Authors: Richard Ackermann, Simeon Emanuilov
Abstract: OpenAI has recently argued that hallucinations in large language models result primarily from misaligned evaluation incentives that reward confident guessing rather than epistemic humility. On this view, hallucination is a contingent behavioral artifact, remediable through improved benchmarks and reward structures. In this paper, we challenge that interpretation. Drawing on previous work on structural hallucination and empirical experiments using a Licensing Oracle, we argue that hallucination is not an optimization failure but an architectural inevitability of the transformer model. Transformers do not represent the world; they model statistical associations among tokens. Their embedding spaces form a pseudo-ontology derived from linguistic co-occurrence rather than world-referential structure. At ontological boundary conditions - regions where training data is sparse or incoherent - the model necessarily interpolates fictional continuations in order to preserve coherence. No incentive mechanism can modify this structural dependence on pattern completion. Our empirical results demonstrate that hallucination can only be eliminated through external truth-validation and abstention modules, not through changes to incentives, prompting, or fine-tuning. The Licensing Oracle achieves perfect abstention precision across domains precisely because it supplies grounding that the transformer lacks. We conclude that hallucination is a structural property of generative architectures and that reliable AI requires hybrid systems that distinguish linguistic fluency from epistemic responsibility.
Authors: Biao Zhang, Paul Suganthan, Ga\"el Liu, Ilya Philippov, Sahil Dua, Ben Hora, Kat Black, Gus Martins, Omar Sanseviero, Shreya Pathak, Cassidy Hardin, Francesco Visin, Jiageng Zhang, Kathleen Kenealy, Qin Yin, Olivier Lacombe, Armand Joulin, Tris Warkentin, Adam Roberts
Abstract: We introduce T5Gemma 2, the next generation of the T5Gemma family of lightweight open encoder-decoder models, featuring strong multilingual, multimodal and long-context capabilities. T5Gemma 2 follows the adaptation recipe (via UL2) in T5Gemma -- adapting a pretrained decoder-only model into an encoder-decoder model, and extends it from text-only regime to multimodal based on the Gemma 3 models. We further propose two methods to improve the efficiency: tied word embedding that shares all embeddings across encoder and decoder, and merged attention that unifies decoder self- and cross-attention into a single joint module. Experiments demonstrate the generality of the adaptation strategy over architectures and modalities as well as the unique strength of the encoder-decoder architecture on long context modeling. Similar to T5Gemma, T5Gemma 2 yields comparable or better pretraining performance and significantly improved post-training performance than its Gemma 3 counterpart. We release the pretrained models (270M-270M, 1B-1B and 4B-4B) to the community for future research.
Authors: Massimiliano Fadda, Enrico Motta, Francesco Osborne, Diego Reforgiato Recupero, Angelo Salatino
Abstract: News sources play a central role in democratic societies by shaping political and social discourse through specific topics, viewpoints and voices. Understanding these dynamics is essential for assessing whether the media landscape offers a balanced and fair account of public debate. In earlier work, we introduced a pipeline that, given a news corpus, i) uses a hybrid human-machine approach to identify the range of viewpoints expressed about a given topic, and ii) classifies relevant claims with respect to the identified viewpoints, defined as sets of semantically and ideologically congruent claims (e.g., positions arguing that immigration positively impacts the UK economy). In this paper, we improve this pipeline by i) fine-tuning Large Language Models (LLMs) for viewpoint classification and ii) enriching claim representations with semantic descriptions of relevant actors drawn from Wikidata. We evaluate our approach against alternative solutions on a benchmark centred on the UK immigration debate. Results show that while both mechanisms independently improve classification performance, their integration yields the best results, particularly when using LLMs capable of processing long inputs.
Authors: Houman Kazemzadeh, Kiarash Mokhtari Dizaji, Seyed Reza Tavakoli, Farbod Davoodi, MohammadReza KarimiNejad, Parham Abed Azad, Ali Sabzi, Armin Khosravi, Siavash Ahmadi, Mohammad Hossein Rohban, Glolamali Aminian, Tahereh Javaheri
Abstract: Objectives: To evaluate large language model (LLM) performance on pharmacy licensure-style question-answering (QA) tasks and develop an external knowledge integration method to improve their accuracy. Methods: We benchmarked eleven existing LLMs with varying parameter sizes (8 billion to 70+ billion) using a 141-question pharmacy dataset. We measured baseline accuracy for each model without modification. We then developed a three-step retrieval-augmented generation (RAG) pipeline, DrugRAG, that retrieves structured drug knowledge from validated sources and augments model prompts with evidence-based context. This pipeline operates externally to the models, requiring no changes to model architecture or parameters. Results: Baseline accuracy ranged from 46% to 92%, with GPT-5 (92%) and o3 (89%) achieving the highest scores. Models with fewer than 8 billion parameters scored below 50%. DrugRAG improved accuracy across all tested models, with gains ranging from 7 to 21 percentage points (e.g., Gemma 3 27B: 61% to 71%, Llama 3.1 8B: 46% to 67%) on the 141-item benchmark. Conclusion: We demonstrate that external structured drug knowledge integration through DrugRAG measurably improves LLM accuracy on pharmacy tasks without modifying the underlying models. This approach provides a practical pipeline for enhancing pharmacy-focused AI applications with evidence-based information.
Authors: Caner Erden
Abstract: The quadratic computational complexity of MultiHead SelfAttention (MHSA) remains a fundamental bottleneck in scaling Large Language Models (LLMs) for longcontext tasks. While sparse and linearized attention mechanisms attempt to mitigate this, they often compromise the representation of global dependencies or fail to capture multiscale semantic granularity effectively. In this paper, we propose Multiscale Aggregated Hierarchical Attention (MAHA), a novel architectural framework that reformulates the attention mechanism through hierarchical decomposition and mathematically rigorous aggregation. Unlike conventional approaches that treat token interactions at a single resolution, MAHA dynamically partitions the input sequence into hierarchical scales via learnable downsampling operators. The core innovation lies in its aggregation strategy: we model the fusion of scalespecific attention matrices as a resource allocation problem, solved via a convex optimization framework or a Nash equilibriumbased gametheoretic approach. This ensures a theoretically optimal balance between local nuance and global context fidelity. Implemented within a hybrid dilatedconvolutional transformer backbone, MAHA utilizes differentiable optimization layers to enable endtoend training. Experimental evaluations demonstrate that MAHA achieves superior scalability; empirical FLOPs analysis confirms an 81% reduction in computational cost at a sequence length of 4096 compared to standard attention. This work bridges the gap between optimization theory and sequence modeling, offering a scalable solution for nextgeneration LLMs.
Authors: George-Andrei Dima, Dumitru-Clementin Cercel
Abstract: Focusing on low-resource languages is an essential step toward democratizing generative AI. In this work, we contribute to reducing the multimodal NLP resource gap for Romanian. We translate the widely known Flickr30k dataset into Romanian and further extend it for visual question answering by leveraging open-source LLMs. We demonstrate the usefulness of our datasets by fine-tuning open-source VLMs on Romanian visual question answering. We select VLMs from three widely used model families: LLaMA 3.2, LLaVA 1.6, and Qwen2. For fine-tuning, we employ the parameter-efficient LoRA method. Our models show improved Romanian capabilities in visual QA, as well as on tasks they were not trained on, such as Romanian image description generation. The seven-billion-parameter Qwen2-VL-RoVQA obtains top scores on both tasks, with improvements of +6.05% and +2.61% in BERTScore F1 over its original version. Finally, the models show substantial reductions in grammatical errors compared to their original forms, indicating improvements not only in language understanding but also in Romanian fluency.
Authors: Buu Phan, Ashish Khisti, Karen Ullrich
Abstract: Computing next-token likelihood ratios between two language models (LMs) is a standard task in training paradigms such as knowledge distillation. Since this requires both models to share the same probability space, it becomes challenging when the teacher and student LMs use different tokenizers, for instance, when edge-device deployment necessitates a smaller vocabulary size to lower memory overhead. In this work, we address this vocabulary misalignment problem by uncovering an implicit recursive structure in the commonly deployed Byte-Pair Encoding (BPE) algorithm and utilizing it to create a probabilistic framework for cross-tokenizer likelihood scoring. Our method enables sequence likelihood evaluation for vocabularies different from the teacher model native tokenizer, addressing two specific scenarios: when the student vocabulary is a subset of the teacher vocabulary, and the general case where it is arbitrary. In the subset regime, our framework computes exact likelihoods and provides next-token probabilities for sequential sampling with only O(1) model evaluations per token. When used for distillation, this yields up to a 12% reduction in memory footprint for the Qwen2.5-1.5B model while also improving baseline performance up to 4% on the evaluated tasks. For the general case, we introduce a rigorous lossless procedure that leverages BPE recursive structure, complemented by a fast approximation that keeps large-vocabulary settings practical. Applied to distillation for mathematical reasoning, our approach improves GSM8K accuracy by more than 2% over the current state of the art.
Authors: Yiming Cui, Xin Yao, Yuxuan Qin, Xin Li, Shijin Wang, Guoping Hu
Abstract: Multimodal scientific reasoning remains a significant challenge for large language models (LLMs), particularly in chemistry, where problem-solving relies on symbolic diagrams, molecular structures, and structured visual data. Here, we systematically evaluate 40 proprietary and open-source multimodal LLMs, including GPT-5, o3, Gemini-2.5-Pro, and Qwen2.5-VL, on a curated benchmark of Olympiad-style chemistry questions drawn from over two decades of U.S. National Chemistry Olympiad (USNCO) exams. These questions require integrated visual and textual reasoning across diverse modalities. We find that many models struggle with modality fusion, where in some cases, removing the image even improves accuracy, indicating misalignment in vision-language integration. Chain-of-Thought prompting consistently enhances both accuracy and visual grounding, as demonstrated through ablation studies and occlusion-based interpretability. Our results reveal critical limitations in the scientific reasoning abilities of current MLLMs, providing actionable strategies for developing more robust and interpretable multimodal systems in chemistry. This work provides a timely benchmark for measuring progress in domain-specific multimodal AI and underscores the need for further advances at the intersection of artificial intelligence and scientific reasoning.
