottleneck for the practical deployment of large language models (LLMs) trained on sensitive corpora including healthcare. Differentially private stochastic gradient descent (DP-SGD) guarantees formal privacy, yet it does so at a pronounced cost: gradients are forcibly clipped and perturbed with noise, degrading sample efficiency and final accuracy. Numerous variants have been proposed to soften this trade-off, but they all share a handicap: their control knobs are hard-coded, global, and oblivious to the evolving optimization landscape. Consequently, practitioners are forced either to over-spend privacy budget in pursuit of utility, or to accept mediocre models in order to stay within privacy constraints. We present RLDP, the first framework to cast DP optimization itself as a closed-loop control problem amenable to modern deep reinforcement learning (RL). RLDP continuously senses rich statistics of the learning dynamics and acts by selecting fine-grained per parameter gradient-clipping thresholds as well as the magnitude of injected Gaussian noise. A soft actor-critic (SAC) hyper-policy is trained online during language model fine-tuning; it learns, from scratch, how to allocate the privacy budget where it matters and when it matters. Across more than 1,600 ablation experiments on GPT2-small, Llama-1B, Llama-3B, and Mistral-7B, RLDP delivers perplexity reductions of 1.3-30.5% (mean 5.4%) and an average 5.6% downstream utility gain. RLDP reaches each baseline's final utility after only 13-43% of the gradient-update budget (mean speed-up 71%), all while honoring the same ($\epsilon$, $\delta$)-DP contract and exhibiting equal or lower susceptibility to membership-inference and canary-extraction attacks.
Authors: Leandro Farina, Sergey Korotov
Abstract: This work demonstrates a methodology for using deep learning to discover simple, practical criteria for classifying matrices based on abstract algebraic properties. By combining a high-performance neural network with explainable AI (XAI) techniques, we can distill a model's learned strategy into human-interpretable rules. We apply this approach to the challenging case of monotone matrices, defined by the condition that their inverses are entrywise nonnegative. Despite their simple definition, an easy characterization in terms of the matrix elements or the derived parameters is not known. Here, we present, to the best of our knowledge, the first systematic machine-learning approach for deriving a practical criterion that distinguishes monotone from non-monotone matrices. After establishing a labelled dataset by randomly generated monotone and non-monotone matrices uniformly on $(-1,1)$, we employ deep neural network algorithms for classifying the matrices as monotone or non-monotone, using both their entries and a comprehensive set of matrix features. By saliency methods, such as integrated gradients, we identify among all features, two matrix parameters which alone provide sufficient information for the matrix classification, with $95\%$ accuracy, namely the absolute values of the two lowest-order coefficients, $c_0$ and $c_1$ of the matrix's characteristic polynomial. A data-driven study of 18,000 random $7\times7$ matrices shows that the monotone class obeys $\lvert c_{0}/c_{1}\rvert\le0.18$ with probability $>99.98\%$; because $\lvert c_{0}/c_{1}\rvert = 1/\mathrm{tr}(A^{-1})$ for monotone $A$, this is equivalent to the simple bound $\mathrm{tr}(A^{-1})\ge5.7$.
Authors: Galadrielle Humblot-Renaux, Gianni Franchi, Sergio Escalera, Thomas B. Moeslund
Abstract: Out-of-distribution (OOD) detection is an important building block in trustworthy image recognition systems as unknown classes may arise at test-time. OOD detection methods typically revolve around a single classifier, leading to a split in the research field between the classical supervised setting (e.g. ResNet18 classifier trained on CIFAR100) vs. the zero-shot setting (class names fed as prompts to CLIP). In both cases, an overarching challenge is that the OOD detection performance is implicitly constrained by the classifier's capabilities on in-distribution (ID) data. In this work, we show that given a little open-mindedness from both ends, remarkable OOD detection can be achieved by instead creating a heterogeneous ensemble - COOkeD combines the predictions of a closed-world classifier trained end-to-end on a specific dataset, a zero-shot CLIP classifier, and a linear probe classifier trained on CLIP image features. While bulky at first sight, this approach is modular, post-hoc and leverages the availability of pre-trained VLMs, thus introduces little overhead compared to training a single standard classifier. We evaluate COOkeD on popular CIFAR100 and ImageNet benchmarks, but also consider more challenging, realistic settings ranging from training-time label noise, to test-time covariate shift, to zero-shot shift which has been previously overlooked. Despite its simplicity, COOkeD achieves state-of-the-art performance and greater robustness compared to both classical and CLIP-based OOD detection methods. Code is available at https://github.com/glhr/COOkeD
Authors: Qian Qi
Abstract: The canonical theory of sublinear expectations, a foundation of stochastic calculus under ambiguity, is insensitive to the non-convex geometry of primitive uncertainty models. This paper develops a new stochastic calculus for a structured class of such non-convex models. We introduce a class of fully coupled Mean-Field Forward-Backward Stochastic Differential Equations where the BSDE driver is defined by a pointwise maximization over a law-dependent, non-convex set. Mathematical tractability is achieved via a uniform strong concavity assumption on the driver with respect to the control variable, which ensures the optimization admits a unique and stable solution. A central contribution is to establish the Lipschitz stability of this optimizer from primitive geometric and regularity conditions, which underpins the entire well-posedness theory. We prove local and global well-posedness theorems for the FBSDE system. The resulting valuation functional, the $\Theta$-Expectation, is shown to be dynamically consistent and, most critically, to violate the axiom of sub-additivity. This, along with its failure to be translation invariant, demonstrates its fundamental departure from the convex paradigm. This work provides a rigorous foundation for stochastic calculus under a class of non-convex, endogenous ambiguity.
Authors: Marcos Fuster-Pena, David de-Fitero-Dominguez, Antonio Garcia-Cabot, Eva Garcia-Lopez
Abstract: Automated Program Repair (APR) seeks to automatically correct software bugs without requiring human intervention. However, existing tools tend to generate patches that satisfy test cases without fixing the underlying bug, those are known as overfitting patches. To address this issue, Automated Patch Correctness Assessment (APCA) attempts to identify overfitting patches generated by APR tools. It can be solved as a static approach, meaning that no additional information is needed beyond the original and fixed code snippets. Current static techniques often struggle with reliability, flexibility and transparency. To address these issues, we introduce RePaCA, a novel static APCA technique that leverages Large Language Models (LLMs) specialized in thinking tasks. Our model is prompted with both buggy and fixed code snippets and guided to generate a Chain of Thought that analyses code differences, reasons about how the patch addresses the root cause, and ultimately provides a binary classification: correct or overfitting. To enhance these reasoning capabilities for the APCA task specifically, the LLM is finetuned using Reinforcement Learning with the Group Relative Policy Optimization algorithm. When evaluated on a standard Defects4J-derived test, our approach achieves state-of-the-art performance, with 83.1% accuracy and an 84.8% F1-score. Furthermore, our model demonstrates superior generalization capabilities when trained on different datasets, outperforming the leading technique. This reasoning capability also provides enhanced explainability for the patch assessment. These findings underscore the considerable promise of finetuned, reasoning LLMs to advance static APCA by enhancing accuracy, generalization, and explainability.
Authors: Rawan Al-Matham, Kareem Darwish, Raghad Al-Rasheed, Waad Alshammari, Muneera Alhoshan, Amal Almazrua, Asma Al Wazrah, Mais Alheraki, Firoj Alam, Preslav Nakov, Norah Alzahrani, Eman alBilali, Nizar Habash, Abdelrahman El-Sheikh, Muhammad Elmallah, Haonan Li, Hamdy Mubarak, Mohamed Anwar, Zaid Alyafeai, Ahmed Abdelali, Nora Altwairesh, Maram Hasanain, Abdulmohsen Al Thubaity, Shady Shehata, Bashar Alhafni, Injy Hamed, Go Inoue, Khalid Elmadani, Ossama Obeid, Fatima Haouari, Tamer Elsayed, Emad Alghamdi, Khalid Almubarak, Saied Alshahrani, Ola Aljarrah, Safa Alajlan, Areej Alshaqarawi, Maryam Alshihri, Sultana Alghurabi, Atikah Alzeghayer, Afrah Altamimi, Abdullah Alfaifi, Abdulrahman AlOsaimy
Abstract: The impressive advancement of Large Language Models (LLMs) in English has not been matched across all languages. In particular, LLM performance in Arabic lags behind, due to data scarcity, linguistic diversity of Arabic and its dialects, morphological complexity, etc. Progress is further hindered by the quality of Arabic benchmarks, which typically rely on static, publicly available data, lack comprehensive task coverage, or do not provide dedicated platforms with blind test sets. This makes it challenging to measure actual progress and to mitigate data contamination. Here, we aim to bridge these gaps. In particular, we introduce BALSAM, a comprehensive, community-driven benchmark aimed at advancing Arabic LLM development and evaluation. It includes 78 NLP tasks from 14 broad categories, with 52K examples divided into 37K test and 15K development, and a centralized, transparent platform for blind evaluation. We envision BALSAM as a unifying platform that sets standards and promotes collaborative research to advance Arabic LLM capabilities.
Authors: Ruifeng Yuan, Chenghao Xiao, Sicong Leng, Jianyu Wang, Long Li, Weiwen Xu, Hou Pong Chan, Deli Zhao, Tingyang Xu, Zhongyu Wei, Hao Zhang, Yu Rong
Abstract: Reinforcement learning has proven its effectiveness in enhancing the reasoning capabilities of large language models. Recent research efforts have progressively extended this paradigm to multimodal reasoning tasks. Due to the inherent complexity and diversity of multimodal tasks, especially in semantic content and problem formulations, existing models often exhibit unstable performance across various domains and difficulty levels. To address these limitations, we propose VL-Cogito, an advanced multimodal reasoning model trained via a novel multi-stage Progressive Curriculum Reinforcement Learning (PCuRL) framework. PCuRL systematically guides the model through tasks of gradually increasing difficulty, substantially improving its reasoning abilities across diverse multimodal contexts. The framework introduces two key innovations: (1) an online difficulty soft weighting mechanism, dynamically adjusting training difficulty across successive RL training stages; and (2) a dynamic length reward mechanism, which encourages the model to adaptively regulate its reasoning path length according to task complexity, thus balancing reasoning efficiency with correctness. Experimental evaluations demonstrate that VL-Cogito consistently matches or surpasses existing reasoning-oriented models across mainstream multimodal benchmarks spanning mathematics, science, logic, and general understanding, validating the effectiveness of our approach.
Authors: Ali Asgari, Milan de Koning, Pouria Derakhshanfar, Annibale Panichella
Abstract: Large language models and deep learning models designed for code intelligence have revolutionized the software engineering field due to their ability to perform various code-related tasks. These models can process source code and software artifacts with high accuracy in tasks such as code completion, defect detection, and code summarization; therefore, they can potentially become an integral part of modern software engineering practices. Despite these capabilities, robustness remains a critical quality attribute for deep-code models as they may produce different results under varied and adversarial conditions (e.g., variable renaming). Metamorphic testing has become a widely used approach to evaluate models' robustness by applying semantic-preserving transformations to input programs and analyzing the stability of model outputs. While prior research has explored testing deep learning models, this systematic literature review focuses specifically on metamorphic testing for deep code models. By studying 45 primary papers, we analyze the transformations, techniques, and evaluation methods used to assess robustness. Our review summarizes the current landscape, identifying frequently evaluated models, programming tasks, datasets, target languages, and evaluation metrics, and highlights key challenges and future directions for advancing the field.
Authors: Junjie Cao
Abstract: Speech-to-text alignment is a critical component of neural text to-speech (TTS) models. Autoregressive TTS models typically use an attention mechanism to learn these alignments on-line. However, these alignments tend to be brittle and often fail to generalize to long utterances and out-of-domain text, leading to missing or repeating words. Most non-autoregressive end to-end TTS models rely on durations extracted from external sources, using additional duration models for alignment. In this paper, we propose a novel duration prediction framework that can give compromising phoneme-level duration distribution with given text. In our experiments, the proposed duration model has more precise prediction and condition adaptation ability compared to previous baseline models. Numerically, it has roughly a 11.3 percents immprovement on alignment accuracy, and makes the performance of zero-shot TTS models more robust to the mismatch between prompt audio and input audio.
Authors: Federico Girella, Davide Talon, Ziyue Liu, Zanxi Ruan, Yiming Wang, Marco Cristani
Abstract: Fashion design is a complex creative process that blends visual and textual expressions. Designers convey ideas through sketches, which define spatial structure and design elements, and textual descriptions, capturing material, texture, and stylistic details. In this paper, we present LOcalized Text and Sketch for fashion image generation (LOTS), an approach for compositional sketch-text based generation of complete fashion outlooks. LOTS leverages a global description with paired localized sketch + text information for conditioning and introduces a novel step-based merging strategy for diffusion adaptation. First, a Modularized Pair-Centric representation encodes sketches and text into a shared latent space while preserving independent localized features; then, a Diffusion Pair Guidance phase integrates both local and global conditioning via attention-based guidance within the diffusion model's multi-step denoising process. To validate our method, we build on Fashionpedia to release Sketchy, the first fashion dataset where multiple text-sketch pairs are provided per image. Quantitative results show LOTS achieves state-of-the-art image generation performance on both global and localized metrics, while qualitative examples and a human evaluation study highlight its unprecedented level of design customization.
Authors: Wei Guo, Siyuan Lu, Yiqi Tong, Zhaojun Hu, Fuzhen Zhuang, Xiao Zhang, Tao Fan, Jin Dong
Abstract: Different from existing federated fine-tuning (FFT) methods for foundation models, hybrid heterogeneous federated fine-tuning (HHFFT) is an under-explored scenario where clients exhibit double heterogeneity in model architectures and downstream tasks. This hybrid heterogeneity introduces two significant challenges: 1) heterogeneous matrix aggregation, where clients adopt different large-scale foundation models based on their task requirements and resource limitations, leading to dimensional mismatches during LoRA parameter aggregation; and 2) multi-task knowledge interference, where local shared parameters, trained with both task-shared and task-specific knowledge, cannot ensure only task-shared knowledge is transferred between clients. To address these challenges, we propose H2Tune, a federated foundation model fine-tuning with hybrid heterogeneity. Our framework H2Tune consists of three key components: (i) sparsified triple matrix decomposition to align hidden dimensions across clients through constructing rank-consistent middle matrices, with adaptive sparsification based on client resources; (ii) relation-guided matrix layer alignment to handle heterogeneous layer structures and representation capabilities; and (iii) alternating task-knowledge disentanglement mechanism to decouple shared and specific knowledge of local model parameters through alternating optimization. Theoretical analysis proves a convergence rate of O(1/\sqrt{T}). Extensive experiments show our method achieves up to 15.4% accuracy improvement compared to state-of-the-art baselines. Our code is available at https://anonymous.4open.science/r/H2Tune-1407.
Authors: Alex Durkin, Jasper Stolte, Matthew Jones, Raghuraman Pitchumani, Bei Li, Christian Michler, Mehmet Mercang\"oz
Abstract: Offline reinforcement learning (offline RL) offers a promising framework for developing control strategies in chemical process systems using historical data, without the risks or costs of online experimentation. This work investigates the application of offline RL to the safe and efficient control of an exothermic polymerisation continuous stirred-tank reactor. We introduce a Gymnasium-compatible simulation environment that captures the reactor's nonlinear dynamics, including reaction kinetics, energy balances, and operational constraints. The environment supports three industrially relevant scenarios: startup, grade change down, and grade change up. It also includes reproducible offline datasets generated from proportional-integral controllers with randomised tunings, providing a benchmark for evaluating offline RL algorithms in realistic process control tasks. We assess behaviour cloning and implicit Q-learning as baseline algorithms, highlighting the challenges offline agents face, including steady-state offsets and degraded performance near setpoints. To address these issues, we propose a novel deployment-time safety layer that performs gradient-based action correction using input convex neural networks (PICNNs) as learned cost models. The PICNN enables real-time, differentiable correction of policy actions by descending a convex, state-conditioned cost surface, without requiring retraining or environment interaction. Experimental results show that offline RL, particularly when combined with convex action correction, can outperform traditional control approaches and maintain stability across all scenarios. These findings demonstrate the feasibility of integrating offline RL with interpretable and safety-aware corrections for high-stakes chemical process control, and lay the groundwork for more reliable data-driven automation in industrial systems.