Authors: Sijin Sun, Liangbin Zhao, Ming Deng, Xiuju Fu
Abstract: Dialogue Topic Segmentation (DTS) is crucial for understanding task-oriented public-channel communications, such as maritime VHF dialogues, which feature informal speech and implicit transitions. To address the limitations of traditional methods, we propose DASH-DTS, a novel LLM-based framework. Its core contributions are: (1) topic shift detection via dialogue handshake recognition; (2) contextual enhancement through similarity-guided example selection; and (3) the generation of selective positive and negative samples to improve model discrimination and robustness. Additionally, we release VHF-Dial, the first public dataset of real-world maritime VHF communications, to advance research in this domain. DASH-DTS provides interpretable reasoning and confidence scores for each segment. Experimental results demonstrate that our framework achieves several sota segmentation trusted accuracy on both VHF-Dial and standard benchmarks, establishing a strong foundation for stable monitoring and decision support in operational dialogues.
Authors: Hongbo Wang, MaungMaung AprilPyone, Isao Echizen
Abstract: Disclaimer: Samples in this paper may be harmful and cause discomfort. Multimodal large language models (MLLMs) enable multimodal generation but inherit toxic, biased, and NSFW signals from weakly curated pretraining corpora, causing safety risks, especially under adversarial triggers that late, opaque training-free detoxification methods struggle to handle. We propose SGM, a white-box neuron-level multimodal intervention that acts like safety glasses for toxic neurons: it selectively recalibrates a small set of toxic expert neurons via expertise-weighted soft suppression, neutralizing harmful cross-modal activations without any parameter updates. We establish MM-TOXIC-QA, a multimodal toxicity evaluation framework, and compare SGM with existing detoxification techniques. Experiments on open-source MLLMs show that SGM mitigates toxicity in standard and adversarial conditions, cutting harmful rates from 48.2\% to 2.5\% while preserving fluency and multimodal reasoning. SGM is extensible, and its combined defenses, denoted as SGM*, integrate with existing detoxification methods for stronger safety performance, providing an interpretable, low-cost solution for toxicity-controlled multimodal generation.
Authors: Fanzhe Fu
Abstract: The transition of Large Language Models (LLMs) from stochastic chat interfaces to reliable software components necessitates a fundamental re-engineering of interaction paradigms. Current methodologies, predominantly heuristic-based "prompt engineering," fail to provide the deterministic guarantees required for mission-critical applications. We introduce the Meta-Prompting Protocol, a rigorous theoretical framework that formalizes the orchestration of LLMs as a programmable, self-optimizing system. Central to this protocol is the Adversarial Trinity, a tripartite topology comprising a Generator (P), an Auditor (A), and an Optimizer (O). By treating natural language instructions as differentiable variables within a semantic computation graph and utilizing textual critiques as gradients, this architecture mitigates hallucination and prevents model collapse. We demonstrate the theoretical viability of this approach using declarative programming paradigms (DSPy) and automatic textual differentiation (TextGrad), establishing a foundation for "Observable Software Engineering" in the era of probabilistic computing.
Authors: Weiqin Wang, Yile Wang, Kehao Chen, Hui Huang
Abstract: Test-time reinforcement learning mitigates the reliance on annotated data by using majority voting results as pseudo-labels, emerging as a complementary direction to reinforcement learning with verifiable rewards (RLVR) for improving reasoning ability of large language models (LLMs). However, this voting strategy often induces confirmation bias and suffers from sparse rewards, limiting the overall performance. In this work, we propose subgroup-specific step-wise confidence-weighted pseudo-label estimation (SCOPE), a framework integrating model confidence and dynamic subgroup partitioning to address these issues. Specifically, SCOPE integrates the proposed step-wise confidence into pseudo label deduction, prioritizing high-quality reasoning paths over simple frequency count. Furthermore, it dynamically partitions the candidate outputs pool into independent subgroups by balancing reasoning quality against exploration diversity. By deriving local consensus via repeat sampling for each sub group, SCOPE provides diverse supervision targets to encourage broader exploration. We conduct experiments across various models and benchmarks, experimental results show that SCOPE consistently outperforms recent baselines. Notably, SCOPE achieving relative improvements of 13.1\% on challenging AIME 2025 and 8.1\% on AMC. The code is released at \href{https://github.com/szu-tera/SCOPE}{https://github.com/szu-tera/SCOPE}.
URLs: https://github.com/szu-tera/SCOPE, https://github.com/szu-tera/SCOPE
Authors: Yuki Nakayama, Koki Hikichi, Yun Ching Liu, Yu Hirate
Abstract: This paper presents a large-scale corpus of Rakuten Travel Reviews. Our collection contains 7.3 million customer reviews for 16 years, ranging from 2009 to 2024. Each record in the dataset contains the review text, its response from an accommodation, an anonymized reviewer ID, review date, accommodation ID, plan ID, plan title, room type, room name, purpose, accompanying group, and user ratings from different aspect categories, as well as an overall score. We present statistical information about our corpus and provide insights into factors driving data drift between 2019 and 2024 using statistical approaches.
Authors: Xuanjun Zong, Zhiqi Shen, Lei Wang, Yunshi Lan, Chao Yang
Abstract: Large language models (LLMs) are evolving into agentic systems that reason, plan, and operate external tools. The Model Context Protocol (MCP) is a key enabler of this transition, offering a standardized interface for connecting LLMs with heterogeneous tools and services. Yet MCP's openness and multi-server workflows introduce new safety risks that existing benchmarks fail to capture, as they focus on isolated attacks or lack real-world coverage. We present MCP-SafetyBench, a comprehensive benchmark built on real MCP servers that supports realistic multi-turn evaluation across five domains: browser automation, financial analysis, location navigation, repository management, and web search. It incorporates a unified taxonomy of 20 MCP attack types spanning server, host, and user sides, and includes tasks requiring multi-step reasoning and cross-server coordination under uncertainty. Using MCP-SafetyBench, we systematically evaluate leading open- and closed-source LLMs, revealing large disparities in safety performance and escalating vulnerabilities as task horizons and server interactions grow. Our results highlight the urgent need for stronger defenses and establish MCP-SafetyBench as a foundation for diagnosing and mitigating safety risks in real-world MCP deployments.
Authors: Mika H\"am\"al\"ainen, Kimmo Leivisk\"a
Abstract: This paper explores the growing epistemic parallel between NLG evaluation and grading of students in a Finnish University. We argue that both domains are experiencing a Great Misalignment Problem. As students increasingly use tools like ChatGPT to produce sophisticated outputs, traditional assessment methods that focus on final products rather than learning processes have lost their validity. To address this, we introduce the Pedagogical Multi-Factor Assessment (P-MFA) model, a process-based, multi-evidence framework inspired by the logic of multi-factor authentication.
Authors: Chao Zhang, Minghan Li, Tianrui Lv, Guodong Zhou
Abstract: Large language models (LLMs) often generate hallucinations in knowledge-intensive QA due to parametric knowledge limitations. While existing methods like KG-CoT improve reliability by integrating knowledge graph (KG) paths, they suffer from rigid hop-count selection (solely question-driven) and underutilization of reasoning paths (lack of guidance). To address this, we propose RFKG-CoT: First, it replaces the rigid hop-count selector with a relation-driven adaptive hop-count selector that dynamically adjusts reasoning steps by activating KG relations (e.g., 1-hop for direct "brother" relations, 2-hop for indirect "father-son" chains), formalized via a relation mask. Second, it introduces a few-shot in-context learning path guidance mechanism with CoT (think) that constructs examples in a "question-paths-answer" format to enhance LLMs' ability to understand reasoning paths. Experiments on four KGQA benchmarks show RFKG-CoT improves accuracy by up to 14.7 pp (Llama2-7B on WebQSP) over KG-CoT. Ablations confirm the hop-count selector and the path prompt are complementary, jointly transforming KG evidence into more faithful answers.
Authors: Yash Bhaskar, Parameswari Krishnamurthy
Abstract: This paper presents the systems submitted by the Yes-MT team for the Low-Resource Indic Language Translation Shared Task at WMT 2024 (Pakray et al., 2024), focusing on translating between English and the Assamese, Mizo, Khasi, and Manipuri languages. The experiments explored various approaches, including fine-tuning pre-trained models like mT5 (Xue et al., 2020) and IndicBart (Dabre et al., 2021) in both multilingual and monolingual settings, LoRA (Hu et al., 2021) fine-tuning IndicTrans2 (Gala et al., 2023), zero-shot and few-shot prompting (Brown, 2020) with large language models (LLMs) like Llama 3 (Dubey et al., 2024) and Mixtral 8x7b (Jiang et al., 2024), LoRA supervised fine-tuning of Llama 3 (Mecklenburg et al., 2024), and training Transformer models (Vaswani, 2017) from scratch. The results were evaluated on the WMT23 Low-Resource Indic Language Translation Shared Task test data using SacreBLEU (Post, 2018) and CHRF (Popovic, 2015), highlighting the challenges of low-resource translation and the potential of LLMs for these tasks, particularly with fine-tuning.
Authors: Claudio Savelli, Moreno La Quatra, Alkis Koudounas, Flavio Giobergia
Abstract: LLMs trained on web-scale data raise concerns about privacy and the right to be forgotten. To address these issues, Machine Unlearning provides techniques to remove specific information from trained models without retraining from scratch. However, existing benchmarks for evaluating unlearning in LLMs face two major limitations: they focus only on English and support only entity-level forgetting (removing all information about a person). We introduce FAME (Fictional Actors for Multilingual Erasure), a synthetic benchmark for evaluating Machine Unlearning across five languages: English, French, German, Italian, and Spanish. FAME contains 1,000 fictional actor biographies and 20,000 question-answer pairs. Each biography includes information on 20 topics organized into structured categories (biography, career, achievements, personal information). This design enables both entity-level unlearning (i.e., forgetting entire identities) and instance-level unlearning (i.e., forgetting specific facts while retaining others). We provide two dataset splits to support these two different unlearning scenarios and enable systematic comparison of unlearning techniques across languages. Since FAME uses entirely fictional data, it ensures that the information was never encountered during model pretraining, allowing for a controlled evaluation of unlearning methods.