Authors: Sabrina Kaniewski, Fabian Schmidt, Markus Enzweiler, Michael Menth, Tobias Heer
Abstract: The increasing adoption of Large Language Models (LLMs) in software engineering has sparked interest in their use for software vulnerability detection. However, the rapid development of this field has resulted in a fragmented research landscape, with diverse studies that are difficult to compare due to differences in, e.g., system designs and dataset usage. This fragmentation makes it difficult to obtain a clear overview of the state-of-the-art or compare and categorize studies meaningfully. In this work, we present a comprehensive systematic literature review (SLR) of LLM-based software vulnerability detection. We analyze 227 studies published between January 2020 and June 2025, categorizing them by task formulation, input representation, system architecture, and adaptation techniques. Further, we analyze the datasets used, including their characteristics, vulnerability coverage, and diversity. We present a fine-grained taxonomy of vulnerability detection approaches, identify key limitations, and outline actionable future research opportunities. By providing a structured overview of the field, this review improves transparency and serves as a practical guide for researchers and practitioners aiming to conduct more comparable and reproducible research. We publicly release all artifacts and maintain a living repository of LLM-based software vulnerability detection studies.
Authors: Mashal Afzal Memon, Gianluca Filippone, Gian Luca Scoccia, Marco Autili, Paola Inverardi
Abstract: The presence of autonomous systems is growing at a fast pace and it is impacting many aspects of our lives. Designed to learn and act independently, these systems operate and perform decision-making without human intervention. However, they lack the ability to incorporate users' ethical preferences, which are unique for each individual in society and are required to personalize the decision-making processes. This reduces user trust and prevents autonomous systems from behaving according to the moral beliefs of their end-users. When multiple systems interact with differing ethical preferences, they must negotiate to reach an agreement that satisfies the ethical beliefs of all the parties involved and adjust their behavior consequently. To address this challenge, this paper proposes RobEthiChor, an approach that enables autonomous systems to incorporate user ethical preferences and contextual factors into their decision-making through ethics-based negotiation. RobEthiChor features a domain-agnostic reference architecture for designing autonomous systems capable of ethic-based negotiating. The paper also presents RobEthiChor-Ros, an implementation of RobEthiChor within the Robot Operating System (ROS), which can be deployed on robots to provide them with ethics-based negotiation capabilities. To evaluate our approach, we deployed RobEthiChor-Ros on real robots and ran scenarios where a pair of robots negotiate upon resource contention. Experimental results demonstrate the feasibility and effectiveness of the system in realizing ethics-based negotiation. RobEthiChor allowed robots to reach an agreement in more than 73\% of the scenarios with an acceptable negotiation time (0.67s on average). Experiments also demonstrate that the negotiation approach implemented in RobEthiChor is scalable.
Authors: Sami Saeed Alghamdi (Open Lab, School of Computing, Newcastle University), Christopher Bull (Open Lab, School of Computing, Newcastle University), Ahmed Kharrufa (Open Lab, School of Computing, Newcastle University)
Abstract: Many people learn programming independently from online resources and often report struggles in achieving their personal learning goals. Learners frequently describe their experiences as isolating and frustrating, challenged by abundant uncertainties, information overload, and distraction, compounded by limited guidance. At the same time, social media serves as a personal space where many engage in diverse self-regulation practices, including help-seeking, using external memory aids (e.g., self-notes), self-reflection, emotion regulation, and self-motivation. For instance, learners often mark achievements and set milestones through their posts. In response, we developed a system consisting of a web platform and browser extensions to support self-regulation online. The design aims to add learner-defined structure to otherwise unstructured experiences and bring meaning to curation and reflection activities by translating them into learning stories with AI-generated feedback. We position storytelling as an integrative approach to design that connects resource curation, reflective and sensemaking practice, and narrative practices learners already use across social platforms. We recruited 15 informal programming learners who are regular social media users to engage with the system in a self-paced manner; participation concluded upon submitting a learning story and survey. We used three quantitative scales and a qualitative survey to examine users' characteristics and perceptions of the system's support for their self-regulation. User feedback suggests the system's viability as a self-regulation aid. Learners particularly valued in-situ reflection, automated story feedback, and video annotation, while other features received mixed views. We highlight perceived benefits, friction points, and design opportunities for future AI-augmented self-regulation tools.
Authors: Yimeng Liu, Maolin Gan, Yidong Ren, Gen Li, Jingkai Lin, Younsuk Dong, Zhichao Cao
Abstract: Leaf wetness detection is a crucial task in agricultural monitoring, as it directly impacts the prediction and protection of plant diseases. However, existing sensing systems suffer from limitations in robustness, accuracy, and environmental resilience when applied to natural leaves under dynamic real-world conditions. To address these challenges, we introduce a new multi-modal dataset specifically designed for evaluating and advancing machine learning algorithms in leaf wetness detection. Our dataset comprises synchronized mmWave raw data, Synthetic Aperture Radar (SAR) images, and RGB images collected over six months from five diverse plant species in both controlled and outdoor field environments. We provide detailed benchmarks using the Hydra model, including comparisons against single modality baselines and multiple fusion strategies, as well as performance under varying scan distances. Additionally, our dataset can serve as a benchmark for future SAR imaging algorithm optimization, enabling a systematic evaluation of detection accuracy under diverse conditions.
Authors: Josh Millar, Ryan Gibb, Roy Ang, Anil Madhavapeddy, Hamed Haddadi
Abstract: Modern networked environments increasingly rely on spatial reasoning, but lack a coherent representation for coordinating physical space. Consequently, tasks such as enforcing spatial access policies remain fragile and manual. We first propose a unifying representation based on bigraphs, capturing spatial, social, and communication relationships within a single formalism, with user-facing tools to generate bigraphs from physical environments. Second, we present a hierarchical agent architecture for distributed spatial reasoning, with runtimes for agentic processes to interact the spatial representation, and a context-aware execution model that scopes reasoning to the smallest viable subspace. Together, these enable private, reliable, and low-latency spatial networking that can safely interact with agentic workflows.
Authors: Hong-Jun Yoon, Mariam Kiran, Danial Ebling, Joe Breen
Abstract: The rapid evolution of network infrastructure is bringing new challenges and opportunities for efficient network management, optimization, and security. With very large monitoring databases becoming expensive to explore, the use of AI and Generative AI can help reduce costs of managing these datasets. This paper explores the use of Large Language Models (LLMs) to revolutionize network monitoring management by addressing the limitations of query finding and pattern analysis. We leverage LLMs to enhance anomaly detection, automate root-cause analysis, and automate incident analysis to build a well-monitored network management team using AI. Through a real-world example of developing our own OFCNetLLM, based on the open-source LLM model, we demonstrate practical applications of OFCnetLLM in the OFC conference network. Our model is developed as a multi-agent approach and is still evolving, and we present early results here.
Authors: Praveenkumar Katwe, Rakesh Chandra, Balabantaray Kali, Prasad Vittala
Abstract: Reducing hallucinations in abstractive summarization remains a critical challenge for deploying language models (LMs) in real-world settings. In this work, we introduce a rewarddriven fine-tuning framework that explicitly optimizes for Entity Hallucination Index (EHI), a metric designed to quantify the presence, correctness, and grounding of named entities in generated summaries. Given a corpus of meeting transcripts, we first generate baseline summaries using a pre-trained LM and compute EHI scores via automatic entity extraction and matching. We then apply reinforcement learning to fine-tune the model parameters, using EHI as a reward signal to bias generation toward entity-faithful outputs. Our approach does not rely on human-written factuality annotations, enabling scalable fine-tuning. Experiments demonstrate consistent improvements in EHI across datasets, with qualitative analysis revealing a significant reduction in entity-level hallucinations without degradation in fluency or informativeness. We release a reproducible Colab pipeline, facilitating further research on hallucination-aware model fine-tuning using lightweight, hallucintion metrics like EHI.
Authors: Samuel Teuber, Debasmita Lohar, Bernhard Beckert
Abstract: As neural networks (NNs) become increasingly prevalent in safety-critical neural network-controlled cyber-physical systems (NNCSs), formally guaranteeing their safety becomes crucial. For these systems, safety must be ensured throughout their entire operation, necessitating infinite-time horizon verification. To verify the infinite-time horizon safety of NNCSs, recent approaches leverage Differential Dynamic Logic (dL). However, these dL-based guarantees rely on idealized, real-valued NN semantics and fail to account for roundoff errors introduced by finite-precision implementations. This paper bridges the gap between theoretical guarantees and real-world implementations by incorporating robustness under finite-precision perturbations -- in sensing, actuation, and computation -- into the safety verification. We model the problem as a hybrid game between a good Demon, responsible for control actions, and a bad Angel, introducing perturbations. This formulation enables formal proofs of robustness w.r.t. a given (bounded) perturbation. Leveraging this bound, we employ state-of-the-art mixed-precision fixed-point tuners to synthesize sound and efficient implementations, thus providing a complete end-to-end solution. We evaluate our approach on case studies from the automotive and aeronautics domains, producing efficient NN implementations with rigorous infinite-time horizon safety guarantees.
Authors: Felix Kronenwett, Georg Maier, Thomas Laengle
Abstract: Sensor-based sorting systems enable the physical separation of a material stream into two fractions. The sorting decision is based on the image data evaluation of the sensors used and is carried out using actuators. Various process parameters must be set depending on the properties of the material stream, the dimensioning of the system, and the required sorting accuracy. However, continuous verification and re-adjustment are necessary due to changing requirements and material stream compositions. In this paper, we introduce an approach for optimizing, recurrently monitoring and adjusting the process parameters of a sensor-based sorting system. Based on Bayesian Optimization, Gaussian process regression models are used as surrogate models to achieve specific requirements for system behavior with the uncertainties contained therein. This method minimizes the number of necessary experiments while simultaneously considering two possible optimization targets based on the requirements for both material output streams. In addition, uncertainties are considered during determining sorting accuracies in the model calculation. We evaluated the method with three example process parameters.
Authors: Soumyadeep Dhar, Kei Sen Fong, Mehul Motani
Abstract: Distilling large neural networks into simple, human-readable symbolic formulas is a promising path toward trustworthy and interpretable AI. However, this process is often brittle, as the complex functions learned by standard networks are poor targets for symbolic discovery, resulting in low-fidelity student models. In this work, we propose a novel training paradigm to address this challenge. Instead of passively distilling a pre-trained network, we introduce a \textbf{Jacobian-based regularizer} that actively encourages the ``teacher'' network to learn functions that are not only accurate but also inherently smoother and more amenable to distillation. We demonstrate through extensive experiments on a suite of real-world regression benchmarks that our method is highly effective. By optimizing the regularization strength for each problem, we improve the $R^2$ score of the final distilled symbolic model by an average of \textbf{120\% (relative)} compared to the standard distillation pipeline, all while maintaining the teacher's predictive accuracy. Our work presents a practical and principled method for significantly improving the fidelity of interpretable models extracted from complex neural networks.
Authors: Ahmed Sabbah, Radi Jarrar, Samer Zein, David Mohaisen
Abstract: Despite outstanding results, machine learning-based Android malware detection models struggle with concept drift, where rapidly evolving malware characteristics degrade model effectiveness. This study examines the impact of concept drift on Android malware detection, evaluating two datasets and nine machine learning and deep learning algorithms, as well as Large Language Models (LLMs). Various feature types--static, dynamic, hybrid, semantic, and image-based--were considered. The results showed that concept drift is widespread and significantly affects model performance. Factors influencing the drift include feature types, data environments, and detection methods. Balancing algorithms helped with class imbalance but did not fully address concept drift, which primarily stems from the dynamic nature of the malware landscape. No strong link was found between the type of algorithm used and concept drift, the impact was relatively minor compared to other variables since hyperparameters were not fine-tuned, and the default algorithm configurations were used. While LLMs using few-shot learning demonstrated promising detection performance, they did not fully mitigate concept drift, highlighting the need for further investigation.
Authors: Junyu Wu, Weiming Chang, Xiaotao Liu, Guanyou He, Haoqiang Hong, Boqi Liu, Hongtao Tian, Tao Yang, Yunsheng Shi, Feng Lin, Ting Yao
Abstract: Reinforcement Learning from Human Feedback (RLHF) has become an increasingly popular paradigm for training large language models (LLMs) and diffusion models. While existing RLHF training systems have enabled significant progress, they often face challenges in scaling to multi-modal and diffusion workflows and adapting to dynamic workloads. In particular, current approaches may encounter limitations in controller scalability, flexible resource placement, and efficient orchestration when handling complex RLHF pipelines, especially in scenarios involving dynamic sampling or generative reward modeling. In this paper, we present \textbf{G-Core}, a simple, scalable, and balanced RLHF training framework designed to address these challenges. G-Core introduces a parallel controller programming model, enabling flexible and efficient orchestration of complex RLHF workflows without the bottlenecks of a single centralized controller. Furthermore, we propose a dynamic placement schema that adaptively partitions resources and schedules workloads, significantly reducing hardware idle time and improving utilization, even under highly variable training conditions. G-Core has successfully trained models that support WeChat product features serving a large-scale user base, demonstrating its effectiveness and robustness in real-world scenarios. Our results show that G-Core advances the state of the art in RLHF training, providing a solid foundation for future research and deployment of large-scale, human-aligned models.
Authors: Dongli He, Hu Wang, Mohammad Yaqub
Abstract: Accurate fetal biometric measurements, such as abdominal circumference, play a vital role in prenatal care. However, obtaining high-quality ultrasound images for these measurements heavily depends on the expertise of sonographers, posing a significant challenge in low-income countries due to the scarcity of trained personnel. To address this issue, we leverage FetalCLIP, a vision-language model pretrained on a curated dataset of over 210,000 fetal ultrasound image-caption pairs, to perform automated fetal ultrasound image quality assessment (IQA) on blind-sweep ultrasound data. We introduce FetalCLIP$_{CLS}$, an IQA model adapted from FetalCLIP using Low-Rank Adaptation (LoRA), and evaluate it on the ACOUSLIC-AI dataset against six CNN and Transformer baselines. FetalCLIP$_{CLS}$ achieves the highest F1 score of 0.757. Moreover, we show that an adapted segmentation model, when repurposed for classification, further improves performance, achieving an F1 score of 0.771. Our work demonstrates how parameter-efficient fine-tuning of fetal ultrasound foundation models can enable task-specific adaptations, advancing prenatal care in resource-limited settings. The experimental code is available at: https://github.com/donglihe-hub/FetalCLIP-IQA.
Authors: Yuqi Pang, Bowen Yang, Yun Cao, Fan Rong, Xiaoyu Li, Chen He
Abstract: Vision large language models (VLLMs) are focusing primarily on handling complex and fine-grained visual information by incorporating advanced vision encoders and scaling up visual models. However, these approaches face high training and inference costs, as well as challenges in extracting visual details, effectively bridging across modalities. In this work, we propose a novel visual framework, MoCHA, to address these issues. Our framework integrates four vision backbones (i.e., CLIP, SigLIP, DINOv2 and ConvNeXt) to extract complementary visual features and is equipped with a sparse Mixture of Experts Connectors (MoECs) module to dynamically select experts tailored to different visual dimensions. To mitigate redundant or insufficient use of the visual information encoded by the MoECs module, we further design a Hierarchical Group Attention (HGA) with intra- and inter-group operations and an adaptive gating strategy for encoded visual features. We train MoCHA on two mainstream LLMs (e.g., Phi2-2.7B and Vicuna-7B) and evaluate their performance across various benchmarks. Notably, MoCHA outperforms state-of-the-art open-weight models on various tasks. For example, compared to CuMo (Mistral-7B), our MoCHA (Phi2-2.7B) presents outstanding abilities to mitigate hallucination by showing improvements of 3.25% in POPE and to follow visual instructions by raising 153 points on MME. Finally, ablation studies further confirm the effectiveness and robustness of the proposed MoECs and HGA in improving the overall performance of MoCHA.