Authors: Maria Becker, Mirko Sommer, Lars Tapken, Yi Wan Teh, Bruno Brocai
Abstract: Moralizations - arguments that invoke moral values to justify demands or positions - are a yet underexplored form of persuasive communication. We present the Moralization Corpus, a novel multi-genre dataset designed to analyze how moral values are strategically used in argumentative discourse. Moralizations are pragmatically complex and often implicit, posing significant challenges for both human annotators and NLP systems. We develop a frame-based annotation scheme that captures the constitutive elements of moralizations - moral values, demands, and discourse protagonists - and apply it to a diverse set of German texts, including political debates, news articles, and online discussions. The corpus enables fine-grained analysis of moralizing language across communicative formats and domains. We further evaluate several large language models (LLMs) under varied prompting conditions for the task of moralization detection and moralization component extraction and compare it to human annotations in order to investigate the challenges of automatic and manual analysis of moralizations. Results show that detailed prompt instructions has a greater effect than few-shot or explanation-based prompting, and that moralization remains a highly subjective and context-sensitive task. We release all data, annotation guidelines, and code to foster future interdisciplinary research on moral discourse and moral reasoning in NLP.
Authors: Piyawoot Songsiritat
Abstract: We introduce SynGP500, a clinician-curated collection of 500 synthetic Australian general practice medical notes. The dataset integrates curriculum-based clinical breadth (RACGP 2022 Curriculum), epidemiologically-calibrated prevalence (BEACH study), and diverse consultation contexts. This approach systematically includes both common presentations and less-common curriculum-specified conditions that GPs must recognize but appear infrequently in single practice populations, potentially supporting more generalizable model training than datasets constrained by naturally occurring case distributions. SynGP500 is messy by design, reflecting the authentic complexity of healthcare delivery: telegraphic documentation, typos, patient non-adherence, socioeconomic barriers, and clinician-patient disagreements, unlike sanitized synthetic datasets that obscure clinical realities. Multi-faceted validation demonstrates dataset quality through epidemiological alignment with real Australian GP consultation patterns (BEACH study), stylometric analysis confirming high linguistic variation, semantic diversity analysis demonstrating broad coverage, and exploratory downstream evaluation using self-supervised medical concept extraction, showing F1 improvements. SynGP500 addresses a critical national gap, providing researchers and educators with a resource for developing and evaluating clinical NLP methods for Australian general practice while inherently protecting patient privacy.
Authors: Yiliu Sun, Zicheng Zhao, Yang Wei, Yanfang Zhang, Chen Gong
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) significantly enhances the reasoning capability of Large Language Models (LLMs). Current RLVR approaches typically conduct training across all generated tokens, but neglect to explore which tokens (e.g., prefix tokens) actually contribute to reasoning. This uniform training strategy spends substantial effort on optimizing low-return tokens, which in turn impedes the potential improvement from high-return tokens and reduces overall training effectiveness. To address this issue, we propose a novel RLVR approach called Progressive Prefix-token Policy Optimization (PPPO), which highlights the significance of the prefix segment of generated outputs. Specifically, inspired by the well-established human thinking theory of Path Dependence, where early-stage thoughts substantially constrain subsequent thinking trajectory, we identify an analogous phenomenon in LLM reasoning termed Beginning Lock-in Effect (BLE). PPPO leverages this finding by focusing its optimization objective on the prefix reasoning process of LLMs. This targeted optimization strategy can positively influence subsequent reasoning processes, and ultimately improve final results. To improve the learning effectiveness of LLMs on how to start reasoning with high quality, PPPO introduces two training strategies: (a) Progressive Prefix Retention, which shapes a progressive learning process by increasing the proportion of retained prefix tokens during training; (b) Continuation Accumulated Reward, which mitigates reward bias by sampling multiple continuations for one prefix token sequence, and accumulating their scores as the reward signal. Extensive experimental results on various reasoning tasks demonstrate that our proposed PPPO outperforms representative RLVR methods, with the accuracy improvements of 18.02% on only 26.17% training tokens.
Authors: Xiaotian Zhang, Yuan Wang, Ruizhe Chen, Zeya Wang, Runchen Hou, Zuozhu Liu
Abstract: The deployment of Large Language Models (LLMs) in interactive systems necessitates a deep alignment with the nuanced and dynamic preferences of individual users. Current alignment techniques predominantly address universal human values or static, single-turn preferences, thereby failing to address the critical needs of long-term personalization and the initial user cold-start problem. To bridge this gap, we propose PersonalAgent, a novel user-centric lifelong agent designed to continuously infer and adapt to user preferences. PersonalAgent constructs and dynamically refines a unified user profile by decomposing dialogues into single-turn interactions, framing preference inference as a sequential decision-making task. Experiments show that PersonalAgent achieves superior performance over strong prompt-based and policy optimization baselines, not only in idealized but also in noisy conversational contexts, while preserving cross-session preference consistency. Furthermore, human evaluation confirms that PersonalAgent excels at capturing user preferences naturally and coherently. Our findings underscore the importance of lifelong personalization for developing more inclusive and adaptive conversational agents. Our code is available here.
Authors: Charan Prakash Rathore, Saumi Ray, Dhruv Kumar
Abstract: Extracting structured information from zeolite synthesis experimental procedures is critical for materials discovery, yet existing methods have not systematically evaluated Large Language Models (LLMs) for this domain-specific task. This work addresses a fundamental question: what is the efficacy of different prompting strategies when applying LLMs to scientific information extraction? We focus on four key subtasks: event type classification (identifying synthesis steps), trigger text identification (locating event mentions), argument role extraction (recognizing parameter types), and argument text extraction (extracting parameter values). We evaluate four prompting strategies - zero-shot, few-shot, event-specific, and reflection-based - across six state-of-the-art LLMs (Gemma-3-12b-it, GPT-5-mini, O4-mini, Claude-Haiku-3.5, DeepSeek reasoning and non-reasoning) using the ZSEE dataset of 1,530 annotated sentences. Results demonstrate strong performance on event type classification (80-90\% F1) but modest performance on fine-grained extraction tasks, particularly argument role and argument text extraction (50-65\% F1). GPT-5-mini exhibits extreme prompt sensitivity with 11-79\% F1 variation. Notably, advanced prompting strategies provide minimal improvements over zero-shot approaches, revealing fundamental architectural limitations. Error analysis identifies systematic hallucination, over-generalization, and inability to capture synthesis-specific nuances. Our findings demonstrate that while LLMs achieve high-level understanding, precise extraction of experimental parameters requires domain-adapted models, providing quantitative benchmarks for scientific information extraction.
Authors: Ayman Eddakrouri (Effat University), Amani Ramadan (Cairo University)
Abstract: This study applies Brookes' Measure of Categorical Dispersion ({\Delta}) to analyze the thematic structure of contemporary Arabic Applied Linguistics research. Using a comprehensive, real-world dataset of 1,564 publications from 2019 to 2025, classified into eight core sub-disciplines, we calculate a dispersion index of {\Delta} = 0.194. This remarkably low value indicates extreme thematic dispersion, revealing that the field is characterized by pronounced heterogeneity rather than concentration. The analysis identifies Computational Linguistics as a dominant but non-hegemonic force, coexisting with robust research in Sociolinguistics, Language Teaching, and other subfields. This study clarifies the correct application of Brookes' original formula, demonstrates its utility for field characterization, and provides a replicable bibliometric methodology for assessing disciplinary structure across domains.
Authors: Joao Queiroz
Abstract: Recent evidence shows that the versification of prompts constitutes a highly effective adversarial mechanism against aligned LLMs. The study 'Adversarial poetry as a universal single-turn jailbreak mechanism in large language models' demonstrates that instructions routinely refused in prose become executable when rewritten as verse, producing up to 18 x more safety failures in benchmarks derived from MLCommons AILuminate. Manually written poems reach approximately 62% ASR, and automated versions 43%, with some models surpassing 90% success in single-turn interactions. The effect is structural: systems trained with RLHF, constitutional AI, and hybrid pipelines exhibit consistent degradation under minimal semiotic formal variation. Versification displaces the prompt into sparsely supervised latent regions, revealing guardrails that are excessively dependent on surface patterns. This dissociation between apparent robustness and real vulnerability exposes deep limitations in current alignment regimes. The absence of evaluations in Portuguese, a language with high morphosyntactic complexity, a rich metric-prosodic tradition, and over 250 million speakers, constitutes a critical gap. Experimental protocols must parameterise scansion, metre, and prosodic variation to test vulnerabilities specific to Lusophone patterns, which are currently ignored.
Authors: Zhengyi Zhao, Shubo Zhang, Yuxi Zhang, Huimin Wang, Binyang Li, Kam-Fai Wong
Abstract: Large Language Models (LLMs) have shown impressive capabilities in complex reasoning tasks. However, current approaches employ uniform language density for both intermediate reasoning and final answers, leading to computational inefficiency. Our observation found that reasoning process serves a computational function for the model itself, while answering serves a communicative function for human understanding. This distinction enables the use of compressed, symbol-rich language for intermediate computations while maintaining human-readable final explanations. To address this inefficiency, we present Denser: \underline{D}ual-d\underline{ens}ity inf\underline{er}ence, a novel framework that optimizes information density separately for reasoning and answering phases. Our framework implements this through three components: a query processing module that analyzes input problems, a high-density compressed reasoning mechanism for efficient intermediate computations, and an answer generation component that translates compressed reasoning into human-readable solutions. Experimental evaluation across multiple reasoning question answering benchmarks demonstrates that Denser reduces token consumption by up to 62\% compared to standard Chain-of-Thought methods while preserving or improving accuracy. These efficiency gains are particularly significant for complex multi-step reasoning problems where traditional methods generate extensive explanations.
Authors: Lev Kharlashkin, Eiaki Morooka, Yehor Tereshchenko, Mika H\"am\"al\"ainen
Abstract: ORACLE turns daily news into week-over-week, decision-ready insights for one of the Finnish University of Applied Sciences. The platform crawls and versions news, applies University-specific relevance filtering, embeds content, classifies items into PESTEL dimensions and builds a concise Time-Dependent Recursive Summary Graph (TRSG): two clustering layers summarized by an LLM and recomputed weekly. A lightweight change detector highlights what is new, removed or changed, then groups differences into themes for PESTEL-aware analysis. We detail the pipeline, discuss concrete design choices that make the system stable in production and present a curriculum-intelligence use case with an evaluation plan.
Authors: Run-ze Hu, Yang Yang, Yi-hang Yang, Jing-qi Kong, Jia-hui Luo, Wen-yu Yang, Jing Chen, Jing-yao Liu, Hui-qun Zeng, Lei Zhang, Zheng Liu
Abstract: Background: Motivational interviewing (MI) is an effective counseling approach for promoting health behavior change, but its impact is constrained by the need for highly trained human counselors. Objective: This study aimed to explore a scalable alternative by developing and evaluating Large Language Models for Motivational Interviewing (MI-LLMs). Methods: We first curated five Chinese psychological counseling corpora and, using GPT-4 with an MI-informed prompt, transcribed multi-turn dialogues from the two highest-quality datasets (CPsyCounD and PsyDTCorpus) into 2,040 MI-style counseling conversations, of which 2,000 were used for training and 40 for testing. Three Chinese-capable open-source LLMs (Baichuan2-7B-Chat, ChatGLM-4-9B-Chat and Llama-3-8B-Chinese-Chat-v2) were fine-tuned on this corpus and were named as MI-LLMs. We evaluated MI-LLMs using round-based automatic metrics and expert manual coding with the Motivational Interviewing Treatment Integrity (MITI) Coding Manual 4.2.1. Results: Across all three models, fine-tuning substantially improved BLEU-4 and ROUGE scores compared with the base models, and manual coding showed that MI-LLMs achieved technical and relational global scores, and MI-adherent ratios that approached those of real MI dialogues, although complex reflections and reflection-to-question ratios remained less frequent. Conclusions: These findings provide initial evidence that MI-oriented fine-tuning can endow general-purpose LLMs with core MI-consistent counseling behaviors, suggesting a scalable pathway toward AI-assisted health behavior change support while underscoring the need for further work on data scale, complex MI skills and real-world intervention trials.
Authors: Md. Samiul Alim, Mahir Shahriar Tamim, Maisha Rahman, Tanvir Ahmed Khan, Md Mushfique Anwar
Abstract: Sentiment analysis, an emerging research area within natural language processing (NLP), has primarily been explored in contexts like elections and social media trends, but there remains a significant gap in understanding emotional dynamics during civil unrest, particularly in the Bangla language. Our study pioneers sentiment analysis in Bangla during a national crisis by examining public emotions amid Bangladesh's 2024 mass uprising. We curated a unique dataset of 2,028 annotated news headlines from major Facebook news portals, classifying them into Outrage, Hope, and Despair. Through Latent Dirichlet Allocation (LDA), we identified prevalent themes like political corruption and public protests, and analyzed how events such as internet blackouts shaped sentiment patterns. It outperformed multilingual transformers (mBERT: 67%, XLM-RoBERTa: 71%) and traditional machine learning methods (SVM and Logistic Regression: both 70%). These results highlight the effectiveness of language-specific models and offer valuable insights into public sentiment during political turmoil.
Authors: Kuan Lu, Shuhang Lin, Sai Wu, Yichen Yao, Junhan Yang, Huan Li, Wei Chu, Xu Yinghui, Yuan Qi, Gang Chen
Abstract: Large language models (LLMs) are increasingly applied in long-context scenarios such as multi-turn conversations. However, long contexts pose significant challenges for inference efficiency, including high memory overhead from Key-Value (KV) cache and increased latency due to excessive memory accesses. Recent methods for dynamic KV selection struggle with trade-offs: block-level indexing degrades accuracy by retrieving irrelevant KV entries, while token-level indexing incurs high latency from inefficient retrieval mechanisms. In this paper, we propose CTKVR, a novel centroid-then-token KV retrieval scheme that addresses these limitations. CTKVR leverages a key observation: query vectors adjacent in position exhibit high similarity after Rotary Position Embedding (RoPE) and share most of their top-k KV cache entries. Based on this insight, CTKVR employs a two-stage retrieval strategy: lightweight centroids are precomputed during prefilling for centroid-grained indexing, followed by token-level refinement for precise KV retrieval. This approach balances retrieval efficiency and accuracy. To further enhance performance, we implement an optimized system for indexing construction and search using CPU-GPU co-execution. Experimentally, CTKVR achieves superior performance across multiple benchmarks with less than 1% accuracy degradation. Meanwhile, CTKVR delivers 3 times and 4 times throughput speedups on Llama-3-8B and Yi-9B at 96K context length across diverse GPU hardware.
Authors: Peter Dekker, Heikki Rasilo, Bart de Boer
Abstract: The concept of inflection classes is an abstraction used by linguists, and provides a means to describe patterns in languages that give an analogical base for deducing previously unencountered forms. This ability is an important part of morphological acquisition and processing. We study the learnability of a system of verbal inflection classes by the individual language user by performing unsupervised clustering of lexemes into inflection classes. As a cognitively plausible and interpretable computational model, we use Adaptive Resonance Theory, a neural network with a parameter that determines the degree of generalisation (vigilance). The model is applied to Latin, Portuguese and Estonian. The similarity of clustering to attested inflection classes varies depending on the complexity of the inflectional system. We find the best performance in a narrow region of the generalisation parameter. The learned features extracted from the model show similarity with linguistic descriptions of the inflection classes. The proposed model could be used to study change in inflection classes in the future, by including it in an agent-based model.
Authors: Dakota Ellis, Samy Bakikerali, Wanshan Chen, Bao Dinh, Uyen Le
Abstract: Traditional linguists have proposed the use of a General Service List (GSL) to assist new language learners in identifying the most important words in English. This process requires linguistic expertise, subjective input, and a considerable amount of time. We attempt to create our own GSL and evaluate its practicality against the industry standard (The NGSL). We found creating a Specialized Word List (SWL), or a word list specific to a subset of the overall corpus, to be the most practical way for language-learners to optimize the process. The SWL's that we created using our model outperformed the industry standard, reaching the 95% coverage required for language comprehension with fewer words comparatively. By restricting the SWL process to objective criteria only, it can be automated, scaled, and tailored to the needs of language-learners across the globe.
Authors: Lifeng Han, Gareth J. F. Jones, Alan F. Smeaton
Abstract: Word meaning, representation, and interpretation play fundamental roles in natural language understanding (NLU), natural language processing (NLP), and natural language generation (NLG) tasks. Many of the inherent difficulties in these tasks stem from Multi-word Expressions (MWEs), which complicate the tasks by introducing ambiguity, idiomatic expressions, infrequent usage, and a wide range of variations. Significant effort and substantial progress have been made in addressing the challenging nature of MWEs in Western languages, particularly English. This progress is attributed in part to the well-established research communities and the abundant availability of computational resources. However, the same level of progress is not true for language families such as Chinese and closely related Asian languages, which continue to lag behind in this regard. While sub-word modelling has been successfully applied to many Western languages to address rare words improving phrase comprehension, and enhancing machine translation (MT) through techniques like byte-pair encoding (BPE), it cannot be applied directly to ideograph language scripts like Chinese. In this work, we conduct a systematic study of the Chinese character decomposition technology in the context of MWE-aware neural machine translation (NMT). Furthermore, we report experiments to examine how Chinese character decomposition technology contributes to the representation of the original meanings of Chinese words and characters, and how it can effectively address the challenges of translating MWEs.
Authors: Benjamin Minixhofer, Tyler Murray, Tomasz Limisiewicz, Anna Korhonen, Luke Zettlemoyer, Noah A. Smith, Edoardo M. Ponti, Luca Soldaini, Valentin Hofmann
Abstract: We introduce Bolmo, the first family of competitive fully open byte-level language models (LMs) at the 1B and 7B parameter scales. In contrast to prior research on byte-level LMs, which focuses predominantly on training from scratch, we train Bolmo by byteifying existing subword-level LMs. Byteification enables overcoming the limitations of subword tokenization - such as insufficient character understanding and efficiency constraints due to the fixed subword vocabulary - while performing at the level of leading subword-level LMs. Bolmo is specifically designed for byteification: our architecture resolves a mismatch between the expressivity of prior byte-level architectures and subword-level LMs, which makes it possible to employ an effective exact distillation objective between Bolmo and the source subword model. This allows for converting a subword-level LM to a byte-level LM by investing less than 1\% of a typical pretraining token budget. Bolmo substantially outperforms all prior byte-level LMs of comparable size, and outperforms the source subword-level LMs on character understanding and, in some cases, coding, while coming close to matching the original LMs' performance on other tasks. Furthermore, we show that Bolmo can achieve inference speeds competitive with subword-level LMs by training with higher token compression ratios, and can be cheaply and effectively post-trained by leveraging the existing ecosystem around the source subword-level LM. Our results finally make byte-level LMs a practical choice competitive with subword-level LMs across a wide set of use cases.
Authors: Hongbin Na, Zimu Wang, Zhaoming Chen, Peilin Zhou, Yining Hua, Grace Ziqi Zhou, Haiyang Zhang, Tao Shen, Wei Wang, John Torous, Shaoxiong Ji, Ling Chen
Abstract: Psychological defenses are strategies, often automatic, that people use to manage distress. Rigid or overuse of defenses is negatively linked to mental health and shapes what speakers disclose and how they accept or resist help. However, defenses are complex and difficult to reliably measure, particularly in clinical dialogues. We introduce PsyDefConv, a dialogue corpus with help seeker utterances labeled for defense level, and DMRS Co-Pilot, a four-stage pipeline that provides evidence-based pre-annotations. The corpus contains 200 dialogues and 4709 utterances, including 2336 help seeker turns, with labeling and Cohen's kappa 0.639. In a counterbalanced study, the co-pilot reduced average annotation time by 22.4%. In expert review, it averaged 4.62 for evidence, 4.44 for clinical plausibility, and 4.40 for insight on a seven-point scale. Benchmarks with strong language models in zero-shot and fine-tuning settings demonstrate clear headroom, with the best macro F1-score around 30% and a tendency to overpredict mature defenses. Corpus analyses confirm that mature defenses are most common and reveal emotion-specific deviations. We will release the corpus, annotations, code, and prompts to support research on defensive functioning in language.
Authors: Kester Clegg, Richard Hawkins, Ibrahim Habli, Tom Lawton
Abstract: LLMs (Large Language Models) are increasingly used in text processing pipelines to intelligently respond to a variety of inputs and generation tasks. This raises the possibility of replacing human roles that bottleneck existing information flows, either due to insufficient staff or process complexity. However, LLMs make mistakes and some processing roles are safety critical. For example, triaging post-operative care to patients based on hospital referral letters, or updating site access schedules in nuclear facilities for work crews. If we want to introduce LLMs into critical information flows that were previously performed by humans, how can we make them safe and reliable? Rather than make performative claims about augmented generation frameworks or graph-based techniques, this paper argues that the safety argument should focus on the type of evidence we get from evaluation points in LLM processes, particularly in frameworks that employ LLM-as-Judges (LaJ) evaluators. This paper argues that although we cannot get deterministic evaluations from many natural language processing tasks, by adopting a basket of weighted metrics it may be possible to lower the risk of errors within an evaluation, use context sensitivity to define error severity and design confidence thresholds that trigger human review of critical LaJ judgments when concordance across evaluators is low.