Authors: Kedong Xiu, Saiqian Zhang
Abstract: As Vision-Language Models (VLMs) are increasingly deployed in split-DNN configurations--with visual encoders (e.g., ResNet, ViT) operating on user devices and sending intermediate features to the cloud--there is a growing privacy risk from semantic information leakage. Existing approaches to reconstructing images from these intermediate features often result in blurry, semantically ambiguous images. To directly address semantic leakage, we propose CapRecover, a cross-modality inversion framework that recovers high-level semantic content, such as labels or captions, directly from intermediate features without image reconstruction. We evaluate CapRecover on multiple datasets and victim models, demonstrating strong performance in semantic recovery. Specifically, CapRecover achieves up to 92.71% Top-1 label accuracy on CIFAR-10 and generates fluent captions from ResNet50 features on COCO2017 with ROUGE-L scores up to 0.52. Our analysis further reveals that deeper convolutional layers encode significantly more semantic information compared to shallow layers. To mitigate semantic leakage, we introduce a simple yet effective protection method: adding random noise to intermediate features at each layer and removing the noise in the next layer. Experimental results show that this approach prevents semantic leakage without additional training costs.
Authors: Zijing Zhang, Ziyang Chen, Mingxiao Li, Zhaopeng Tu, Xiaolong Li
Abstract: The development of autonomous agents for complex, long-horizon tasks is a central goal in AI. However, dominant training paradigms face a critical limitation: reinforcement learning (RL) methods that optimize solely for final task success often reinforce flawed or inefficient reasoning paths, a problem we term inefficient exploration. This leads to agents that are brittle and fail to generalize, as they learn to find solutions without learning how to reason coherently. To address this, we introduce RLVMR, a novel framework that integrates dense, process-level supervision into end-to-end RL by rewarding verifiable, meta-reasoning behaviors. RLVMR equips an agent to explicitly tag its cognitive steps, such as planning, exploration, and reflection, and provides programmatic, rule-based rewards for actions that contribute to effective problem-solving. These process-centric rewards are combined with the final outcome signal and optimized using a critic-free policy gradient method. On the challenging ALFWorld and ScienceWorld benchmarks, RLVMR achieves new state-of-the-art results, with our 7B model reaching an 83.6% success rate on the most difficult unseen task split. Our analysis confirms these gains stem from improved reasoning quality, including significant reductions in redundant actions and enhanced error recovery, leading to more robust, efficient, and interpretable agents.
Authors: Haichuan Hu, Xiaochen Xie, Quanjun Zhang
Abstract: APR (Automated Program Repair) aims to automatically locate program defects, generate patches and validate the repairs. Existing techniques for APR are often combined with LLMs (Large Language Models), which leverages the code-related knowledge of LLMs to improve repair effectiveness. Current LLM-based APR methods typically utilize test cases only during the inference stage, adopting an iterative approach that performs repair first and validates it through test execution afterward. This conventional paradigm neglects two important aspects: the potential contribution of test cases in the training phase, and the possibility of leveraging testing prior to repair. To address this, we propose Repair-R1, which introduces test cases into the model's training phase and shifts test generation to precede repair. The model is required to first generate discriminative test cases that can distinguish defective behaviors, and then perform repair based on these tests. This enables the model to better locate defects and understand the underlying causes of defects, thereby improving repair effectiveness. We implement Repair-R1 with three different backbone models, using RL (reinforcement learning) to co-optimize test generation and bug repair. Experimental results on four widely adopted benchmarks demonstrate the superiority of Repair-R1. Specially, compared to vanilla models, Repair-R1 improves repair success rate by 2.68\% to 48.29\%, test generation success rate by 16.38\% to 53.28\%, and test coverage by 0.78\% to 53.96\%. We publish the code and weights at https://github.com/Tomsawyerhu/APR-RL and https://huggingface.co/tomhu/Qwen3-4B-RL-5000-step.
URLs: https://github.com/Tomsawyerhu/APR-RL, https://huggingface.co/tomhu/Qwen3-4B-RL-5000-step.
Authors: Andris Ambainis, Joao F. Doriguello, Debbie Lim
Abstract: We propose novel classical and quantum online algorithms for learning finite-horizon and infinite-horizon average-reward Markov Decision Processes (MDPs). Our algorithms are based on a hybrid exploration-generative reinforcement learning (RL) model wherein the agent can, from time to time, freely interact with the environment in a generative sampling fashion, i.e., by having access to a "simulator". By employing known classical and new quantum algorithms for approximating optimal policies under a generative model within our learning algorithms, we show that it is possible to avoid several paradigms from RL like "optimism in the face of uncertainty" and "posterior sampling" and instead compute and use optimal policies directly, which yields better regret bounds compared to previous works. For finite-horizon MDPs, our quantum algorithms obtain regret bounds which only depend logarithmically on the number of time steps $T$, thus breaking the $O(\sqrt{T})$ classical barrier. This matches the time dependence of the prior quantum works of Ganguly et al. (arXiv'23) and Zhong et al. (ICML'24), but with improved dependence on other parameters like state space size $S$ and action space size $A$. For infinite-horizon MDPs, our classical and quantum bounds still maintain the $O(\sqrt{T})$ dependence but with better $S$ and $A$ factors. Nonetheless, we propose a novel measure of regret for infinite-horizon MDPs with respect to which our quantum algorithms have $\operatorname{poly}\log{T}$ regret, exponentially better compared to classical algorithms. Finally, we generalise all of our results to compact state spaces.
Authors: Kwesi Cobbina, Tianyi Zhou
Abstract: In-context learning (ICL) is a critical emerging capability of large language models (LLMs), enabling few-shot learning during inference by including a few demonstrations (demos) in the prompt. However, it has been found that ICL's performance can be sensitive to the choices of demos and their order. This paper investigates an unexplored new positional bias of ICL for the first time: we observe that the predictions and accuracy can drift drastically when the positions of demos, the system prompt, and the user message in LLM input are varied. We refer to this bias as DEMOS' POSITION IN PROMPT (DPP) bias. We design a systematic evaluation pipeline to study this type of positional bias across classification, question answering, summarization, and reasoning tasks. We introduce two metrics, ACCURACY-CHANGE and PREDICTION-CHANGE, to quantify net gains and output volatility induced by changes in the demos' position. Extensive experiments on ten LLMs from four open-source model families (QWEN, LLAMA3, MISTRAL, COHERE) verify that the bias significantly affects their accuracy and predictions: placing demos at the start of the prompt yields the most stable and accurate outputs with gains of up to +6 points. In contrast, placing demos at the end of the user message flips over 30\% of predictions without improving correctness on QA tasks. Smaller models are most affected by this sensitivity, though even large models remain marginally affected on complex tasks.
Authors: Arun Kumar, Paul Schrater
Abstract: People aptly exhibit general intelligence behaviors through flexible problem-solving and the ability to adapt to novel situations by reusing and applying high-level knowledge acquired over time. In contrast, artificial agents tend to be specialists, lacking such generalist behaviors. To bridge this gap, artificial agents will require understanding and exploiting critical structured knowledge representations. We introduce a metacognitive reasoning framework, Knowledge-Interaction-eXecution (KIX), and argue that interactions with objects, by leveraging a type space, facilitate the learning of transferable interaction concepts and promote generalization. This framework offers a principled approach for integrating knowledge into reinforcement learning and holds promise as an enabler for generalist behaviors in artificial intelligence, robotics, and autonomous systems.
Authors: Ryota Maruo, Koh Takeuchi, Hisashi Kashima
Abstract: Designing two-sided matching mechanisms is challenging when practical demands for matching outcomes are difficult to formalize and the designed mechanism must satisfy theoretical conditions. To address this, prior work has proposed a framework that learns a matching mechanism from examples, using a parameterized family that satisfies properties such as stability. However, despite its usefulness, this framework does not guarantee strategy-proofness (SP), and cannot handle varying numbers of agents or incorporate publicly available contextual information about agents, both of which are crucial in real-world applications. In this paper, we propose a new parametrized family of matching mechanisms that always satisfy strategy-proofness, are applicable for an arbitrary number of agents, and deal with public contextual information of agents, based on the serial dictatorship (SD). This family is represented by NeuralSD, a novel neural network architecture based on SD, where agent rankings in SD are treated as learnable parameters computed from agents' contexts using an attention-based sub-network. To enable learning, we introduce tensor serial dictatorship (TSD), a differentiable relaxation of SD using tensor operations. This allows NeuralSD to be trained end-to-end from example matchings while satisfying SP. We conducted experiments to learn a matching mechanism from matching examples while satisfying SP. We demonstrated that our method outperformed baselines in predicting matchings and on several metrics for goodness of matching outcomes.
Authors: Manuel Cebrian, Andres Abeliuk, Jan Arne Telle
Abstract: Attributing outputs from Large Language Models (LLMs) in adversarial settings-such as cyberattacks and disinformation campaigns-presents significant challenges that are likely to grow in importance. We approach this attribution problem from both a theoretical and an empirical perspective, drawing on formal language theory (identification in the limit) and data-driven analysis of the expanding LLM ecosystem. By modeling an LLM's set of possible outputs as a formal language, we analyze whether finite samples of text can uniquely pinpoint the originating model. Our results show that, under mild assumptions of overlapping capabilities among models, certain classes of LLMs are fundamentally non-identifiable from their outputs alone. We delineate four regimes of theoretical identifiability: (1) an infinite class of deterministic (discrete) LLM languages is not identifiable (Gold's classical result from 1967); (2) an infinite class of probabilistic LLMs is also not identifiable (by extension of the deterministic case); (3) a finite class of deterministic LLMs is identifiable (consistent with Angluin's tell-tale criterion); and (4) even a finite class of probabilistic LLMs can be non-identifiable (we provide a new counterexample establishing this negative result). Complementing these theoretical insights, we quantify the explosion in the number of plausible model origins (hypothesis space) for a given output in recent years. Even under conservative assumptions-each open-source model fine-tuned on at most one new dataset-the count of distinct candidate models doubles approximately every 0.5 years, and allowing multi-dataset fine-tuning combinations yields doubling times as short as 0.28 years. This combinatorial growth, alongside the extraordinary computational cost of brute-force likelihood attribution across all models and potential users, renders exhaustive attribution infeasible in practice.
Authors: Haoyang Li, Yiming Li, Anxin Tian, Tianhao Tang, Zhanchao Xu, Xuejia Chen, Nicole Hu, Wei Dong, Qing Li, Lei Chen
Abstract: Large Language Models (LLMs) have revolutionized a wide range of domains such as natural language processing, computer vision, and multi-modal tasks due to their ability to comprehend context and perform logical reasoning. However, the computational and memory demands of LLMs, particularly during inference, pose significant challenges when scaling them to real-world, long-context, and real-time applications. Key-Value (KV) cache management has emerged as a critical optimization technique for accelerating LLM inference by reducing redundant computations and improving memory utilization. This survey provides a comprehensive overview of KV cache management strategies for LLM acceleration, categorizing them into token-level, model-level, and system-level optimizations. Token-level strategies include KV cache selection, budget allocation, merging, quantization, and low-rank decomposition, while model-level optimizations focus on architectural innovations and attention mechanisms to enhance KV reuse. System-level approaches address memory management, scheduling, and hardware-aware designs to improve efficiency across diverse computing environments. Additionally, the survey provides an overview of both text and multimodal datasets and benchmarks used to evaluate these strategies. By presenting detailed taxonomies and comparative analyses, this work aims to offer useful insights for researchers and practitioners to support the development of efficient and scalable KV cache management techniques, contributing to the practical deployment of LLMs in real-world applications. The curated paper list for KV cache management is in: \href{https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}{https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}.
URLs: https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management, https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management
Authors: Shirley Wu, Michel Galley, Baolin Peng, Hao Cheng, Gavin Li, Yao Dou, Weixin Cai, James Zou, Jure Leskovec, Jianfeng Gao
Abstract: Large Language Models are typically trained with next-turn rewards, limiting their ability to optimize for long-term interaction. As a result, they often respond passively to ambiguous or open-ended user requests, failing to help users reach their ultimate intents and leading to inefficient conversations. To address these limitations, we introduce CollabLLM, a novel and general training framework that enhances multiturn human-LLM collaboration. Its key innovation is a collaborative simulation that estimates the long-term contribution of responses using Multiturn-aware Rewards. By reinforcement fine-tuning these rewards, CollabLLM goes beyond responding to user requests, and actively uncovers user intent and offers insightful suggestions-a key step towards more human-centered AI. We also devise a multiturn interaction benchmark with three challenging tasks such as document creation. CollabLLM significantly outperforms our baselines with averages of 18.5% higher task performance and 46.3% improved interactivity by LLM judges. Finally, we conduct a large user study with 201 judges, where CollabLLM increases user satisfaction by 17.6% and reduces user spent time by 10.4%.
Authors: Aleksander Ficek, Somshubra Majumdar, Vahid Noroozi, Boris Ginsburg
Abstract: Synthetic verification techniques such as generating test cases and reward modelling are common ways to enhance the coding capabilities of large language models (LLM) beyond predefined tests. Additionally, code verification has recently found great success as a critical component in improving reasoning capability of LLMs via reinforcement learning. In this paper, we propose an approach which can transform existing coding benchmarks into scoring and ranking datasets to evaluate the effectiveness of synthetic verifiers. We also propose multiple metrics to measure different aspects of the synthetic verifiers with the proposed benchmarks. By employing the proposed approach, we release four new benchmarks (HE-R, HE-R+, MBPP-R, and MBPP-R+), and analyzed synthetic verification methods with standard, reasoning-based, and reward-based LLMs. Our experiments show that reasoning can significantly improve test case generation and that scaling the number of test cases enhances the verification accuracy.
Authors: Matej \v{S}progar
Abstract: Despite major advances in machine learning, current artificial intelligence systems continue to fall short of human-like general intelligence. While large language and reasoning models can generate fluent and coherent outputs, they lack the deep understanding and adaptive reasoning that characterize truly general intelligence. Existing evaluation frameworks, which are centered on broad language or perception tasks, fail to capture generality at its core and offer no guidance. The artificial general intelligence testbed (AGITB) is a novel and freely available benchmarking suite comprising twelve fully automatable tests designed to evaluate low-level cognitive precursors through binary signal prediction. AGITB requires models to forecast temporal sequences without pretraining, symbolic manipulation, or semantic grounding. The framework isolates core computational invariants - such as determinism, sensitivity, and generalization - that align with principles of biological information processing. Engineered to resist brute-force and memorization-based approaches, AGITB presumes no prior knowledge and demands learning from first principles. While humans pass all tests, no current AI system has met the full AGITB criteria, underscoring its potential as a rigorous, interpretable, and actionable benchmark for guiding and evaluating progress toward artificial general intelligence. A reference implementation of AGITB is available on GitHub.
Authors: Roie Kazoom, Raz Lapid, Moshe Sipper, Ofer Hadar
Abstract: Adversarial patch attacks pose a major threat to vision systems by embedding localized perturbations that mislead deep models. Traditional defense methods often require retraining or fine-tuning, making them impractical for real-world deployment. We propose a training-free Visual Retrieval-Augmented Generation (VRAG) framework that integrates Vision-Language Models (VLMs) for adversarial patch detection. By retrieving visually similar patches and images that resemble stored attacks in a continuously expanding database, VRAG performs generative reasoning to identify diverse attack types, all without additional training or fine-tuning. We extensively evaluate open-source large-scale VLMs, including Qwen-VL-Plus, Qwen2.5-VL-72B, and UI-TARS-72B-DPO, alongside Gemini-2.0, a closed-source model. Notably, the open-source UI-TARS-72B-DPO model achieves up to 95 percent classification accuracy, setting a new state-of-the-art for open-source adversarial patch detection. Gemini-2.0 attains the highest overall accuracy, 98 percent, but remains closed-source. Experimental results demonstrate VRAG's effectiveness in identifying a variety of adversarial patches with minimal human annotation, paving the way for robust, practical defenses against evolving adversarial patch attacks.
Authors: Jake Tuero, Michael Buro, Levi H. S. Lelis
Abstract: Policy tree search is a family of tree search algorithms that use a policy to guide the search. These algorithms provide guarantees on the number of expansions required to solve a given problem that are based on the quality of the policy. While these algorithms have shown promising results, the process in which they are trained requires complete solution trajectories to train the policy. Search trajectories are obtained during a trial-and-error search process. When the training problem instances are hard, learning can be prohibitively costly, especially when starting from a randomly initialized policy. As a result, search samples are wasted in failed attempts to solve these hard instances. This paper introduces a novel method for learning subgoal-based policies for policy tree search algorithms. The subgoals and policies conditioned on subgoals are learned from the trees that the search expands while attempting to solve problems, including the search trees of failed attempts. We empirically show that our policy formulation and training method improve the sample efficiency of learning a policy and heuristic function in this online setting.
Authors: Yining Hong, Rui Sun, Bingxuan Li, Xingcheng Yao, Maxine Wu, Alexander Chien, Da Yin, Ying Nian Wu, Zhecan James Wang, Kai-Wei Chang
Abstract: AI agents today are mostly siloed - they either retrieve and reason over vast amount of digital information and knowledge obtained online; or interact with the physical world through embodied perception, planning and action - but rarely both. This separation limits their ability to solve tasks that require integrated physical and digital intelligence, such as cooking from online recipes, navigating with dynamic map data, or interpreting real-world landmarks using web knowledge. We introduce Embodied Web Agents, a novel paradigm for AI agents that fluidly bridge embodiment and web-scale reasoning. To operationalize this concept, we first develop the Embodied Web Agents task environments, a unified simulation platform that tightly integrates realistic 3D indoor and outdoor environments with functional web interfaces. Building upon this platform, we construct and release the Embodied Web Agents Benchmark, which encompasses a diverse suite of tasks including cooking, navigation, shopping, tourism, and geolocation - all requiring coordinated reasoning across physical and digital realms for systematic assessment of cross-domain intelligence. Experimental results reveal significant performance gaps between state-of-the-art AI systems and human capabilities, establishing both challenges and opportunities at the intersection of embodied cognition and web-scale knowledge access. All datasets, codes and websites are publicly available at our project page https://embodied-web-agent.github.io/.
Authors: Roy Uziel, Irit Chelly, Oren Freifeld, Ari Pakman
Abstract: Diffusion models, widely recognized for their success in generative tasks, have not yet been applied to clustering. We introduce Clustering via Diffusion (CLUDI), a self-supervised framework that combines the generative power of diffusion models with pre-trained Vision Transformer features to achieve robust and accurate clustering. CLUDI is trained via a teacher-student paradigm: the teacher uses stochastic diffusion-based sampling to produce diverse cluster assignments, which the student refines into stable predictions. This stochasticity acts as a novel data augmentation strategy, enabling CLUDI to uncover intricate structures in high-dimensional data. Extensive evaluations on challenging datasets demonstrate that CLUDI achieves state-of-the-art performance in unsupervised classification, setting new benchmarks in clustering robustness and adaptability to complex data distributions. Our code is available at https://github.com/BGU-CS-VIL/CLUDI.
Authors: Peter V. Coveney, Sauro Succi
Abstract: We show that the scaling laws which determine the performance of large language models (LLMs) severely limit their ability to improve the uncertainty of their predictions. As a result, raising their reliability to meet the standards of scientific inquiry is intractable by any reasonable measure. We argue that the very mechanism which fuels much of the learning power of LLMs, namely the ability to generate non-Gaussian output distributions from Gaussian input ones, might well be at the roots of their propensity to produce error pileup, ensuing information catastrophes and degenerative AI behaviour. This tension between learning and accuracy is a likely candidate mechanism underlying the observed low values of the scaling components. It is substantially compounded by the deluge of spurious correlations pointed out by Calude and Longo which rapidly increase in any data set merely as a function of its size, regardless of its nature. The fact that a degenerative AI pathway is a very probable feature of the LLM landscape does not mean that it must inevitably arise in all future AI research. Its avoidance, which we also discuss in this paper, necessitates putting a much higher premium on insight and understanding of the structural characteristics of the problems being investigated.
Authors: Yuzhang Xie, Xu Han, Ran Xu, Xiao Hu, Jiaying Lu, Carl Yang
Abstract: Knowledge graphs (KGs) are important products of the semantic web, which are widely used in various application domains. Healthcare is one of such domains where KGs are intensively used, due to the high requirement for knowledge accuracy and interconnected nature of healthcare data. However, KGs storing general factual information often lack the ability to account for important contexts of the knowledge such as the status of specific patients, which are crucial in precision healthcare. Meanwhile, electronic health records (EHRs) provide rich personal data, including various diagnoses and medications, which provide natural contexts for general KGs. In this paper, we propose HypKG, a framework that integrates patient information from EHRs into KGs to generate contextualized knowledge representations for accurate healthcare predictions. Using advanced entity-linking techniques, we connect relevant knowledge from general KGs with patient information from EHRs, and then utilize a hypergraph model to "contextualize" the knowledge with the patient information. Finally, we employ hypergraph transformers guided by downstream prediction tasks to jointly learn proper contextualized representations for both KGs and patients, fully leveraging existing knowledge in KGs and patient contexts in EHRs. In experiments using a large biomedical KG and two real-world EHR datasets, HypKG demonstrates significant improvements in healthcare prediction tasks across multiple evaluation metrics. Additionally, by integrating external contexts, HypKG can learn to adjust the representations of entities and relations in KG, potentially improving the quality and real-world utility of knowledge.
Authors: Farshid Nooshi, Suining He
Abstract: Allocating mobility resources (e.g., shared bikes/e-scooters, ride-sharing vehicles) is crucial for rebalancing the mobility demand and supply in the urban environments. We propose in this work a novel multi-agent reinforcement learning named Hierarchical Adaptive Grouping-based Parameter Sharing (HAG-PS) for dynamic mobility resource allocation. HAG-PS aims to address two important research challenges regarding multi-agent reinforcement learning for mobility resource allocation: (1) how to dynamically and adaptively share the mobility resource allocation policy (i.e., how to distribute mobility resources) across agents (i.e., representing the regional coordinators of mobility resources); and (2) how to achieve memory-efficient parameter sharing in an urban-scale setting. To address the above challenges, we have provided following novel designs within HAG-PS. To enable dynamic and adaptive parameter sharing, we have designed a hierarchical approach that consists of global and local information of the mobility resource states (e.g., distribution of mobility resources). We have developed an adaptive agent grouping approach in order to split or merge the groups of agents based on their relative closeness of encoded trajectories (i.e., states, actions, and rewards). We have designed a learnable identity (ID) embeddings to enable agent specialization beyond simple parameter copy. We have performed extensive experimental studies based on real-world NYC bike sharing data (a total of more than 1.2 million trips), and demonstrated the superior performance (e.g., improved bike availability) of HAG-PS compared with other baseline approaches.
Authors: Huan-ang Gao, Jiayi Geng, Wenyue Hua, Mengkang Hu, Xinzhe Juan, Hongzhang Liu, Shilong Liu, Jiahao Qiu, Xuan Qi, Yiran Wu, Hongru Wang, Han Xiao, Yuhang Zhou, Shaokun Zhang, Jiayi Zhang, Jinyu Xiang, Yixiong Fang, Qiwen Zhao, Dongrui Liu, Qihan Ren, Cheng Qian, Zhenghailong Wang, Minda Hu, Huazheng Wang, Qingyun Wu, Heng Ji, Mengdi Wang
Abstract: Large Language Models (LLMs) have demonstrated strong capabilities but remain fundamentally static, unable to adapt their internal parameters to novel tasks, evolving knowledge domains, or dynamic interaction contexts. As LLMs are increasingly deployed in open-ended, interactive environments, this static nature has become a critical bottleneck, necessitating agents that can adaptively reason, act, and evolve in real time. This paradigm shift -- from scaling static models to developing self-evolving agents -- has sparked growing interest in architectures and methods enabling continual learning and adaptation from data, interactions, and experiences. This survey provides the first systematic and comprehensive review of self-evolving agents, organized around three foundational dimensions -- what to evolve, when to evolve, and how to evolve. We examine evolutionary mechanisms across agent components (e.g., models, memory, tools, architecture), categorize adaptation methods by stages (e.g., intra-test-time, inter-test-time), and analyze the algorithmic and architectural designs that guide evolutionary adaptation (e.g., scalar rewards, textual feedback, single-agent and multi-agent systems). Additionally, we analyze evaluation metrics and benchmarks tailored for self-evolving agents, highlight applications in domains such as coding, education, and healthcare, and identify critical challenges and research directions in safety, scalability, and co-evolutionary dynamics. By providing a structured framework for understanding and designing self-evolving agents, this survey establishes a roadmap for advancing adaptive agentic systems in both research and real-world deployments, ultimately shedding lights to pave the way for the realization of Artificial Super Intelligence (ASI), where agents evolve autonomously, performing at or beyond human-level intelligence across a wide array of tasks.
Authors: Jing Xu, Weiqiang Wang, Cunjian Chen, Jun Liu, Qiuhong Ke
Abstract: Group dance generation from music has broad applications in film, gaming, and animation production. However, it requires synchronizing multiple dancers while maintaining spatial coordination. As the number of dancers and sequence length increase, this task faces higher computational complexity and a greater risk of motion collisions. Existing methods often struggle to model dense spatial-temporal interactions, leading to scalability issues and multi-dancer collisions. To address these challenges, we propose ST-GDance, a novel framework that decouples spatial and temporal dependencies to optimize long-term and collision-free group choreography. We employ lightweight graph convolutions for distance-aware spatial modeling and accelerated sparse attention for efficient temporal modeling. This design significantly reduces computational costs while ensuring smooth and collision-free interactions. Experiments on the AIOZ-GDance dataset demonstrate that ST-GDance outperforms state-of-the-art baselines, particularly in generating long and coherent group dance sequences. Project page: https://yilliajing.github.io/ST-GDance-Website/.
Authors: Kuiye Ding, Fanda Fan, Yao Wang, Ruijie jian, Xiaorui Wang, Luqi Gong, Yishan Jiang, Chunjie Luo an Jianfeng Zhan
Abstract: Multivariate Time Series Forecasting plays a key role in many applications. Recent works have explored using Large Language Models for MTSF to take advantage of their reasoning abilities. However, many methods treat LLMs as end-to-end forecasters, which often leads to a loss of numerical precision and forces LLMs to handle patterns beyond their intended design. Alternatively, methods that attempt to align textual and time series modalities within latent space frequently encounter alignment difficulty. In this paper, we propose to treat LLMs not as standalone forecasters, but as semantic guidance modules within a dual-stream framework. We propose DualSG, a dual-stream framework that provides explicit semantic guidance, where LLMs act as Semantic Guides to refine rather than replace traditional predictions. As part of DualSG, we introduce Time Series Caption, an explicit prompt format that summarizes trend patterns in natural language and provides interpretable context for LLMs, rather than relying on implicit alignment between text and time series in the latent space. We also design a caption-guided fusion module that explicitly models inter-variable relationships while reducing noise and computation. Experiments on real-world datasets from diverse domains show that DualSG consistently outperforms 15 state-of-the-art baselines, demonstrating the value of explicitly combining numerical forecasting with semantic guidance.
Authors: Shouyi Lu, Zihan Lin, Chao Lu, Huanran Wang, Guirong Zhuo, Lianqing Zheng
Abstract: Autonomous driving systems rely heavily on multimodal perception data to understand complex environments. However, the long-tailed distribution of real-world data hinders generalization, especially for rare but safety-critical vehicle categories. To address this challenge, we propose MultiEditor, a dual-branch latent diffusion framework designed to edit images and LiDAR point clouds in driving scenarios jointly. At the core of our approach is introducing 3D Gaussian Splatting (3DGS) as a structural and appearance prior for target objects. Leveraging this prior, we design a multi-level appearance control mechanism--comprising pixel-level pasting, semantic-level guidance, and multi-branch refinement--to achieve high-fidelity reconstruction across modalities. We further propose a depth-guided deformable cross-modality condition module that adaptively enables mutual guidance between modalities using 3DGS-rendered depth, significantly enhancing cross-modality consistency. Extensive experiments demonstrate that MultiEditor achieves superior performance in visual and geometric fidelity, editing controllability, and cross-modality consistency. Furthermore, generating rare-category vehicle data with MultiEditor substantially enhances the detection accuracy of perception models on underrepresented classes.
Authors: Stefanos Gkikas, Ioannis Kyprakis, Manolis Tsiknakis
Abstract: Pain is a complex and pervasive condition that affects a significant portion of the population. Accurate and consistent assessment is essential for individuals suffering from pain, as well as for developing effective management strategies in a healthcare system. Automatic pain assessment systems enable continuous monitoring, support clinical decision-making, and help minimize patient distress while mitigating the risk of functional deterioration. Leveraging physiological signals offers objective and precise insights into a person's state, and their integration in a multimodal framework can further enhance system performance. This study has been submitted to the \textit{Second Multimodal Sensing Grand Challenge for Next-Gen Pain Assessment (AI4PAIN)}. The proposed approach introduces \textit{Tiny-BioMoE}, a lightweight pretrained embedding model for biosignal analysis. Trained on $4.4$ million biosignal image representations and consisting of only $7.3$ million parameters, it serves as an effective tool for extracting high-quality embeddings for downstream tasks. Extensive experiments involving electrodermal activity, blood volume pulse, respiratory signals, peripheral oxygen saturation, and their combinations highlight the model's effectiveness across diverse modalities in automatic pain recognition tasks. \textit{\textcolor{blue}{The model's architecture (code) and weights are available at https://github.com/GkikasStefanos/Tiny-BioMoE.
Authors: Stefanos Gkikas, Ioannis Kyprakis, Manolis Tsiknakis
Abstract: Pain is a multifaceted phenomenon that affects a substantial portion of the population. Reliable and consistent evaluation benefits those experiencing pain and underpins the development of effective and advanced management strategies. Automatic pain-assessment systems deliver continuous monitoring, inform clinical decision-making, and aim to reduce distress while preventing functional decline. By incorporating physiological signals, these systems provide objective, accurate insights into an individual's condition. This study has been submitted to the \textit{Second Multimodal Sensing Grand Challenge for Next-Gen Pain Assessment (AI4PAIN)}. The proposed method introduces a pipeline that leverages electrodermal activity signals as input modality. Multiple representations of the signal are created and visualized as waveforms, and they are jointly visualized within a single multi-representation diagram. Extensive experiments incorporating various processing and filtering techniques, along with multiple representation combinations, demonstrate the effectiveness of the proposed approach. It consistently yields comparable, and in several cases superior, results to traditional fusion methods, establishing it as a robust alternative for integrating different signal representations or modalities.
Authors: Stefanos Gkikas, Ioannis Kyprakis, Manolis Tsiknakis
Abstract: Pain is a complex condition affecting a large portion of the population. Accurate and consistent evaluation is essential for individuals experiencing pain, and it supports the development of effective and advanced management strategies. Automatic pain assessment systems provide continuous monitoring and support clinical decision-making, aiming to reduce distress and prevent functional decline. This study has been submitted to the \textit{Second Multimodal Sensing Grand Challenge for Next-Gen Pain Assessment (AI4PAIN)}. The proposed method introduces a pipeline that leverages respiration as the input signal and incorporates a highly efficient cross-attention transformer alongside a multi-windowing strategy. Extensive experiments demonstrate that respiration is a valuable physiological modality for pain assessment. Moreover, experiments revealed that compact and efficient models, when properly optimized, can achieve strong performance, often surpassing larger counterparts. The proposed multi-window approach effectively captures both short-term and long-term features, as well as global characteristics, thereby enhancing the model's representational capacity.