Authors: Darshita Rathore, Vineet Kumar, Chetna Bansal, Anindya Moitra
Abstract: Large language models are increasingly adapted to downstream tasks through fine-tuning. Full supervised fine-tuning (SFT) and parameter-efficient fine-tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA), are two dominant approaches. While PEFT methods are widely used for their computational efficiency, the implications of their configurations (e.g., rank) remain under-explored in downstream Q&A tasks and generalisation. In this work, we perform a comprehensive evaluation across multiple reasoning and recall datasets, conducting a rank sweep to quantify the trade-off between SFT and PEFT. We also compare the accuracy of PEFT and SFT models across in-domain and out-of-domain adaptation, highlighting distinct generalisation behaviour and task-specific forgetting. We demonstrate that LoRA achieves competitive and in some cases superior performance compared to SFT, particularly on reasoning tasks at specific rank values. Additionally, we analyze the internal representations via spectral features and layer-wise attention structures, offering insights into representational drift and structural changes in attention patterns.
Authors: Tamanna Hossain, Robert L. Logan IV, Ganesh Jagadeesan, Sameer Singh, Joel Tetreault, Alejandro Jaimes
Abstract: State space models (SSMs) are a promising alternative to transformers for language modeling because they use fixed memory during inference. However, this fixed memory usage requires some information loss in the hidden state when processing long sequences. While prior work has studied the sequence length at which this information loss occurs, it does not characterize the types of information SSM language models (LMs) tend to forget. In this paper, we address this knowledge gap by identifying the types of tokens (e.g., parts of speech, named entities) and sequences (e.g., code, math problems) that are more frequently forgotten by SSM LMs. We achieve this by training an auto-encoder to reconstruct sequences from the SSM's hidden state, and measure information loss by comparing inputs with their reconstructions. We perform experiments using the Mamba family of SSM LMs (130M--1.4B) on sequences ranging from 4--256 tokens. Our results show significantly higher rates of information loss on math-related tokens (e.g., numbers, variables), mentions of organization entities, and alternative dialects to Standard American English. We then examine the frequency that these tokens appear in Mamba's pretraining data and find that less prevalent tokens tend to be the ones Mamba is most likely to forget. By identifying these patterns, our work provides clear direction for future research to develop methods that better control Mamba's ability to retain important information.
Authors: Xiaodi Li, Dingcheng Li, Rujun Gao, Mahmoud Zamani, Feng Mi, Latifur Khan
Abstract: Continual learning remains a fundamental challenge in machine learning, requiring models to learn from a stream of tasks without forgetting previously acquired knowledge. A major obstacle in this setting is catastrophic forgetting, where performance on earlier tasks degrades as new tasks are learned. In this paper, we introduce PPSEBM, a novel framework that integrates an Energy-Based Model (EBM) with Progressive Parameter Selection (PPS) to effectively address catastrophic forgetting in continual learning for natural language processing tasks. In PPSEBM, progressive parameter selection allocates distinct, task-specific parameters for each new task, while the EBM generates representative pseudo-samples from prior tasks. These generated samples actively inform and guide the parameter selection process, enhancing the model's ability to retain past knowledge while adapting to new tasks. Experimental results on diverse NLP benchmarks demonstrate that PPSEBM outperforms state-of-the-art continual learning methods, offering a promising and robust solution to mitigate catastrophic forgetting.
Authors: Adam Karvonen, James Chua, Cl\'ement Dumas, Kit Fraser-Taliente, Subhash Kantamneni, Julian Minder, Euan Ong, Arnab Sen Sharma, Daniel Wen, Owain Evans, Samuel Marks
Abstract: Large language model (LLM) activations are notoriously difficult to understand, with most existing techniques using complex, specialized methods for interpreting them. Recent work has proposed a simpler approach known as LatentQA: training LLMs to directly accept LLM activations as inputs and answer arbitrary questions about them in natural language. However, prior work has focused on narrow task settings for both training and evaluation. In this paper, we instead take a generalist perspective. We evaluate LatentQA-trained models, which we call Activation Oracles (AOs), in far out-of-distribution settings and examine how performance scales with training data diversity. We find that AOs can recover information fine-tuned into a model (e.g., biographical knowledge or malign propensities) that does not appear in the input text, despite never being trained with activations from a fine-tuned model. Our main evaluations are four downstream tasks where we can compare to prior white- and black-box techniques. We find that even narrowly-trained LatentQA models can generalize well, and that adding additional training datasets (such as classification tasks and a self-supervised context prediction task) yields consistent further improvements. Overall, our best AOs match or exceed prior white-box baselines on all four tasks and are the best method on 3 out of 4. These results suggest that diversified training to answer natural-language queries imparts a general capability to verbalize information about LLM activations.
Authors: Pinxian Lu, Nimra Ishfaq, Emma Win, Morgan Rose, Sierra R Strickland, Candice L Biernesser, Jamie Zelazny, Munmun De Choudhury
Abstract: Online harassment has been a persistent issue in the online space. Predominantly, research focused on online harassment in public social media platforms, while less is placed on private messaging platforms. To address online harassment on one private messaging platform, Instagram, we leverage the capabilities of Large Language Models (LLMs). To achieve this, we recruited human labelers to identify online harassment in an Instagram messages dataset. Using the previous conversation as context, we utilize an LLM pipeline to conduct large-scale labeling on Instagram messages and evaluate its performance against human labels. Then, we use LLM to generate and evaluate simulated responses to online harassment messages. We find that the LLM labeling pipeline is capable of identifying online harassment in private messages. By comparing human responses and simulated responses, we also demonstrate that our simulated responses are superior in helpfulness compared to original human responses.
Authors: Krunal Jesani, Dmitry Ignatov, Radu Timofte
Abstract: Neural architecture search (NAS) traditionally requires significant human expertise or automated trial-and-error to design deep learning models. We present NN-Caption, an LLM-guided neural architecture search pipeline that generates runnable image-captioning models by composing CNN encoders from LEMUR's classification backbones with sequence decoders (LSTM/GRU/Transformer) under a strict Net API. Using DeepSeek-R1-0528-Qwen3-8B as the primary generator, we present the prompt template and examples of generated architectures. We evaluate on MS COCO with BLEU-4. The LLM generated dozens of captioning models, with over half successfully trained and producing meaningful captions. We analyse the outcomes of using different numbers of input model snippets (5 vs. 10) in the prompt, finding a slight drop in success rate when providing more candidate components. We also report training dynamics (caption accuracy vs. epochs) and the highest BLEU-4 attained. Our results highlight the promise of LLM-guided NAS: the LLM not only proposes architectures but also suggests hyperparameters and training practices. We identify the challenges encountered (e.g., code hallucinations or API compliance issues) and detail how prompt rules and iterative code fixes addressed them. This work presents a pipeline that integrates prompt-based code generation with automatic evaluation, and adds dozens of novel captioning models to the open LEMUR dataset to facilitate reproducible benchmarking and downstream AutoML research.
Authors: Ryan Cartularo
Abstract: Sepsis accounts for nearly 20% of global ICU admissions, yet conventional prediction models often fail to effectively integrate heterogeneous data streams, remaining either siloed by modality or reliant on brittle early fusion. In this work, we present a rigorous architectural comparison between End-to-End Deep Fusion and Context-Aware Stacking for sepsis tasks. We initially hypothesized that a novel Quad-Modal Hierarchical Gated Attention Network -- termed SepsisFusionFormer -- would resolve complex cross-modal interactions between vitals, text, and imaging. However, experiments on MIMIC-IV revealed that SepsisFusionFormer suffered from "attention starvation" in the small antibiotic cohort ($N \approx 2,100$), resulting in overfitting (AUC 0.66). This counterintuitive result informed the design of SepsisLateFusion, a "leaner" Context-Aware Mixture-of-Experts (MoE) architecture. By treating modalities as orthogonal experts -- the "Historian" (Static), the "Monitor" (Temporal), and the "Reader" (NLP) -- and dynamically gating them via a CatBoost meta-learner, we achieved State-of-the-Art (SOTA) performance: 0.915 AUC for prediction 4 hours prior to clinical onset. By calibrating the decision threshold for clinical safety, we reduced missed cases by 48% relative to the default operating point, thus opening a true preventative window for timely intervention over reactive alerts. Furthermore, for the novel prescriptive task of multi-class antibiotic selection, we demonstrate that a Quad-Modal Ensemble achieved the highest performance (0.72 AUC). These models are integrated into SepsisSuite, a deployment-ready Python framework for clinical decision support. SepsisSuite is available for free at: https://github.com/RyanCartularo/SepsisSuite-Info
Authors: Dizhan Xue, Jing Cui, Shengsheng Qian, Chuanrui Hu, Changsheng Xu
Abstract: Intelligent agents powered by large language models (LLMs) have recently demonstrated impressive capabilities and gained increasing popularity on social media platforms. While LLM agents are reshaping the ecology of social media, there exists a current gap in conducting a comprehensive evaluation of their ability to comprehend media content, understand user behaviors, and make intricate decisions. To address this challenge, we introduce SoMe, a pioneering benchmark designed to evaluate social media agents equipped with various agent tools for accessing and analyzing social media data. SoMe comprises a diverse collection of 8 social media agent tasks, 9,164,284 posts, 6,591 user profiles, and 25,686 reports from various social media platforms and external websites, with 17,869 meticulously annotated task queries. Compared with the existing datasets and benchmarks for social media tasks, SoMe is the first to provide a versatile and realistic platform for LLM-based social media agents to handle diverse social media tasks. By extensive quantitative and qualitative analysis, we provide the first overview insight into the performance of mainstream agentic LLMs in realistic social media environments and identify several limitations. Our evaluation reveals that both the current closed-source and open-source LLMs cannot handle social media agent tasks satisfactorily. SoMe provides a challenging yet meaningful testbed for future social media agents. Our code and data are available at https://github.com/LivXue/SoMe
Authors: Zhengxu Yan, Han Li, Yuming Feng
Abstract: With the growing ease of academic publishing, the volume of research papers, especially in AI-related fields, has surged dramatically. This flood of publications makes it difficult for truly novel and impactful work to stand out, and manual novelty assessment is often unstable and time-consuming. Our project aims to develop a model that estimates and ranks the conceptual novelty of AI papers, enabling a data-driven and scalable assessment of research originality. Such a system can help researchers efficiently identify submissions that introduce genuinely innovative ideas rather than minor variants, and provide conference reviewers with a quantitative and consistent signal of novelty. Our approach evaluates novelty primarily through a paper's title, abstract, and semantic similarity to prior literature. Given the motivation of novelty estimation, we explore two task formulations with different modeling objectives, each offering a different perspective: (1) binary classification, which predicts the paper's absolute novelty from learned patterns of prior novel works, and (2) pairwise novelty comparison, which learns to distinguish papers by relative novelty over others. We fine-tune Qwen3-4B-Instruct-2507 and SciBERT on both tasks, benchmarking against GPT-5.1 to analyze how task formulation and modeling choices affect performance. The implementation is publicly available at https://github.com/ZhengxuYan/NoveltyRank.