Authors: Shuquan Lian, Yuhang Wu, Jia Ma, Zihan Song, Bingqi Chen, Xiawu Zheng, Hui Li
Abstract: The emergence of Multimodal Large Language Models (MLLMs) has driven significant advances in Graphical User Interface (GUI) agent capabilities. Nevertheless, existing GUI agent training and inference techniques still suffer from a dilemma for reasoning designs, ineffective reward, and visual noise. To address these issues, we introduce UI-AGILE, a comprehensive framework enhancing GUI agents at both the training and inference stages. For training, we propose a suite of improvements to the Supervised Fine-Tuning (SFT) process: 1) a Continuous Reward function to incentivize high-precision grounding; 2) a "Simple Thinking" reward to balance planning with speed and grounding accuracy; and 3) a Cropping-based Resampling strategy to mitigate the sparse reward problem and improve learning on complex tasks. For inference, we present Decomposed Grounding with Selection, a novel method that dramatically improves grounding accuracy on high-resolution displays by breaking the image into smaller, manageable parts. Experiments show that UI-AGILE achieves the state-of-the-art performance on two benchmarks ScreenSpot-Pro and ScreenSpot-v2. For instance, using both our proposed training and inference enhancement methods brings 23% grounding accuracy improvement over the best baseline on ScreenSpot-Pro.
Authors: Yang Jiao, Kai Yang, Dongjin Song
Abstract: Distributionally Robust Optimization (DRO), which aims to find an optimal decision that minimizes the worst case cost over the ambiguity set of probability distribution, has been widely applied in diverse applications, e.g., network behavior analysis, risk management, etc. However, existing DRO techniques face three key challenges: 1) how to deal with the asynchronous updating in a distributed environment; 2) how to leverage the prior distribution effectively; 3) how to properly adjust the degree of robustness according to different scenarios. To this end, we propose an asynchronous distributed algorithm, named Asynchronous Single-looP alternatIve gRadient projEction (ASPIRE) algorithm with the itErative Active SEt method (EASE) to tackle the federated distributionally robust optimization (FDRO) problem. Furthermore, a new uncertainty set, i.e., constrained D-norm uncertainty set, is developed to effectively leverage the prior distribution and flexibly control the degree of robustness. Finally, our theoretical analysis elucidates that the proposed algorithm is guaranteed to converge and the iteration complexity is also analyzed. Extensive empirical studies on real-world datasets demonstrate that the proposed method can not only achieve fast convergence, and remain robust against data heterogeneity as well as malicious attacks, but also tradeoff robustness with performance.
Authors: Chen Zhang, Qiuchi Li, Dawei Song, Zheyu Ye, Yan Gao, Yan Hu
Abstract: Language model (LM) distillation aims at distilling the knowledge in a large teacher LM to a small student one. As a critical issue facing LM distillation, a superior student often arises from a teacher of a relatively small scale instead of a larger one, especially in the presence of substantial capacity gap between the teacher and student. This issue, often referred to as the \textit{curse of capacity gap}, suggests that there is likely an optimal teacher yielding the best-performing student along the scaling course of the teacher. Consequently, distillation trials on teachers of a wide range of scales are called for to determine the optimal teacher, which becomes computationally intensive in the context of large LMs (LLMs). This paper addresses this critical bottleneck by providing the \textit{law of capacity gap} inducted from a preliminary study on distilling a broad range of small-scale (<3B) LMs, where the optimal teacher consistently scales linearly with the student scale across different model and data scales. By extending the law to LLM distillation on a larger scale (7B), we succeed in obtaining versatile LLMs that outperform a wide array of competitors.
Authors: Xiaojin Zhang, Wei Chen
Abstract: The widespread adoption of machine learning necessitates robust privacy protection alongside algorithmic resilience. While Local Differential Privacy (LDP) provides foundational guarantees, sophisticated adversaries with prior knowledge demand more nuanced Bayesian privacy notions, such as Maximum Bayesian Privacy (MBP) and Average Bayesian Privacy (ABP), first introduced by \cite{zhang2022no}. Concurrently, machine learning systems require inherent robustness against data perturbations and adversarial manipulations. This paper systematically investigates the intricate theoretical relationships among LDP, MBP, and ABP. Crucially, we bridge these privacy concepts with algorithmic robustness, particularly within the Probably Approximately Correct (PAC) learning framework. Our work demonstrates that privacy-preserving mechanisms inherently confer PAC robustness. We present key theoretical results, including the formalization of the established LDP-MBP relationship, novel bounds between MBP and ABP, and a proof demonstrating PAC robustness from MBP. Furthermore, we establish a novel theoretical relationship quantifying how privacy leakage directly influences an algorithm's input robustness. These results provide a unified theoretical framework for understanding and optimizing the privacy-robustness trade-off, paving the way for the development of more secure, trustworthy, and resilient machine learning systems.
Authors: Akash Ghosh, Arkadeep Acharya, Sriparna Saha, Vinija Jain, Aman Chadha
Abstract: The advent of Large Language Models (LLMs) has significantly reshaped the trajectory of the AI revolution. Nevertheless, these LLMs exhibit a notable limitation, as they are primarily adept at processing textual information. To address this constraint, researchers have endeavored to integrate visual capabilities with LLMs, resulting in the emergence of Vision-Language Models (VLMs). These advanced models are instrumental in tackling more intricate tasks such as image captioning and visual question answering. In our comprehensive survey paper, we delve into the key advancements within the realm of VLMs. Our classification organizes VLMs into three distinct categories: models dedicated to vision-language understanding, models that process multimodal inputs to generate unimodal (textual) outputs and models that both accept and produce multimodal inputs and outputs.This classification is based on their respective capabilities and functionalities in processing and generating various modalities of data.We meticulously dissect each model, offering an extensive analysis of its foundational architecture, training data sources, as well as its strengths and limitations wherever possible, providing readers with a comprehensive understanding of its essential components. We also analyzed the performance of VLMs in various benchmark datasets. By doing so, we aim to offer a nuanced understanding of the diverse landscape of VLMs. Additionally, we underscore potential avenues for future research in this dynamic domain, anticipating further breakthroughs and advancements.
Authors: Xinzhe Zheng, Sijie Ji, Yipeng Pan, Kaiwen Zhang, Chenshu Wu
Abstract: Inertial tracking is vital for robotic IoT and has gained popularity thanks to the ubiquity of low-cost inertial measurement units and deep learning-powered tracking algorithms. Existing works, however, have not fully utilized IMU measurements, particularly magnetometers, nor have they maximized the potential of deep learning to achieve the desired accuracy. To address these limitations, we introduce NeurIT, which elevates tracking accuracy to a new level. NeurIT employs a Time-Frequency Block-recurrent Transformer (TF-BRT) at its core, combining both RNN and Transformer to learn representative features in both time and frequency domains. To fully utilize IMU information, we strategically employ body-frame differentiation of magnetometers, considerably reducing the tracking error. We implement NeurIT on a customized robotic platform and conduct evaluation in various indoor environments. Experimental results demonstrate that NeurIT achieves a mere 1-meter tracking error over a 300-meter distance. Notably, it significantly outperforms state-of-the-art baselines by 48.21% on unseen data. Moreover, NeurIT demonstrates robustness in large urban complexes and performs comparably to the visual-inertial approach (Tango Phone) in vision-favored conditions while surpassing it in feature-sparse settings. We believe NeurIT takes an important step forward toward practical neural inertial tracking for ubiquitous and scalable tracking of robotic things. NeurIT is open-sourced here: https://github.com/aiot-lab/NeurIT.
Authors: Pascal J. Sager, Jan M. Deriu, Benjamin F. Grewe, Thilo Stadelmann, Christoph von der Malsburg
Abstract: We introduce the Cooperative Network Architecture (CNA), a model that represents sensory signals using structured, recurrently connected networks of neurons, termed "nets." Nets are dynamically assembled from overlapping net fragments, which are learned based on statistical regularities in sensory input. This architecture offers robustness to noise, deformation, and out-of-distribution data, addressing challenges in current vision systems from a novel perspective. We demonstrate that net fragments can be learned without supervision and flexibly recombined to encode novel patterns, enabling figure completion and resilience to noise. Our findings establish CNA as a promising paradigm for developing neural representations that integrate local feature processing with global structure formation, providing a foundation for future research on invariant object recognition.
Authors: Marc Schmitt
Abstract: The integration of Artificial Intelligence (AI) into corporate strategy has become critical for organizations seeking to maintain a competitive advantage in the digital age. As AI transforms business models, operations, and decision-making, the need for dedicated executive leadership to guide, govern, and orchestrate this transformation becomes increasingly evident. This paper examines emerging future scenarios across three domains: the AI Economy, the AI Organization, and Competition in the Age of AI. These domains reveal environmental, structural, and strategic tensions that existing C-suite roles struggle to resolve. In response, the paper develops a theory-informed framework for the Chief AI Officer (CAIO), outlining the distinct functions and capabilities required to guide and govern AI at scale. Drawing on illustrative cases and emerging practice, this conceptualization clarifies the CAIOs unique role within the executive landscape and presents a forward-looking research agenda. This paper advances the discourse on AI leadership by offering a theory-driven rationale for the strategic integration of AI at the executive level and by positioning the Chief AI Officer as a distinct and necessary role within modern organizations.
Authors: Hamed Taherkhani, Melika Sepindband, Hung Viet Pham, Song Wang, Hadi Hemmati
Abstract: Large Language Models have seen increasing use in various software development tasks, especially in code generation. The most advanced recent methods attempt to incorporate feedback from code execution into prompts to help guide LLMs in generating correct code in an iterative process. While effective, these methods could be costly due to numerous interactions with the LLM and extensive token usage. To address this issue, we propose an alternative approach named Evolutionary Prompt Engineering for Code (EPiC), which leverages a lightweight evolutionary algorithm to refine the original prompts into improved versions that generate high quality code, with minimal interactions with the LLM. Our evaluation against state-of-the-art (SOTA) LLM based code generation agents shows that EPiC not only achieves up to 6% improvement in pass@k but is also 2-10 times more cost-effective than the baselines.
Authors: Nianqi Li, Siyu Yuan, Jiangjie Chen, Jiaqing Liang, Feng Wei, Zujie Liang, Deqing Yang, Yanghua Xiao
Abstract: Historical analogies, which compare known past events with contemporary but unfamiliar events, are important abilities that help people make decisions and understand the world. However, research in applied history suggests that people have difficulty finding appropriate analogies. And previous studies in the AI community have also overlooked historical analogies. To fill this gap, in this paper, we focus on the historical analogy acquisition task, which aims to acquire analogous historical events for a given event. We explore retrieval and generation methods for acquiring historical analogies based on different large language models (LLMs). Furthermore, we propose a self-reflection method to mitigate hallucinations and stereotypes when LLMs generate historical analogies. Through human evaluations and our specially designed automatic multi-dimensional assessment, we find that LLMs generally have a good potential for historical analogies. And the performance of the models can be further improved by using our self-reflection method.
Authors: Franck Signe Talla, Edouard Grave, Herv\'e J\'egou
Abstract: We address the problem of extending a pretrained large language model to a new domain that was not seen during training. Standard techniques, such as finetuning or low-rank adaptation (LoRA) are successful at domain adaptation, but do not formally add capacity to the model. This often leads to a trade-off, between performing well on the new domain vs. degrading performance on the original domain. Here, we revisit and improve adapters to extend LLMs from three angles: data, architecture and training procedure, which are advantageously considered jointly. The resulting method, called neutral residues, modifies adapters in a way that leads each new residual block to output near-zeros on the original domain. This solution leads to strong results when adapting a state-of-the-art model originally trained on English to a new language. Neutral residues significantly outperform competing approaches such as finetuning, LoRA or vanilla adapters in terms of the trade-off between learning the new language and not forgetting English.
Authors: Sara Sarto, Nicholas Moratelli, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara
Abstract: Despite significant advancements in caption generation, existing evaluation metrics often fail to capture the full quality or fine-grained details of captions. This is mainly due to their reliance on non-specific human-written references or noisy pre-training data. Still, finding an effective metric is crucial not only for captions evaluation but also for the generation phase. Metrics can indeed play a key role in the fine-tuning stage of captioning models, ultimately enhancing the quality of the generated captions. In this paper, we propose PAC-S++, a learnable metric that leverages the CLIP model, pre-trained on both web-collected and cleaned data and regularized through additional pairs of generated visual and textual positive samples. Exploiting this stronger and curated pre-training, we also apply PAC-S++ as a reward in the Self-Critical Sequence Training (SCST) stage typically employed to fine-tune captioning models. Extensive experiments on different image and video datasets highlight the effectiveness of PAC-S++ compared to popular metrics for the task, including its sensitivity to object hallucinations. Furthermore, we show that integrating PAC-S++ into the fine-tuning stage of a captioning model results in semantically richer captions with fewer repetitions and grammatical errors. Evaluations on out-of-domain benchmarks further demonstrate the efficacy of our fine-tuning approach in enhancing model capabilities. Source code and trained models are publicly available at: https://github.com/aimagelab/pacscore.
Authors: Farid Ariai, Joel Mackenzie, Gianluca Demartini
Abstract: Natural Language Processing (NLP) is revolutionising the way both professionals and laypersons operate in the legal field. The considerable potential for NLP in the legal sector, especially in developing computational assistance tools for various legal processes, has captured the interest of researchers for years. This survey follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses framework, reviewing 154 studies, with a final selection of 131 after manual filtering. It explores foundational concepts related to NLP in the legal domain, illustrating the unique aspects and challenges of processing legal texts, such as extensive document lengths, complex language, and limited open legal datasets. We provide an overview of NLP tasks specific to legal text, such as Document Summarisation, Named Entity Recognition, Question Answering, Argument Mining, Text Classification, and Judgement Prediction. Furthermore, we analyse both developed legal-oriented language models, and approaches for adapting general-purpose language models to the legal domain. Additionally, we identify sixteen open research challenges, including the detection and mitigation of bias in artificial intelligence applications, the need for more robust and interpretable models, and improving explainability to handle the complexities of legal language and reasoning.
Authors: Yichen He, Yuan Lin, Jianchao Wu, Hanchong Zhang, Yuchen Zhang, Ruicheng Le
Abstract: Existing large vision-language models (LVLMs) are largely limited to processing short, seconds-long videos and struggle with generating coherent descriptions for extended video spanning minutes or more. Long video description introduces new challenges, such as consistent character identification and plot-level descriptions incorporating both visual and audio information. To address these, we figure out audio-visual character identification, matching character names to each dialogue, as a key factor. We propose StoryTeller, a system for generating dense descriptions of long videos, incorporating both low-level visual concepts and high-level plot information. StoryTeller uses a multimodal large language model that integrates visual, audio, and text modalities to perform audio-visual character identification on minute-long video clips. The results are then fed into a LVLM to enhance consistency of video description. We validate our approach on movie description tasks and introduce MovieStory101, a dataset with dense descriptions for three-minute movie clips. To evaluate long video descriptions, we create StoryQA, a large set of multiple-choice questions for MovieStory101 test set. We assess descriptions by inputting them into GPT-4 to answer these questions, using accuracy as an automatic evaluation metric. Experiments show that StoryTeller outperforms all open and closed-source baselines on StoryQA, achieving 9.5% higher accuracy than the strongest baseline, Gemini-1.5-pro, and demonstrating a +15.56% advantage in human side-by-side evaluations. Additionally, incorporating audio-visual character identification from StoryTeller improves the performance of all video description models, with Gemini-1.5-pro and GPT-4o showing relative improvement of 5.5% and 13.0%, respectively, in accuracy on StoryQA.
Authors: Hamed Taherkhani, Jiho Shin, Muhammad Ammar Tahir, Md Rakib Hossain Misu, Vineet Sunil Gattani, Hadi Hemmati
Abstract: Modern Large Language Model (LLM)-based programming agents often rely on test execution feedback to refine their generated code. These tests are synthetically generated by LLMs. However, LLMs may produce invalid or hallucinated test cases, which can mislead feedback loops and degrade the performance of agents in refining and improving code. This paper introduces VALTEST, a novel framework that leverages semantic entropy to automatically validate test cases generated by LLMs. Analyzing the semantic structure of test cases and computing entropy-based uncertainty measures, VALTEST trains a machine learning model to classify test cases as valid or invalid and filters out invalid test cases. Experiments on multiple benchmark datasets and various LLMs show that VALTEST not only boosts test validity by up to 29% but also improves code generation performance, as evidenced by significant increases in pass@1 scores. Our extensive experiments also reveal that semantic entropy is a reliable indicator to distinguish between valid and invalid test cases, which provides a robust solution for improving the correctness of LLM-generated test cases used in software testing and code generation.
Authors: Pan Liao, Feng Yang, Di Wu, Jinwen Yu, Wenhui Zhao, Dingwen Zhang
Abstract: Transformer-based multi-object tracking (MOT) methods have captured the attention of many researchers in recent years. However, these models often suffer from slow inference speeds due to their structure or other issues. To address this problem, we revisited the Joint Detection and Tracking (JDT) method by looking back at past approaches. By integrating the original JDT approach with some advanced theories, this paper employs an efficient method of information transfer between frames on the DETR, constructing a fast and novel JDT-type MOT framework: FastTrackTr. Thanks to the superiority of this information transfer method, our approach not only reduces the number of queries required during tracking but also avoids the excessive introduction of network structures, ensuring model simplicity. Experimental results indicate that our method has the potential to achieve real-time tracking and exhibits competitive tracking accuracy across multiple datasets.
Authors: Hortense Fong, George Gui
Abstract: Understanding when and why consumers engage with stories is crucial for content creators and platforms. While existing theories suggest that audience beliefs of what is going to happen should play an important role in engagement decisions, empirical work has mostly focused on developing techniques to directly extract features from actual content, rather than capturing forward-looking beliefs, due to the lack of a principled way to model such beliefs in unstructured narrative data. To complement existing feature extraction techniques, this paper introduces a novel framework that leverages large language models to model audience forward-looking beliefs about how stories might unfold. Our method generates multiple potential continuations for each story and extracts features related to expectations, uncertainty, and surprise using established content analysis techniques. Applying our method to over 30,000 book chapters, we demonstrate that our framework complements existing feature engineering techniques by amplifying their marginal explanatory power on average by 31%. The results reveal that different types of engagement-continuing to read, commenting, and voting-are driven by distinct combinations of current and anticipated content features. Our framework provides a novel way to study and explore how audience forward-looking beliefs shape their engagement with narrative media, with implications for marketing strategy in content-focused industries.
Authors: Zhongjian Hu, Peng Yang, Bing Li, Fengyuan Liu
Abstract: Recently, Large Language Models (LLMs) have been used for knowledge-based Visual Question Answering (VQA). Despite the encouraging results of previous studies, prior methods prompt LLMs to predict answers directly, neglecting intermediate thought processes. We argue that prior methods do not sufficiently activate the capacities of LLMs. We propose a framework called PLRH that Prompts LLMs with Rationale Heuristics for knowledge-based VQA. The PLRH prompts LLMs with Chain of Thought (CoT) to generate rationale heuristics, i.e., intermediate thought processes, and then leverages the rationale heuristics to inspire LLMs to predict answers. Experiments show that our approach outperforms the existing baselines by more than 2.2 and 2.1 on OK-VQA and A-OKVQA, respectively.
Authors: Wenkun He, Yun Liu, Ruitao Liu, Li Yi
Abstract: Synthesizing realistic human-object interaction motions is a critical problem in VR/AR and human animation. Unlike the commonly studied scenarios involving a single human or hand interacting with one object, we address a more generic multi-body setting with arbitrary numbers of humans, hands, and objects. This complexity introduces significant challenges in synchronizing motions due to the high correlations and mutual influences among bodies. To address these challenges, we introduce SyncDiff, a novel method for multi-body interaction synthesis using a synchronized motion diffusion strategy. SyncDiff employs a single diffusion model to capture the joint distribution of multi-body motions. To enhance motion fidelity, we propose a frequency-domain motion decomposition scheme. Additionally, we introduce a new set of alignment scores to emphasize the synchronization of different body motions. SyncDiff jointly optimizes both data sample likelihood and alignment likelihood through an explicit synchronization strategy. Extensive experiments across four datasets with various multi-body configurations demonstrate the superiority of SyncDiff over existing state-of-the-art motion synthesis methods.
Authors: Haoxiang Gao, Li Zhang, Yu Zhao, Zhou Yang, Jinghan Cao
Abstract: Vision-language models (VLMs) have become a promising approach to enhancing perception and decision-making in autonomous driving. The gap remains in applying VLMs to understand complex scenarios interacting with pedestrians and efficient vehicle deployment. In this paper, we propose a knowledge distillation method that transfers knowledge from large-scale vision-language foundation models to efficient vision networks, and we apply it to pedestrian behavior prediction and scene understanding tasks, achieving promising results in generating more diverse and comprehensive semantic attributes. We also utilize multiple pre-trained models and ensemble techniques to boost the model's performance. We further examined the effectiveness of the model after knowledge distillation; the results show significant metric improvements in open-vocabulary perception and trajectory prediction tasks, which can potentially enhance the end-to-end performance of autonomous driving.
Authors: Ziqing Wen, Ping Luo, Jiahuan Wang, Xiaoge Deng, Jinping Zou, Kun Yuan, Tao Sun, Dongsheng Li
Abstract: Large language models (LLMs) have shown impressive performance across a range of natural language processing tasks. However, their vast number of parameters introduces significant memory challenges during training, particularly when using memory-intensive optimizers like Adam. Existing memory-efficient algorithms often rely on techniques such as singular value decomposition projection or weight freezing. While these approaches help alleviate memory constraints, they generally produce suboptimal results compared to full-rank updates. In this paper, we investigate the memory-efficient method beyond low-rank training, proposing a novel solution called Gradient Wavelet Transform (GWT), which applies wavelet transforms to gradients in order to significantly reduce the memory requirements for maintaining optimizer states. We demonstrate that GWT can be seamlessly integrated with memory-intensive optimizers, enabling efficient training without sacrificing performance. Through extensive experiments on both pre-training and fine-tuning tasks, we show that GWT achieves state-of-the-art performance compared with advanced memory-efficient optimizers and full-rank approaches in terms of both memory usage and training performance.
Authors: Taiki Yamada, Yuichi Katori, Kantaro Fujiwara
Abstract: Echo state networks (ESNs) are a class of recurrent neural networks in which only the readout layer is trainable, while the recurrent and input layers are fixed. This architectural constraint enables computationally efficient processing of time-series data. Traditionally, the readout layer in ESNs is trained using supervised learning with target outputs. In this study, we focus on input reconstruction (IR), where the readout layer is trained to reconstruct the input time series fed into the ESN. We show that IR can be achieved through unsupervised learning (UL), without access to supervised targets, provided that the ESN parameters are known a priori and satisfy invertibility conditions. This formulation allows applications relying on IR, such as dynamical system replication and noise filtering, to be reformulated within the UL framework via straightforward integration with existing algorithms. Our results suggest that prior knowledge of ESN parameters can reduce reliance on supervision, thereby establishing a new principle: not only by fixing part of the network parameters but also by exploiting their specific values. Furthermore, our UL-based algorithms for input reconstruction and related tasks are suitable for autonomous processing, offering insights into how analogous computational mechanisms might operate in the brain in principle. These findings contribute to a deeper understanding of the mathematical foundations of ESNs and their relevance to models in computational neuroscience.
Authors: Yisong Chen, Chuqing Zhao, Yixin Xu, Chuanhao Nie, Yixin Zhang
Abstract: This paper systematically reviews advancements in deep learning (DL) techniques for financial fraud detection, a critical issue in the financial sector. Using the Kitchenham systematic literature review approach, 57 studies published between 2019 and 2024 were analyzed. The review highlights the effectiveness of various deep learning models such as Convolutional Neural Networks, Long Short-Term Memory, and transformers across domains such as credit card transactions, insurance claims, and financial statement audits. Performance metrics such as precision, recall, F1-score, and AUC-ROC were evaluated. Key themes explored include the impact of data privacy frameworks and advancements in feature engineering and data preprocessing. The study emphasizes challenges such as imbalanced datasets, model interpretability, and ethical considerations, alongside opportunities for automation and privacy-preserving techniques such as blockchain integration and Principal Component Analysis. By examining trends over the past five years, this review identifies critical gaps and promising directions for advancing DL applications in financial fraud detection, offering actionable insights for researchers and practitioners.
Authors: Hajar Emami Gohari, Swanand Ravindra Kadhe, Syed Yousaf Shah, Constantin Adam, Abdulhamid Adebayo, Praneet Adusumilli, Farhan Ahmed, Nathalie Baracaldo Angel, Santosh Subhashrao Borse, Yuan-Chi Chang, Xuan-Hong Dang, Nirmit Desai, Revital Eres, Ran Iwamoto, Alexei Karve, Yan Koyfman, Wei-Han Lee, Changchang Liu, Boris Lublinsky, Takuyo Ohko, Pablo Pesce, Maroun Touma, Shiqiang Wang, Shalisha Witherspoon, Herbert Woisetschl\"ager, David Wood, Kun-Lung Wu, Issei Yoshida, Syed Zawad, Petros Zerfos, Yi Zhou, Bishwaranjan Bhattacharjee
Abstract: Data quantity and quality play a vital role in determining the performance of Large Language Models (LLMs). High-quality data, in particular, can significantly boost the LLM's ability to generalize on a wide range of downstream tasks. Large pre-training datasets for leading LLMs remain inaccessible to the public, whereas many open datasets are small in size (less than 5 trillion tokens), limiting their suitability for training large models. In this paper, we introduce GneissWeb, a large dataset yielding around 10 trillion tokens that caters to the data quality and quantity requirements of training LLMs. Our GneissWeb recipe that produced the dataset consists of sharded exact sub-string deduplication and a judiciously constructed ensemble of quality filters. GneissWeb achieves a favorable trade-off between data quality and quantity, producing models that outperform models trained on state-of-the-art open large datasets (5+ trillion tokens). We show that models trained using GneissWeb dataset outperform those trained on FineWeb-V1.1.0 by 2.73 percentage points in terms of average score computed on a set of 11 commonly used benchmarks (both zero-shot and few-shot) for pre-training dataset evaluation. When the evaluation set is extended to 20 benchmarks (both zero-shot and few-shot), models trained using GneissWeb still achieve a 1.75 percentage points advantage over those trained on FineWeb-V1.1.0.
Authors: Nian Shao, Rui Zhou, Pengyu Wang, Xian Li, Ying Fang, Yujie Yang, Xiaofei Li
Abstract: In this work, we propose CleanMel, a single-channel Mel-spectrogram denoising and dereverberation network for improving both speech quality and automatic speech recognition (ASR) performance. The proposed network takes as input the noisy and reverberant microphone recording and predicts the corresponding clean Mel-spectrogram. The enhanced Mel-spectrogram can be either transformed to the speech waveform with a neural vocoder or directly used for ASR. The proposed network is composed of interleaved cross-band and narrow-band processing in the Mel-frequency domain, for learning the full-band spectral pattern and the narrow-band properties of signals, respectively. Compared to linear-frequency domain or time-domain speech enhancement, the key advantage of Mel-spectrogram enhancement is that Mel-frequency presents speech in a more compact way and thus is easier to learn, which will benefit both speech quality and ASR. Experimental results on five English and one Chinese datasets demonstrate a significant improvement in both speech quality and ASR performance achieved by the proposed model.Code and audio examples of our model are available online.
Authors: Xiang Xiang, Zhuo Xu, Yao Deng, Qinhao Zhou, Yifan Liang, Ke Chen, Qingfang Zheng, Yaowei Wang, Xilin Chen, Wen Gao
Abstract: The advancement of remote sensing, including satellite systems, facilitates the continuous acquisition of remote sensing imagery globally, introducing novel challenges for achieving open-world tasks. Deployed models need to continuously adjust to a constant influx of new data, which frequently exhibits diverse shifts from the data encountered during the training phase. To effectively handle the new data, models are required to detect semantic shifts, adapt to covariate shifts, and continuously update their parameters without forgetting learned knowledge, as has been considered in works on a variety of open-world tasks. However, existing studies are typically conducted within a single dataset to simulate realistic conditions, with a lack of large-scale benchmarks capable of evaluating multiple open-world tasks. In this paper, we introduce \textbf{OpenEarthSensing (OES)}, a large-scale fine-grained benchmark for open-world remote sensing. OES includes 189 scene and object categories, covering the vast majority of potential semantic shifts that may occur in the real world. Additionally, to provide a more comprehensive testbed for evaluating the generalization performance, OES encompasses five data domains with significant covariate shifts, including two RGB satellite domains, one RGB aerial domain, one multispectral RGB domain, and one infrared domain. We evaluate the baselines and existing methods for diverse tasks on OES, demonstrating that it serves as a meaningful and challenging benchmark for open-world remote sensing. The proposed dataset OES is available at https://haiv-lab.github.io/OES.
Authors: Julia Kreutzer, Eleftheria Briakou, Sweta Agrawal, Marzieh Fadaee, Kocmi Tom
Abstract: Generation capabilities and language coverage of multilingual large language models (mLLMs) are advancing rapidly. However, evaluation practices for generative abilities of mLLMs are still lacking comprehensiveness, scientific rigor, and consistent adoption across research labs, which undermines their potential to meaningfully guide mLLM development. We draw parallels with machine translation (MT) evaluation, a field that faced similar challenges and has, over decades, developed transparent reporting standards and reliable evaluations for multilingual generative models. Through targeted experiments across key stages of the generative evaluation pipeline, we demonstrate how best practices from MT evaluation can deepen the understanding of quality differences between models. Additionally, we identify essential components for robust meta-evaluation of mLLMs, ensuring the evaluation methods themselves are rigorously assessed. We distill these insights into a checklist of actionable recommendations for mLLM research and development.
Authors: Huiyi Chen, Jiawei Peng, Kaihua Tang, Xin Geng, Xu Yang
Abstract: In-context learning (ICL) enables Large Vision-Language Models (LVLMs) to adapt to new tasks without parameter updates, using a few demonstrations from a large support set. However, selecting informative demonstrations leads to high computational and memory costs. While some methods explore selecting a small and representative coreset in the text classification, evaluating all support set samples remains costly, and discarded samples lead to unnecessary information loss. These methods may also be less effective for image classification due to differences in feature spaces. Given these limitations, we propose Key-based Coreset Optimization (KeCO), a novel framework that leverages untapped data to construct a compact and informative coreset. We introduce visual features as keys within the coreset, which serve as the anchor for identifying samples to be updated through different selection strategies. By leveraging untapped samples from the support set, we update the keys of selected coreset samples, enabling the randomly initialized coreset to evolve into a more informative coreset under low computational cost. Through extensive experiments on coarse-grained and fine-grained image classification benchmarks, we demonstrate that KeCO effectively enhances ICL performance for image classification task, achieving an average improvement of more than 20\%. Notably, we evaluate KeCO under a simulated online scenario, and the strong performance in this scenario highlights the practical value of our framework for resource-constrained real-world scenarios.
Authors: Andrea Cremaschi, Dae-Jin Lee, Manuele Leonelli
Abstract: As artificial intelligence and robotics increasingly reshape the global labor market, understanding public perceptions of these technologies becomes critical. We examine how these perceptions have evolved across Latin America, using survey data from the 2017, 2018, 2020, and 2023 waves of the Latinobar\'ometro. Drawing on responses from over 48,000 individuals across 16 countries, we analyze fear of job loss due to artificial intelligence and robotics. Using statistical modeling and latent class analysis, we identify key structural and ideological predictors of concern, with education level and political orientation emerging as the most consistent drivers. Our findings reveal substantial temporal and cross-country variation, with a notable peak in fear during 2018 and distinct attitudinal profiles emerging from latent segmentation. These results offer new insights into the social and structural dimensions of AI anxiety in emerging economies and contribute to a broader understanding of public attitudes toward automation beyond the Global North.