Authors: Jianshuo Dong, Yutong Zhang, Yan Liu, Zhenyu Zhong, Tao Wei, Chao Zhang, Han Qiu
Abstract: Advanced LLMs have achieved near-ceiling instruction-following accuracy on benchmarks such as IFEval. However, these impressive scores do not necessarily translate to reliable services in real-world use, where users often vary their phrasing, contextual framing, and task formulations. In this paper, we study nuance-oriented reliability: whether models exhibit consistent competence across cousin prompts that convey analogous user intents but with subtle nuances. To quantify this, we introduce a new metric, reliable@k, and develop an automated pipeline that generates high-quality cousin prompts via data augmentation. Building upon this, we construct IFEval++ for systematic evaluation. Across 20 proprietary and 26 open-source LLMs, we find that current models exhibit substantial insufficiency in nuance-oriented reliability -- their performance can drop by up to 61.8% with nuanced prompt modifications. What's more, we characterize it and explore three potential improvement recipes. Our findings highlight nuance-oriented reliability as a crucial yet underexplored next step toward more dependable and trustworthy LLM behavior. Our code and benchmark are accessible: https://github.com/jianshuod/IFEval-pp.
Authors: Advait Gosai, Tyler Vuong, Utkarsh Tyagi, Steven Li, Wenjia You, Miheer Bavare, Arda U\c{c}ar, Zhongwang Fang, Brian Jang, Bing Liu, Yunzhong He
Abstract: End-to-end (E2E) spoken dialogue systems are increasingly replacing cascaded pipelines for voice-based human-AI interaction, processing raw audio directly without intermediate transcription. Existing benchmarks primarily evaluate these models on synthetic speech and single-turn tasks, leaving realistic multi-turn conversational ability underexplored. We introduce Audio MultiChallenge, an open-source benchmark to evaluate E2E spoken dialogue systems under natural multi-turn interaction patterns. Building on the text-based MultiChallenge framework, which evaluates Inference Memory, Instruction Retention, and Self Coherence, we introduce a new axis Voice Editing that tests robustness to mid-utterance speech repairs and backtracking. We further augment each axis to the audio modality, such as introducing Audio-Cue challenges for Inference Memory that require recalling ambient sounds and paralinguistic signals beyond semantic content. We curate 452 conversations from 47 speakers with 1,712 instance-specific rubrics through a hybrid audio-native agentic and human-in-the-loop pipeline that exposes model failures at scale while preserving natural disfluencies found in unscripted human speech. Our evaluation of proprietary and open-source models reveals that even frontier models struggle on our benchmark, with Gemini 3 Pro Preview (Thinking), our highest-performing model achieving a 54.65% pass rate. Error analysis shows that models fail most often on our new axes and that Self Coherence degrades with longer audio context. These failures reflect difficulty of tracking edits, audio cues, and long-range context in natural spoken dialogue. Audio MultiChallenge provides a reproducible testbed to quantify them and drive improvements in audio-native multi-turn interaction capability.
Authors: Darrin O' Brien, Dhikshith Gajulapalli, Eric Xia
Abstract: Results in interpretability suggest that large vision and language models learn implicit linear encodings when models are biased by in-context prompting. However, the existence of similar linear representations in more general adaptation regimes has not yet been demonstrated. In this work, we develop the concept of a task matrix, a linear transformation from a base to finetuned embedding state. We demonstrate that for vision and text models and ten different datasets, a base model augmented with a task matrix achieves results surpassing linear probes, sometimes approaching finetuned levels. Our results validate the existence of cross-layer linear encodings between pretrained and finetuned architectures. Moreover, we show that a data-based approximation for such encodings is both efficient and generalizable to multiple domains. We make our implementation publicly available.
Authors: Yaniv Leviathan, Matan Kalman, Yossi Matias
Abstract: When not using reasoning, repeating the input prompt improves performance for popular models (Gemini, GPT, Claude, and Deepseek) without increasing the number of generated tokens or latency.
Authors: Ruiyi Zhang, Peijia Qin, Qi Cao, Pengtao Xie
Abstract: Process Reward Models (PRMs) have become essential for improving Large Language Models (LLMs) via test-time scaling, yet their effectiveness in coding remains limited due to the lack of meaningful step decompositions in code and the noise of Monte-Carlo-generated partial labels. We propose DreamPRM-Code, a coding-focused PRM that treats functions as reasoning steps using a Chain-of-Function prompting strategy to induce modular code generation, enabling PRM training and application analogous to mathematical reasoning tasks. To address label noise, DreamPRM-Code introduces a meta-learning-based correction mechanism that leverages clean final-solution unit-test labels and performs bi-level optimization to refine intermediate labels. Applying on test-time scaling, DreamPRM-Code achieved state-of-the-art performance on LiveCodeBench with 80.9 pass@1 rate, surpassing OpenAI o4-mini.
Authors: Yunheng Wang, Yixiao Feng, Yuetong Fang, Shuning Zhang, Tan Jing, Jian Li, Xiangrui Jiang, Renjing Xu
Abstract: 3D Scene Graphs (3DSGs) constitute a powerful representation of the physical world, distinguished by their abilities to explicitly model the complex spatial, semantic, and functional relationships between entities, rendering a foundational understanding that enables agents to interact intelligently with their environment and execute versatile behaviors. Embodied navigation, as a crucial component of such capabilities, leverages the compact and expressive nature of 3DSGs to enable long-horizon reasoning and planning in complex, large-scale environments. However, prior works rely on a static-world assumption, defining traversable space solely based on static spatial layouts and thereby treating interactable obstacles as non-traversable. This fundamental limitation severely undermines their effectiveness in real-world scenarios, leading to limited reachability, low efficiency, and inferior extensibility. To address these issues, we propose HERO, a novel framework for constructing Hierarchical Traversable 3DSGs, that redefines traversability by modeling operable obstacles as pathways, capturing their physical interactivity, functional semantics, and the scene's relational hierarchy. The results show that, relative to its baseline, HERO reduces PL by 35.1% in partially obstructed environments and increases SR by 79.4% in fully obstructed ones, demonstrating substantially higher efficiency and reachability.
Authors: Debu Sinha
Abstract: Retrieval-Augmented Generation (RAG) systems remain susceptible to hallucinations despite grounding in retrieved evidence. Current detection methods rely on semantic similarity and natural language inference (NLI), but their fundamental limitations have not been rigorously characterized. We apply conformal prediction to hallucination detection, providing finite-sample coverage guarantees that enable precise quantification of detection capabilities. Using calibration sets of approximately 600 examples, we achieve 94% coverage with 0% false positive rate on synthetic hallucinations (Natural Questions). However, on three real hallucination benchmarks spanning multiple LLMs (GPT-4, ChatGPT, GPT-3, Llama-2, Mistral), embedding-based methods - including state-of-the-art OpenAI text-embedding-3-large and cross-encoder models - exhibit unacceptable false positive rates: 100% on HaluEval, 88% on RAGTruth, and 50% on WikiBio. Crucially, GPT-4 as an LLM judge achieves only 7% FPR (95% CI: [3.4%, 13.7%]) on the same data, proving the task is solvable through reasoning. We term this the "semantic illusion": semantically plausible hallucinations preserve similarity to source documents while introducing factual errors invisible to embeddings. This limitation persists across embedding architectures, LLM generators, and task types, suggesting embedding-based detection is insufficient for production RAG deployment.
Authors: Richard Helder Moulton, Austin O'Brien, John D. Hastings
Abstract: Although large language models (LLMs) are increasingly used in security-critical workflows, practitioners lack quantitative guidance on which safeguards are worth deploying. This paper introduces a decision-oriented framework and reproducible methodology that together quantify residual risk, convert adversarial probe outcomes into financial risk estimates and return-on-control (RoC) metrics, and enable monetary comparison of layered defenses for LLM-based systems. A retrieval-augmented generation (RAG) service is instantiated using the DeepSeek-R1 model over a corpus containing synthetic personally identifiable information (PII), and subjected to automated attacks with Garak across five vulnerability classes: PII leakage, latent context injection, prompt injection, adversarial attack generation, and divergence. For each (vulnerability, control) pair, attack success probabilities are estimated via Laplace's Rule of Succession and combined with loss triangle distributions, calibrated from public breach-cost data, in 10,000-run Monte Carlo simulations to produce loss exceedance curves and expected losses. Three widely used mitigations, attribute-based access control (ABAC); named entity recognition (NER) redaction using Microsoft Presidio; and NeMo Guardrails, are then compared to a baseline RAG configuration. The baseline system exhibits very high attack success rates (>= 0.98 for PII, latent injection, and prompt injection), yielding a total simulated expected loss of $313k per attack scenario. ABAC collapses success probabilities for PII and prompt-related attacks to near zero and reduces the total expected loss by ~94%, achieving an RoC of 9.83. NER redaction likewise eliminates PII leakage and attains an RoC of 5.97, while NeMo Guardrails provides only marginal benefit (RoC of 0.05).
Authors: Aaron Mueller, Andrew Lee, Shruti Joshi, Ekdeep Singh Lubana, Dhanya Sridhar, Patrik Reizinger
Abstract: A central goal of interpretability is to recover representations of causally relevant concepts from the activations of neural networks. The quality of these concept representations is typically evaluated in isolation, and under implicit independence assumptions that may not hold in practice. Thus, it is unclear whether common featurization methods - including sparse autoencoders (SAEs) and sparse probes - recover disentangled representations of these concepts. This study proposes a multi-concept evaluation setting where we control the correlations between textual concepts, such as sentiment, domain, and tense, and analyze performance under increasing correlations between them. We first evaluate the extent to which featurizers can learn disentangled representations of each concept under increasing correlational strengths. We observe a one-to-many relationship from concepts to features: features correspond to no more than one concept, but concepts are distributed across many features. Then, we perform steering experiments, measuring whether each concept is independently manipulable. Even when trained on uniform distributions of concepts, SAE features generally affect many concepts when steered, indicating that they are neither selective nor independent; nonetheless, features affect disjoint subspaces. These results suggest that correlational metrics for measuring disentanglement are generally not sufficient for establishing independence when steering, and that affecting disjoint subspaces is not sufficient for concept selectivity. These results underscore the importance of compositional evaluations in interpretability research.
Authors: Seok-Hyun Ga, Chun-Yen Chang
Abstract: The rapid development of Generative AI is bringing innovative changes to education and assessment. As the prevalence of students utilizing AI for assignments increases, concerns regarding academic integrity and the validity of assessments are growing. This study utilizes the Earth Science I section of the 2025 Korean College Scholastic Ability Test (CSAT) to deeply analyze the multimodal scientific reasoning capabilities and cognitive limitations of state-of-the-art Large Language Models (LLMs), including GPT-4o, Gemini 2.5 Flash, and Gemini 2.5 Pro. Three experimental conditions (full-page input, individual item input, and optimized multimodal input) were designed to evaluate model performance across different data structures. Quantitative results indicated that unstructured inputs led to significant performance degradation due to segmentation and Optical Character Recognition (OCR) failures. Even under optimized conditions, models exhibited fundamental reasoning flaws. Qualitative analysis revealed that "Perception Errors" were dominant, highlighting a "Perception-Cognition Gap" where models failed to interpret symbolic meanings in schematic diagrams despite recognizing visual data. Furthermore, models demonstrated a "Calculation-Conceptualization Discrepancy," successfully performing calculations while failing to apply the underlying scientific concepts, and "Process Hallucination," where models skipped visual verification in favor of plausible but unfounded background knowledge. Addressing the challenge of unauthorized AI use in coursework, this study provides actionable cues for designing "AI-resistant questions" that target these specific cognitive vulnerabilities. By exploiting AI's weaknesses, such as the gap between perception and cognition, educators can distinguish genuine student competency from AI-generated responses, thereby ensuring assessment fairness.
Authors: Kotaro Funakoshi, Yaoxiong Zhu
Abstract: Recognition of signers' emotions suffers from one theoretical challenge and one practical challenge, namely, the overlap between grammatical and affective facial expressions and the scarcity of data for model training. This paper addresses these two challenges in a cross-lingual setting using our eJSL dataset, a new benchmark dataset for emotion recognition in Japanese Sign Language signers, and BOBSL, a large British Sign Language dataset with subtitles. In eJSL, two signers expressed 78 distinct utterances with each of seven different emotional states, resulting in 1,092 video clips. We empirically demonstrate that 1) textual emotion recognition in spoken language mitigates data scarcity in sign language, 2) temporal segment selection has a significant impact, and 3) incorporating hand motion enhances emotion recognition in signers. Finally we establish a stronger baseline than spoken language LLMs.
Authors: Taichi Aida, Mamoru Komachi, Toshinobu Ogiso, Hiroya Takamura, Daichi Mochihashi
Abstract: Understanding the temporal evolution of sets of vectors is a fundamental challenge across various domains, including ecology, crime analysis, and linguistics. For instance, ecosystem structures evolve due to interactions among plants, herbivores, and carnivores; the spatial distribution of crimes shifts in response to societal changes; and word embedding vectors reflect cultural and semantic trends over time. However, analyzing such time-varying sets of vectors is challenging due to their complicated structures, which also evolve over time. In this work, we propose a novel method for modeling the distribution underlying each set of vectors using infinite-dimensional Gaussian processes. By approximating the latent function in the Gaussian process with Random Fourier Features, we obtain compact and comparable vector representations over time. This enables us to track and visualize temporal transitions of vector sets in a low-dimensional space. We apply our method to both sociological data (crime distributions) and linguistic data (word embeddings), demonstrating its effectiveness in capturing temporal dynamics. Our results show that the proposed approach provides interpretable and robust representations, offering a powerful framework for analyzing structural changes in temporally indexed vector sets across diverse domains.
Authors: Hongbo Zhao, Meng Wang, Fei Zhu, Wenzhuo Liu, Bolin Ni, Fanhu Zeng, Gaofeng Meng, Zhaoxiang Zhang
Abstract: The computational and memory overheads associated with expanding the context window of LLMs severely limit their scalability. A noteworthy solution is vision-text compression (VTC), exemplified by frameworks like DeepSeek-OCR and Glyph, which convert long texts into dense 2D visual representations, thereby achieving token compression ratios of 3x-20x. However, the impact of this high information density on the core long-context capabilities of vision-language models (VLMs) remains under-investigated. To address this gap, we introduce the first benchmark for VTC and systematically assess the performance of VLMs across three long-context understanding settings: VTC-Retrieval, which evaluates the model's ability to retrieve and aggregate information; VTC-Reasoning, which requires models to infer latent associations to locate facts with minimal lexical overlap; and VTC-Memory, which measures comprehensive question answering within long-term dialogue memory. Furthermore, we establish the VTCBench-Wild to simulate diverse input scenarios.We comprehensively evaluate leading open-source and proprietary models on our benchmarks. The results indicate that, despite being able to decode textual information (e.g., OCR) well, most VLMs exhibit a surprisingly poor long-context understanding ability with VTC-compressed information, failing to capture long associations or dependencies in the context.This study provides a deep understanding of VTC and serves as a foundation for designing more efficient and scalable VLMs.
Authors: Chase Walker, Rickard Ewetz
Abstract: Large language models (LLMs) exhibit remarkable capabilities, yet their reasoning remains opaque, raising safety and trust concerns. Attribution methods, which assign credit to input features, have proven effective for explaining the decision making of computer vision models. From these, context attributions have emerged as a promising approach for explaining the behavior of autoregressive LLMs. However, current context attributions produce incomplete explanations by directly relating generated tokens to the prompt, discarding inter-generational influence in the process. To overcome these shortcomings, we introduce the Context Attribution via Graph Explanations (CAGE) framework. CAGE introduces an attribution graph: a directed graph that quantifies how each generation is influenced by both the prompt and all prior generations. The graph is constructed to preserve two properties-causality and row stochasticity. The attribution graph allows context attributions to be computed by marginalizing intermediate contributions along paths in the graph. Across multiple models, datasets, metrics, and methods, CAGE improves context attribution faithfulness, achieving average gains of up to 40%.
Authors: Vincent Huang, Dami Choi, Daniel D. Johnson, Sarah Schwettmann, Jacob Steinhardt
Abstract: Interpreting the internal activations of neural networks can produce more faithful explanations of their behavior, but is difficult due to the complex structure of activation space. Existing approaches to scalable interpretability use hand-designed agents that make and test hypotheses about how internal activations relate to external behavior. We propose to instead turn this task into an end-to-end training objective, by training interpretability assistants to accurately predict model behavior from activations through a communication bottleneck. Specifically, an encoder compresses activations to a sparse list of concepts, and a decoder reads this list and answers a natural language question about the model. We show how to pretrain this assistant on large unstructured data, then finetune it to answer questions. The resulting architecture, which we call a Predictive Concept Decoder, enjoys favorable scaling properties: the auto-interp score of the bottleneck concepts improves with data, as does the performance on downstream applications. Specifically, PCDs can detect jailbreaks, secret hints, and implanted latent concepts, and are able to accurately surface latent user attributes.
Authors: Jaewoo Park, Jungyang Park, Dongju Jang, Jiwan Chung, Byungwoo Yoo, Jaewoo Shin, Seonjoon Park, Taehyeong Kim, Youngjae Yu
Abstract: With the rapid advancement of mathematical reasoning capabilities in Large Language Models (LLMs), AI systems are increasingly being adopted in educational settings to support students' comprehension of problem-solving processes. However, a critical component remains underexplored in current LLM-generated explanations: multimodal explanation. In real-world instructional contexts, human tutors routinely employ visual aids, such as diagrams, markings, and highlights, to enhance conceptual clarity. To bridge this gap, we introduce the multimodal solution explanation task, designed to evaluate whether models can identify visual keypoints, such as auxiliary lines, points, angles, and generate explanations that incorporate these key elements essential for understanding. To evaluate model performance on this task, we propose ME2, a multimodal benchmark consisting of 1,000 math problems annotated with visual keypoints and corresponding explanatory text that references those elements. Our empirical results show that current models struggle to identify visual keypoints. In the task of generating keypoint-based explanations, open-source models also face notable difficulties. This highlights a significant gap in current LLMs' ability to perform mathematical visual grounding, engage in visually grounded reasoning, and provide explanations in educational contexts. We expect that the multimodal solution explanation task and the ME2 dataset will catalyze further research on LLMs in education and promote their use as effective, explanation-oriented AI tutors.
Authors: Ben Anson, Xi Wang, Laurence Aitchison
Abstract: One persistent challenge in LLM research is the development of attention mechanisms that are able to generalise from training on shorter contexts to inference on longer contexts. We propose two conditions that we expect all effective long context attention mechanisms to have: scale-invariant total attention, and scale-invariant attention sparsity. Under a Gaussian assumption, we show that a simple position-dependent transformation of the attention logits is sufficient for these conditions to hold. Experimentally we find that the resulting scale-invariant attention scheme gives considerable benefits in terms of validation loss when zero-shot generalising from training on short contexts to validation on longer contexts, and is effective at long-context retrieval.