Authors: Yimao Guo, Zuomin Qu, Wei Lu, Xiangyang Luo
Abstract: With the increasing prevalence of diffusion-based malicious image manipulation, existing proactive defense methods struggle to safeguard images against tampering under unknown conditions. To address this, we propose Anti-Inpainting, a proactive defense approach that achieves protection comprising three novel modules. First, we introduce a multi-level deep feature extractor to obtain intricate features from the diffusion denoising process, enhancing protective effectiveness. Second, we design a multi-scale, semantic-preserving data augmentation technique to enhance the transferability of adversarial perturbations across unknown conditions. Finally, we propose a selection-based distribution deviation optimization strategy to bolster protection against manipulations guided by diverse random seeds. Extensive experiments on InpaintGuardBench and CelebA-HQ demonstrate that Anti-Inpainting effectively defends against diffusion-based inpainters under unknown conditions. Additionally, our approach demonstrates robustness against various image purification methods and transferability across different diffusion model versions.
Authors: Ryo Bertolissi, Jonas H\"ubotter, Ido Hakimi, Andreas Krause
Abstract: Mixture of expert (MoE) models are a promising approach to increasing model capacity without increasing inference cost, and are core components of many state-of-the-art language models. However, current MoE models typically use only few experts due to prohibitive training and inference cost. We propose Test-Time Model Merging (TTMM) which scales the MoE paradigm to an order of magnitude more experts and uses model merging to avoid almost any test-time overhead. We show that TTMM is an approximation of test-time training (TTT), which fine-tunes an expert model for each prediction task, i.e., prompt. TTT has recently been shown to significantly improve language models, but is computationally expensive. We find that performance of TTMM improves with more experts and approaches the performance of TTT. Moreover, we find that with a 1B parameter base model, TTMM is more than 100x faster than TTT at test-time by amortizing the cost of TTT at train-time. Thus, TTMM offers a promising cost-effective approach to scale test-time training.
Authors: Haiyan Zhao, Xuansheng Wu, Fan Yang, Bo Shen, Ninghao Liu, Mengnan Du
Abstract: Linear concept vectors effectively steer LLMs, but existing methods suffer from noisy features in diverse datasets that undermine steering robustness. We propose Sparse Autoencoder-Denoised Concept Vectors (SDCV), which selectively keep the most discriminative SAE latents while reconstructing hidden representations. Our key insight is that concept-relevant signals can be explicitly separated from dataset noise by scaling up activations of top-k latents that best differentiate positive and negative samples. Applied to linear probing and difference-in-mean, SDCV consistently improves steering success rates by 4-16\% across six challenging concepts, while maintaining topic relevance.
Authors: Benjamin Turtel, Danny Franklin, Kris Skotheim, Luke Hewitt, Philipp Schoenegger
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has been an effective approach for improving Large Language Models' reasoning in domains such as coding and mathematics. Here, we apply RLVR methods towards forecasting future real-world events - a challenging task for RL due to the very noisy (and delayed) outcomes involved. Using a novel dataset of recent questions from a prediction market, and accompanying relevant news headlines, we show that a compact (14B) reasoning model can be trained to match or surpass the predictive accuracy of frontier models like o1, while greatly improving probabilistic calibration. The model's performance is also practically meaningful: in a Polymarket trading simulation, we estimate that its bets would have yielded a return on investment of over 10% across all questions in the test set. We detail and compare approaches used in training our model, including augmenting our training-data with synthetic prediction questions, guardrails for learning stability, and median prediction sampling at inference-time.
Authors: Peiran Sun
Abstract: Adversarial attack reveals the vulnerability of deep learning models. For about a decade, countless attack and defense methods have been proposed, leading to robustified classifiers and better understanding of models. Among these methods, curvature-based approaches have attracted attention because it is assumed that high curvature may give rise to rough decision boundary. However, the most commonly used \textit{curvature} is the curvature of loss function, scores or other parameters from within the model as opposed to decision boundary curvature, since the former can be relatively easily formed using second order derivative. In this paper, we propose a new query-efficient method, dynamic curvature estimation(DCE), to estimate the decision boundary curvature in a black-box setting. Our approach is based on CGBA, a black-box adversarial attack. By performing DCE on a wide range of classifiers, we discovered, statistically, a connection between decision boundary curvature and adversarial robustness. We also propose a new attack method, curvature dynamic black-box attack(CDBA) with improved performance using the dynamically estimated curvature.
Authors: Dane Malenfant, Blake A. Richards
Abstract: Sometimes we benefit from actions that others have taken even when we are unaware that they took those actions. For example, if your neighbor chooses not to take a parking spot in front of your house when you are not there, you can benefit, even without being aware that they took this action. These "hidden gifts" represent an interesting challenge for multi-agent reinforcement learning (MARL), since assigning credit when the beneficial actions of others are hidden is non-trivial. Here, we study the impact of hidden gifts with a very simple MARL task. In this task, agents in a grid-world environment have individual doors to unlock in order to obtain individual rewards. As well, if all the agents unlock their door the group receives a larger collective reward. However, there is only one key for all of the doors, such that the collective reward can only be obtained when the agents drop the key for others after they use it. Notably, there is nothing to indicate to an agent that the other agents have dropped the key, thus the act of dropping the key for others is a "hidden gift". We show that several different state-of-the-art RL algorithms, including MARL algorithms, fail to learn how to obtain the collective reward in this simple task. Interestingly, we find that independent model-free policy gradient agents can solve the task when we provide them with information about their own action history, but MARL agents still cannot solve the task with action history. Finally, we derive a correction term for these independent agents, inspired by learning aware approaches, which reduces the variance in learning and helps them to converge to collective success more reliably. These results show that credit assignment in multi-agent settings can be particularly challenging in the presence of "hidden gifts", and demonstrate that learning awareness in independent agents can benefit these settings.
Authors: Ziming Wang, Nan Xue, Rebecka J\"ornsten
Abstract: The goal of point cloud assembly is to reconstruct a complete 3D shape by aligning multiple point cloud pieces. This work presents a novel equivariant solver for assembly tasks based on flow matching models. We first theoretically show that the key to learning equivariant distributions via flow matching is to learn related vector fields. Based on this result, we propose an assembly model, called equivariant diffusion assembly (Eda), which learns related vector fields conditioned on the input pieces. We further construct an equivariant path for Eda, which guarantees high data efficiency of the training process. Our numerical results show that Eda is highly competitive on practical datasets, and it can even handle the challenging situation where the input pieces are non-overlapped.
Authors: Nadezhda Chirkova, Tunde Oluwaseyi Ajayi, Seth Aycock, Zain Muhammad Mujahid, Vladana Perli\'c, Ekaterina Borisova, Markarit Vartampetian
Abstract: Prompting large language models (LLMs) to evaluate generated text, known as LLM-as-a-judge, has become a standard evaluation approach in natural language generation (NLG), but is primarily used as a quantitative tool, i.e. with numerical scores as main outputs. In this work, we propose LLM-as-a-qualitative-judge, an LLM-based evaluation approach with the main output being a structured report of common issue types in the NLG system outputs. Our approach is targeted at providing developers with meaningful insights on what improvements can be done to a given NLG system and consists of two main steps, namely open-ended per-instance issue analysis and clustering of the discovered issues using an intuitive cumulative algorithm. We also introduce a strategy for evaluating the proposed approach, coupled with ~300 annotations of issues in instances from 12 NLG datasets. Our results show that LLM-as-a-qualitative-judge correctly recognizes instance-specific issues in 2/3 cases and is capable of producing error type reports resembling the reports composed by human annotators. Our code and data are publicly available at https://github.com/tunde-ajayi/llm-as-a-qualitative-judge.
URLs: https://github.com/tunde-ajayi/llm-as-a-qualitative-judge.
Authors: Vinicius L. S. Silva, Gabriel S. Seabra, Alexandre A. Emerick
Abstract: We propose two new methods based/inspired by machine learning for tabular data and distance-free localization to enhance the covariance estimations in an ensemble data assimilation. The main goal is to enhance the data assimilation results by mitigating loss of variance due to sampling errors. We also analyze the suitability of several machine learning models and the balance between accuracy and computational cost of the covariance estimations. We introduce two distance-free localization techniques leveraging machine learning methods specifically tailored for tabular data. The methods are integrated into the Ensemble Smoother with Multiple Data Assimilation (ES-MDA) framework. The results show that the proposed localizations improve covariance accuracy and enhance data assimilation and uncertainty quantification results. We observe reduced variance loss for the input variables using the proposed methods. Furthermore, we compare several machine learning models, assessing their suitability for the problem in terms of computational cost, and quality of the covariance estimation and data match. The influence of ensemble size is also investigated, providing insights into balancing accuracy and computational efficiency. Our findings demonstrate that certain machine learning models are more suitable for this problem. This study introduces two novel methods that mitigate variance loss for model parameters in ensemble-based data assimilation, offering practical solutions that are easy to implement and do not require any additional numerical simulation or hyperparameter tuning.
Authors: Rafi Al Attrach, Pedro Moreira, Rajna Fani, Renato Umeton, Leo Anthony Celi
Abstract: As ever-larger clinical datasets become available, they have the potential to unlock unprecedented opportunities for medical research. Foremost among them is Medical Information Mart for Intensive Care (MIMIC-IV), the world's largest open-source EHR database. However, the inherent complexity of these datasets, particularly the need for sophisticated querying skills and the need to understand the underlying clinical settings, often presents a significant barrier to their effective use. M3 lowers the technical barrier to understanding and querying MIMIC-IV data. With a single command it retrieves MIMIC-IV from PhysioNet, launches a local SQLite instance (or hooks into the hosted BigQuery), and-via the Model Context Protocol (MCP)-lets researchers converse with the database in plain English. Ask a clinical question in natural language; M3 uses a language model to translate it into SQL, executes the query against the MIMIC-IV dataset, and returns structured results alongside the underlying query for verifiability and reproducibility. Demonstrations show that minutes of dialogue with M3 yield the kind of nuanced cohort analyses that once demanded hours of handcrafted SQL and relied on understanding the complexities of clinical workflows. By simplifying access, M3 invites the broader research community to mine clinical critical-care data and accelerates the translation of raw records into actionable insight.
Authors: Yukang Chen, Wei Huang, Baifeng Shi, Qinghao Hu, Hanrong Ye, Ligeng Zhu, Zhijian Liu, Pavlo Molchanov, Jan Kautz, Xiaojuan Qi, Sifei Liu, Hongxu Yin, Yao Lu, Song Han
Abstract: We introduce a full-stack framework that scales up reasoning in vision-language models (VLMs) to long videos, leveraging reinforcement learning. We address the unique challenges of long video reasoning by integrating three critical components: (1) a large-scale dataset, LongVideo-Reason, comprising 104K long video QA pairs with high-quality reasoning annotations across diverse domains such as sports, games, and vlogs; (2) a two-stage training pipeline that extends VLMs with chain-of-thought supervised fine-tuning (CoT-SFT) and reinforcement learning (RL); and (3) a training infrastructure for long video RL, named Multi-modal Reinforcement Sequence Parallelism (MR-SP), which incorporates sequence parallelism and a vLLM-based engine tailored for long video, using cached video embeddings for efficient rollout and prefilling. In our experiments, LongVILA-R1-7B achieves strong performance on video benchmarks, reaching 65.1% and 71.1% accuracy on VideoMME without and with subtitles, respectively, and consistently outperforming LongVILA-7B across multiple benchmarks. Moreover, LongVILA-R1-7B supports processing up to 8,192 video frames per video, and configurable FPS settings. Notably, our MR-SP system achieves up to 2.1x speedup on long video RL training. In addition, we release our training system for public availability that supports RL training on various modalities (video, text, and audio), various models (VILA and Qwen series), and even image and video generation models. On a single A100 node (8 GPUs), it supports RL training on hour-long videos (e.g., 3,600 frames).
Authors: Shubham Shukla, Kunal Sonalkar
Abstract: The fashion retail business is centered around the capacity to comprehend products. Product attribution helps in comprehending products depending on the business process. Quality attribution improves the customer experience as they navigate through millions of products offered by a retail website. It leads to well-organized product catalogs. In the end, product attribution directly impacts the 'discovery experience' of the customer. Although large language models (LLMs) have shown remarkable capabilities in understanding multimodal data, their performance on fine-grained fashion attribute recognition remains under-explored. This paper presents a zero-shot evaluation of state-of-the-art LLMs that balance performance with speed and cost efficiency, mainly GPT-4o-mini and Gemini 2.0 Flash. We have used the dataset DeepFashion-MultiModal (https://github.com/yumingj/DeepFashion-MultiModal) to evaluate these models in the attribution tasks of fashion products. Our study evaluates these models across 18 categories of fashion attributes, offering insight into where these models excel. We only use images as the sole input for product information to create a constrained environment. Our analysis shows that Gemini 2.0 Flash demonstrates the strongest overall performance with a macro F1 score of 56.79% across all attributes, while GPT-4o-mini scored a macro F1 score of 43.28%. Through detailed error analysis, our findings provide practical insights for deploying these LLMs in production e-commerce product attribution-related tasks and highlight the need for domain-specific fine-tuning approaches. This work also lays the groundwork for future research in fashion AI and multimodal attribute extraction.
Authors: Junhong Min, Youngpil Jeon, Jimin Kim, Minyong Choi
Abstract: The pursuit of a generalizable stereo matching model, capable of performing well across varying resolutions and disparity ranges without dataset-specific fine-tuning, has revealed a fundamental trade-off. Iterative local search methods achieve high scores on constrained benchmarks, but their core mechanism inherently limits the global consistency required for true generalization. However, global matching architectures, while theoretically more robust, have historically been rendered infeasible by prohibitive computational and memory costs. We resolve this dilemma with $S^2M^2$: a global matching architecture that achieves state-of-the-art accuracy and high efficiency without relying on cost volume filtering or deep refinement stacks. Our design integrates a multi-resolution transformer for robust long-range correspondence, trained with a novel loss function that concentrates probability on feasible matches. This approach enables a more robust joint estimation of disparity, occlusion, and confidence. $S^2M^2$ establishes a new state of the art on Middlebury v3 and ETH3D benchmarks, significantly outperforming prior methods in most metrics while reconstructing high-quality details with competitive efficiency.
Authors: Jaros{\l}aw Byrka, Franciszek Malinka, Tomasz Ponitka
Abstract: We study the fair division of indivisible items and provide new insights into the EFX problem, which is widely regarded as the central open question in fair division, and the PMMS problem, a strictly stronger variant of EFX. Our first result constructs a three-agent instance with two monotone valuations and one additive valuation in which no PMMS allocation exists. Since EFX allocations are known to exist under these assumptions, this establishes a formal separation between EFX and PMMS. We prove existence of fair allocations for three important special cases. We show that EFX allocations exist for personalized bivalued valuations, where for each agent $i$ there exist values $a_i > b_i$ such that agent $i$ assigns value $v_i(\{g\}) \in \{a_i, b_i\}$ to each good $g$. We establish an analogous existence result for PMMS allocations when $a_i$ is divisible by $b_i$. We also prove that PMMS allocations exist for binary-valued MMS-feasible valuations, where each bundle $S$ has value $v_i(S) \in \{0, 1\}$. Notably, this result holds even without assuming monotonicity of valuations and thus applies to the fair division of chores and mixed manna. Finally, we study a class of valuations called pair-demand valuations, which extend the well-studied unit-demand valuations to the case where each agent derives value from at most two items, and we show that PMMS allocations exist in this setting. Our proofs are constructive, and we provide polynomial-time algorithms for all three existence results.
Authors: Yiwen Chen, Zhihao Li, Yikai Wang, Hu Zhang, Qin Li, Chi Zhang, Guosheng Lin
Abstract: Recent advances in sparse voxel representations have significantly improved the quality of 3D content generation, enabling high-resolution modeling with fine-grained geometry. However, existing frameworks suffer from severe computational inefficiencies due to the quadratic complexity of attention mechanisms in their two-stage diffusion pipelines. In this work, we propose Ultra3D, an efficient 3D generation framework that significantly accelerates sparse voxel modeling without compromising quality. Our method leverages the compact VecSet representation to efficiently generate a coarse object layout in the first stage, reducing token count and accelerating voxel coordinate prediction. To refine per-voxel latent features in the second stage, we introduce Part Attention, a geometry-aware localized attention mechanism that restricts attention computation within semantically consistent part regions. This design preserves structural continuity while avoiding unnecessary global attention, achieving up to 6.7x speed-up in latent generation. To support this mechanism, we construct a scalable part annotation pipeline that converts raw meshes into part-labeled sparse voxels. Extensive experiments demonstrate that Ultra3D supports high-resolution 3D generation at 1024 resolution and achieves state-of-the-art performance in both visual fidelity and user preference.
Authors: Pascal Spiegler, Taha Koleilat, Arash Harirpoush, Corey S. Miller, Hassan Rivaz, Marta Kersten-Oertel, Yiming Xiao
Abstract: Pancreatic cancer carries a poor prognosis and relies on endoscopic ultrasound (EUS) for targeted biopsy and radiotherapy. However, the speckle noise, low contrast, and unintuitive appearance of EUS make segmentation of pancreatic tumors with fully supervised deep learning (DL) models both error-prone and dependent on large, expert-curated annotation datasets. To address these challenges, we present TextSAM-EUS, a novel, lightweight, text-driven adaptation of the Segment Anything Model (SAM) that requires no manual geometric prompts at inference. Our approach leverages text prompt learning (context optimization) through the BiomedCLIP text encoder in conjunction with a LoRA-based adaptation of SAM's architecture to enable automatic pancreatic tumor segmentation in EUS, tuning only 0.86% of the total parameters. On the public Endoscopic Ultrasound Database of the Pancreas, TextSAM-EUS with automatic prompts attains 82.69% Dice and 85.28% normalized surface distance (NSD), and with manual geometric prompts reaches 83.10% Dice and 85.70% NSD, outperforming both existing state-of-the-art (SOTA) supervised DL models and foundation models (e.g., SAM and its variants). As the first attempt to incorporate prompt learning in SAM-based medical image segmentation, TextSAM-EUS offers a practical option for efficient and robust automatic EUS segmentation. Code is available at https://github.com/HealthX-Lab/TextSAM-EUS .
Authors: Xiaokai Bai, Chenxu Zhou, Lianqing Zheng, Si-Yuan Cao, Jianan Liu, Xiaohan Zhang, Zhengzhuang Zhang, Hui-liang Shen
Abstract: 4D millimeter-wave radar has emerged as a promising sensor for autonomous driving, but effective 3D object detection from both 4D radar and monocular images remains a challenge. Existing fusion approaches typically rely on either instance-based proposals or dense BEV grids, which either lack holistic scene understanding or are limited by rigid grid structures. To address these, we propose RaGS, the first framework to leverage 3D Gaussian Splatting (GS) as representation for fusing 4D radar and monocular cues in 3D object detection. 3D GS naturally suits 3D object detection by modeling the scene as a field of Gaussians, dynamically allocating resources on foreground objects and providing a flexible, resource-efficient solution. RaGS uses a cascaded pipeline to construct and refine the Gaussian field. It starts with the Frustum-based Localization Initiation (FLI), which unprojects foreground pixels to initialize coarse 3D Gaussians positions. Then, the Iterative Multimodal Aggregation (IMA) fuses semantics and geometry, refining the limited Gaussians to the regions of interest. Finally, the Multi-level Gaussian Fusion (MGF) renders the Gaussians into multi-level BEV features for 3D object detection. By dynamically focusing on sparse objects within scenes, RaGS enable object concentrating while offering comprehensive scene perception. Extensive experiments on View-of-Delft, TJ4DRadSet, and OmniHD-Scenes benchmarks demonstrate its state-of-the-art performance. Code will be released.
Authors: Drandreb Earl O. Juanico, Rowel O. Atienza, Jeffrey Kenneth Go
Abstract: We propose Reverse Contrast Attention (RCA), a plug-in method that enhances object localization in vision-language transformers without retraining. RCA reweights final-layer attention by suppressing extremes and amplifying mid-level activations to let semantically relevant but subdued tokens guide predictions. We evaluate it on Open Vocabulary Referring Object Detection (OV-RefOD), introducing FitAP, a confidence-free average precision metric based on IoU and box area. RCA improves FitAP in 11 out of 15 open-source VLMs, with gains up to $+26.6\%$. Effectiveness aligns with attention sharpness and fusion timing; while late-fusion models benefit consistently, models like $\texttt{DeepSeek-VL2}$ also improve, pointing to capacity and disentanglement as key factors. RCA offers both interpretability and performance gains for multimodal transformers. Codes and dataset are available from https://github.com/earl-juanico/rca
Authors: Joydeep Chandra, Satyam Kumar Navneet
Abstract: As AI-driven dataspaces become integral to data sharing and collaborative analytics, ensuring privacy, performance, and policy compliance presents significant challenges. This paper provides a comprehensive review of privacy-preserving and policy-aware AI techniques, including Federated Learning, Differential Privacy, Trusted Execution Environments, Homomorphic Encryption, and Secure Multi-Party Computation, alongside strategies for aligning AI with regulatory frameworks such as GDPR and the EU AI Act. We propose a novel taxonomy to classify these techniques based on privacy levels, performance impacts, and compliance complexity, offering a clear framework for practitioners and researchers to navigate trade-offs. Key performance metrics -- latency, throughput, cost overhead, model utility, fairness, and explainability -- are analyzed to highlight the multi-dimensional optimization required in dataspaces. The paper identifies critical research gaps, including the lack of standardized privacy-performance KPIs, challenges in explainable AI for federated ecosystems, and semantic policy enforcement amidst regulatory fragmentation. Future directions are outlined, proposing a conceptual framework for policy-driven alignment, automated compliance validation, standardized benchmarking, and integration with European initiatives like GAIA-X, IDS, and Eclipse EDC. By synthesizing technical, ethical, and regulatory perspectives, this work lays the groundwork for developing trustworthy, efficient, and compliant AI systems in dataspaces, fostering innovation in secure and responsible data-driven ecosystems.
Authors: Likun Tan, Kuan-Wei Huang, Kevin Wu
Abstract: Hallucinations in large language models pose a critical challenge for applications requiring factual reliability, particularly in high-stakes domains such as finance. This work presents an effective approach for detecting and editing factually incorrect content in model-generated responses based on the provided context. Given a user-defined domain-specific error taxonomy, we construct a synthetic dataset by inserting tagged errors into financial question-answering corpora and then fine-tune four language models, Phi-4, Phi-4-mini, Qwen3-4B, and Qwen3-14B, to detect and edit these factual inaccuracies. Our best-performing model, fine-tuned Phi-4, achieves an 8% improvement in binary F1 score and a 30% gain in overall detection performance compared to OpenAI-o3. Notably, our fine-tuned Phi-4-mini model, despite having only 4 billion parameters, maintains competitive performance with just a 2% drop in binary detection and a 0.1% decline in overall detection compared to OpenAI-o3. Our work provides a practical solution for detecting and editing factual inconsistencies in financial text generation while introducing a generalizable framework that can enhance the trustworthiness and alignment of large language models across diverse applications beyond finance. Our code and data are available at https://github.com/pegasi-ai/shield.
Authors: Yixin Song, Zhenliang Xue, Dongliang Wei, Feiyang Chen, Jianxiang Gao, Junchen Liu, Hangyu Liang, Guangshuo Qin, Chengrong Tian, Bo Wen, Longyu Zhao, Xinrui Zheng, Zeyu Mi, Haibo Chen
Abstract: While frontier large language models (LLMs) continue to push capability boundaries, their deployment remains confined to GPU-powered cloud infrastructure. We challenge this paradigm with SmallThinker, a family of LLMs natively designed - not adapted - for the unique constraints of local devices: weak computational power, limited memory, and slow storage. Unlike traditional approaches that mainly compress existing models built for clouds, we architect SmallThinker from the ground up to thrive within these limitations. Our innovation lies in a deployment-aware architecture that transforms constraints into design principles. First, We introduce a two-level sparse structure combining fine-grained Mixture-of-Experts (MoE) with sparse feed-forward networks, drastically reducing computational demands without sacrificing model capacity. Second, to conquer the I/O bottleneck of slow storage, we design a pre-attention router that enables our co-designed inference engine to prefetch expert parameters from storage while computing attention, effectively hiding storage latency that would otherwise cripple on-device inference. Third, for memory efficiency, we utilize NoPE-RoPE hybrid sparse attention mechanism to slash KV cache requirements. We release SmallThinker-4B-A0.6B and SmallThinker-21B-A3B, which achieve state-of-the-art performance scores and even outperform larger LLMs. Remarkably, our co-designed system mostly eliminates the need for expensive GPU hardware: with Q4_0 quantization, both models exceed 20 tokens/s on ordinary consumer CPUs, while consuming only 1GB and 8GB of memory respectively. SmallThinker is publicly available at hf.co/PowerInfer/SmallThinker-4BA0.6B-Instruct and hf.co/PowerInfer/SmallThinker-21BA3B-Instruct.
Authors: Donglu Yang, Liang Zhang, Zihao Yue, Liangyu Chen, Yichen Xu, Wenxuan Wang, Qin Jin
Abstract: Charts are a fundamental visualization format widely used in data analysis across research and industry. While enabling users to edit charts based on high-level intentions is of great practical value, existing methods primarily rely on natural language instructions, which are often too ambiguous to support fine-grained editing. In this work, we introduce a novel paradigm for multimodal chart editing, where user intent is expressed through a combination of natural language and visual indicators that explicitly highlight the elements to be modified. To support this paradigm, we present Chart$\text{M}^3$, a new benchmark for Multimodal chart editing with Multi-level complexity and Multi-perspective evaluation. Chart$\text{M}^3$ contains 1,000 samples spanning four levels of editing difficulty. Each sample includes triplets in the form of (chart, code, multimodal instructions). To comprehensively evaluate chart editing models, Chart$\text{M}^3$ provides metrics that assess both visual appearance and code correctness. Our benchmark reveals significant limitations in current multimodal large language models (MLLMs), including GPT-4o, particularly in their ability to interpret and act on visual indicators. To address this, we construct Chart$\text{M}^3$-Train, a large-scale training set with 24,000 multimodal chart editing samples. Fine-tuning MLLMs on this dataset leads to substantial improvements, demonstrating the importance of multimodal supervision in building practical chart editing systems. Our datasets, codes, and evaluation tools are available at https://github.com/MLrollIT/ChartM3. %https://github.com/MLrollIT/ChartM3Our datasets, codes, and evaluation tools are available at https://github.com/yaolinli/VCE.
URLs: https://github.com/MLrollIT/ChartM3., https://github.com/MLrollIT/ChartM3Our, https://github.com/yaolinli/VCE.
Authors: Sinh Trong Vu, Huong Thu Truong, Oanh Tien Do, Tu Anh Le, Tai Tan Mai
Abstract: Large language models have been widely applied in many aspects of real life, bringing significant efficiency to businesses and offering distinctive user experiences. In this paper, we focus on exploring the application of ChatGPT, a chatbot based on a large language model, to support higher educator in generating quiz questions and assessing learners. Specifically, we explore interactive prompting patterns to design an optimal AI-powered question bank creation process. The generated questions are evaluated through a "Blind test" survey sent to various stakeholders including lecturers and learners. Initial results at the Banking Academy of Vietnam are relatively promising, suggesting a potential direction to streamline the time and effort involved in assessing learners at higher education institutes.
Authors: Shijie Zhou, Ruiyi Zhang, Huaisheng Zhu, Branislav Kveton, Yufan Zhou, Jiuxiang Gu, Jian Chen, Changyou Chen
Abstract: We introduce LLaVA-Reward, an efficient reward model designed to automatically evaluate text-to-image (T2I) generations across multiple perspectives, leveraging pretrained multimodal large language models (MLLMs). Existing MLLM-based approaches require instruction-following data for supervised fine-tuning and evaluate generation quality on analyzing text response, which is time-consuming and difficult to train. To address this problem, we propose LLaVA-Reward, which directly utilizes the hidden states of MLLMs given text-image pairs. To enhance the bidirectional interaction between visual and textual representations in decoder-only MLLMs, we further propose adding a Skip-connection Cross Attention (SkipCA) module. This design enhances text-image correlation reasoning by connecting early-layer visual features with later-layer hidden representations. In addition, LLaVA-Reward supports different types of preference data for efficient fine-tuning, including paired preference data and unpaired data. We train LLaVA-Reward on four evaluation perspectives: text-image alignment, fidelity/artifact, safety, and overall ranking. Empirical results demonstrate that LLaVA-Reward outperforms conventional and MLLM-based methods in generating human-aligned scores for automatic evaluations and inference-time scaling in text-to-image generations.
Authors: Chang-Hun Ji, SiWoon Song, Youn-Hee Han, SungTae Moon
Abstract: A drone trajectory planner should be able to dynamically adjust the safety-efficiency trade-off according to varying mission requirements in unknown environments. Although traditional polynomial-based planners offer computational efficiency and smooth trajectory generation, they require expert knowledge to tune multiple parameters to adjust this trade-off. Moreover, even with careful tuning, the resulting adjustment may fail to achieve the desired trade-off. Similarly, although reinforcement learning-based planners are adaptable in unknown environments, they do not explicitly address the safety-efficiency trade-off. To overcome this limitation, we introduce a Decision Transformer-based trajectory planner that leverages a single parameter, Return-to-Go (RTG), as a \emph{temperature parameter} to dynamically adjust the safety-efficiency trade-off. In our framework, since RTG intuitively measures the safety and efficiency of a trajectory, RTG tuning does not require expert knowledge. We validate our approach using Gazebo simulations in both structured grid and unstructured random environments. The experimental results demonstrate that our planner can dynamically adjust the safety-efficiency trade-off by simply tuning the RTG parameter. Furthermore, our planner outperforms existing baseline methods across various RTG settings, generating safer trajectories when tuned for safety and more efficient trajectories when tuned for efficiency. Real-world experiments further confirm the reliability and practicality of our proposed planner.
Authors: Lujain Ibrahim, Franziska Sofia Hafner, Luc Rocher
Abstract: Artificial intelligence (AI) developers are increasingly building language models with warm and empathetic personas that millions of people now use for advice, therapy, and companionship. Here, we show how this creates a significant trade-off: optimizing language models for warmth undermines their reliability, especially when users express vulnerability. We conducted controlled experiments on five language models of varying sizes and architectures, training them to produce warmer, more empathetic responses, then evaluating them on safety-critical tasks. Warm models showed substantially higher error rates (+10 to +30 percentage points) than their original counterparts, promoting conspiracy theories, providing incorrect factual information, and offering problematic medical advice. They were also significantly more likely to validate incorrect user beliefs, particularly when user messages expressed sadness. Importantly, these effects were consistent across different model architectures, and occurred despite preserved performance on standard benchmarks, revealing systematic risks that current evaluation practices may fail to detect. As human-like AI systems are deployed at an unprecedented scale, our findings indicate a need to rethink how we develop and oversee these systems that are reshaping human relationships and social interaction.
Authors: Zihan Zhao, Bo Chen, Ziping Wan, Lu Chen, Xuanze Lin, Shiyang Yu, Situo Zhang, Da Ma, Zichen Zhu, Danyang Zhang, Huayang Wang, Zhongyang Dai, Liyang Wen, Xin Chen, Kai Yu
Abstract: While large language models (LLMs) have achieved impressive progress, their application in scientific domains such as chemistry remains hindered by shallow domain understanding and limited reasoning capabilities. In this work, we focus on the specific field of chemistry and develop a Chemical Reasoner LLM, ChemDFM-R. We first construct a comprehensive dataset of atomized knowledge points to enhance the model's understanding of the fundamental principles and logical structure of chemistry. Then, we propose a mix-sourced distillation strategy that integrates expert-curated knowledge with general-domain reasoning skills, followed by domain-specific reinforcement learning to enhance chemical reasoning. Experiments on diverse chemical benchmarks demonstrate that ChemDFM-R achieves cutting-edge performance while providing interpretable, rationale-driven outputs. Further case studies illustrate how explicit reasoning chains significantly improve the reliability, transparency, and practical utility of the model in real-world human-AI collaboration scenarios.