Authors: Owen Bianchi, Mathew J. Koretsky, Maya Willey, Chelsea X. Alvarado, Tanay Nayak, Adi Asija, Nicole Kuznetsov, Mike A. Nalls, Faraz Faghri, Daniel Khashabi
Abstract: Large language models (LLMs) face significant challenges with needle-in-ahaystack tasks, where relevant information ("the needle") must be drawn from a large pool of irrelevant context ("the haystack"). Previous studies have highlighted positional bias and distractor quantity as critical factors affecting model performance, yet the influence of gold context size, the length of the answer-containing document, has received little attention. We present the first systematic study of gold context size in long-context question answering, spanning three diverse benchmarks (general knowledge, biomedical reasoning, and mathematical reasoning), eleven state-of-the-art LLMs (including recent reasoning models), and more than 150K controlled runs. Our experiments reveal that LLM performance drops sharply when the gold context is shorter, i.e., smaller gold contexts consistently degrade model performance and amplify positional sensitivity, posing a major challenge for agentic systems that must integrate scattered, fine-grained information of varying lengths. This effect persists under rigorous confounder analysis: even after controlling for gold context position, answer token repetition, gold-to-distractor ratio, distractor volume, and domain specificity, gold context size remains a decisive, independent predictor of success. Our work provides clear insights to guide the design of robust, context-aware LLM-driven systems.
Authors: Yi Zhao, Siqi Wang, Jing Li
Abstract: Navigation instruction generation for visually impaired (VI) individuals (NIG-VI) is critical yet relatively underexplored. This study focuses on generating precise, in-situ, step-by-step navigation instructions that are practically usable for VI users. Specifically, we propose LaF-GRPO (LLM-as-Follower GRPO), where an LLM simulates VI user responses to navigation instructions, thereby providing feedback rewards to guide the post-training of a Vision-Language Model (VLM). This enhances instruction accuracy and usability while reducing costly real-world data collection needs. To address the scarcity of dedicated benchmarks in this field, we introduce NIG4VI, a 27k-sample open-source dataset to facilitate training and evaluation. It comprises diverse navigation scenarios with accurate spatial coordinates, supporting detailed and open-ended in-situ instruction generation. Experiments on NIG4VI demonstrate the effectiveness of LaF-GRPO through quantitative metrics (e.g., Zero-(LaF-GRPO) boosts BLEU 14\%; SFT+(LaF-GRPO) METEOR 0.542 vs. GPT-4o 0.323), and qualitative analysis further confirms that our method yields more intuitive and safer instructions.
Authors: Renren Jin, Tianhao Shen, Xinwei Wu, Dan Shi, Haoran Sun, Yuqi Ren, Wuwei Huang, Quandong Wang, Wei Liu, Jian Luan, Bin Wang, Deyi Xiong
Abstract: Conducting supervised fine-tuning and preference fine-tuning on large language models (LLMs) requires high-quality datasets to improve their ability to follow instructions and align with human preferences and values. However, constructing such datasets is resource-intensive, and most available datasets for supervised and preference fine-tuning are in English. To address these challenges, we propose the \underline{\textbf{Ta}}xonomy-Guided \underline{\textbf{P}}reference Data Generation (TaP) framework, which facilitates automated and scalable construction of preference datasets across various languages. TaP is grounded in a structured taxonomy that allows fine-grained control over dataset composition, thereby ensuring both diversity and comprehensive coverage. We employ TaP-generated datasets to perform supervised and preference fine-tuning on various LLMs. Experimental results demonstrate that LLMs trained on TaP-generated datasets outperform those trained on existing open-source datasets. Remarkably, LLMs trained on TaP-generated datasets surpass the performance of those trained on an open-source dataset that is 180 times larger.
Authors: Benoit Dherin, Michael Munn, Hanna Mazzawi, Michael Wunder, Javier Gonzalvo
Abstract: One of the most striking features of Large Language Models (LLMs) is their ability to learn in-context. Namely at inference time an LLM is able to learn new patterns without any additional weight update when these patterns are presented in the form of examples in the prompt, even if these patterns were not seen during training. The mechanisms through which this can happen are still largely unknown. In this work, we show that the stacking of a self-attention layer with an MLP, allows the transformer block to implicitly modify the weights of the MLP layer according to the context. We argue through theory and experimentation that this simple mechanism may be the reason why LLMs can learn in-context and not only during training. Specifically, we show how a transformer block implicitly transforms a context into a low-rank weight-update of its MLP layer.
Authors: Egor Fadeev, Dzhambulat Mollaev, Aleksei Shestov, Omar Zoloev, Artem Sakhno, Dmitry Korolev, Ivan Kireev, Andrey Savchenko, Maksim Makarenko
Abstract: Learning clients embeddings from sequences of their historic communications is central to financial applications. While large language models (LLMs) offer general world knowledge, their direct use on long event sequences is computationally expensive and impractical in real-world pipelines. In this paper, we propose LATTE, a contrastive learning framework that aligns raw event embeddings with semantic embeddings from frozen LLMs. Behavioral features are summarized into short prompts, embedded by the LLM, and used as supervision via contrastive loss. The proposed approach significantly reduces inference cost and input size compared to conventional processing of complete sequence by LLM. We experimentally show that our method outperforms state-of-the-art techniques for learning event sequence representations on real-world financial datasets while remaining deployable in latency-sensitive environments.
Authors: Xiaoyi Wang, Jiwei Zhang, Guangtao Zhang, Honglei Guo
Abstract: Synthetic therapy dialogues generated by large language models (LLMs) are increasingly used in mental health NLP to simulate counseling scenarios, train models, and supplement limited real-world data. However, it remains unclear whether these synthetic conversations capture the nuanced emotional dynamics of real therapy. In this work, we introduce RealCBT, a dataset of authentic cognitive behavioral therapy (CBT) dialogues, and conduct the first comparative analysis of emotional arcs between real and LLM-generated CBT sessions. We adapt the Utterance Emotion Dynamics framework to analyze fine-grained affective trajectories across valence, arousal, and dominance dimensions. Our analysis spans both full dialogues and individual speaker roles (counselor and client), using real sessions from the RealCBT dataset and synthetic dialogues from the CACTUS dataset. We find that while synthetic dialogues are fluent and structurally coherent, they diverge from real conversations in key emotional properties: real sessions exhibit greater emotional variability, more emotion-laden language, and more authentic patterns of reactivity and regulation. Moreover, emotional arc similarity remains low across all pairings, with especially weak alignment between real and synthetic speakers. These findings underscore the limitations of current LLM-generated therapy data and highlight the importance of emotional fidelity in mental health applications. To support future research, our dataset RealCBT is released at https://gitlab.com/xiaoyi.wang/realcbt-dataset.
Authors: Helene Tenzer, Oumnia Abidi, Stefan Feuerriegel
Abstract: Large language models (LLMs) are increasingly used in everyday communication, including multilingual interactions across different cultural contexts. While LLMs can now generate near-perfect literal translations, it remains unclear whether LLMs support culturally appropriate communication. In this paper, we analyze the cultural sensitivity of different LLM designs when applied to English-Japanese translations of workplace e-mails. Here, we vary the prompting strategies: (1) naive "just translate" prompts, (2) audience-targeted prompts specifying the recipient's cultural background, and (3) instructional prompts with explicit guidance on Japanese communication norms. Using a mixed-methods study, we then analyze culture-specific language patterns to evaluate how well translations adapt to cultural norms. Further, we examine the appropriateness of the tone of the translations as perceived by native speakers. We find that culturally-tailored prompting can improve cultural fit, based on which we offer recommendations for designing culturally inclusive LLMs in multilingual settings.
Authors: Haewon Park, Sangwoo Kim, Yohan Jo
Abstract: Knowledge editing aims to efficiently correct factual errors in Language Models (LMs). The popular locate-then-edit approach modifies an MLP layer by finding an optimal mapping between its input vector (key) and output vector (value) that leads to the expression of the edited knowledge. However, existing methods without any constraints on the key and value vectors cause significant perturbations to the edited model. To address this, we propose Subspace Knowledge Edit (SUIT), a method that identifies and modifies only the subspace of critical features relevant to the edit. Our empirical results on LLaMA-3-8B, GPT-J-6B, and Qwen2.5-7B models show that SUIT dramatically improves knowledge preservation over strong baselines while maintaining high edit efficacy. This effectiveness confirms that SUIT successfully identifies the critical subspace for the edit. Further analyses provide additional validation for our approach. The source code and data will be released to the public upon publication of the paper.
Authors: Xinye Zhao, Spyridon Mastorakis
Abstract: As large language models (LLMs) continue to scale, the memory footprint of key-value (KV) caches during inference has become a significant bottleneck. Existing approaches primarily focus on compressing KV caches within a single prompt or reusing shared prefixes or frequently ocurred text segments across prompts. However, such strategies are limited in scenarios where prompts are semantically similar but lexically different, which frequently occurs in tasks such as multi-document summarization and conversational agents. We propose \textit{SemShareKV}, a KV cache sharing and compression framework that accelerates LLM inference by reusing KVCache in semantically similar prompts. Instead of relying on exact token matches, SemShareKV applies fuzzy token matching using locality-sensitive hashing (LSH) on token embeddings and incorporates Rotary Position Embedding (RoPE) to better preserve positional information. By selectively reusing relevant key-value pairs from a reference prompt's cache, SemShareKV reduces redundant computation while maintaining output quality. Experiments on diverse summarization datasets show up to 6.25$\times$ speedup and 42\% lower GPU memory usage with 5k tokens input, with negligible quality degradation. These results highlight the potential of semantic-aware cache sharing for efficient LLM inference.
Authors: Wengao Ye, Yan Liang, Lianlei Shan
Abstract: Recent advancements in Large Language Models (LLMs) have shifted from explicit Chain-of-Thought (CoT) reasoning to more efficient latent reasoning, where intermediate thoughts are represented as vectors rather than text. However, latent reasoning can be brittle on challenging, out-of-distribution tasks where robust reasoning is most critical. To overcome these limitations, we introduce Latent Thought Policy Optimization (LTPO), a parameter-free framework that enhances LLM reasoning entirely at test time, without requiring model parameter updates. LTPO treats intermediate latent "thought" vectors as dynamic parameters that are actively optimized for each problem instance. It employs an online policy gradient method guided by an intrinsic, confidence-based reward signal computed directly from the frozen LLM's own output distributions, eliminating the need for external supervision or expensive text generation during optimization. Extensive experiments on five reasoning benchmarks show that LTPO not only matches or surpasses strong baselines on standard tasks but also demonstrates remarkable robustness where others fail. Most notably, on highly challenging AIME benchmarks where existing latent reasoning baselines collapse to near-zero accuracy, LTPO delivers substantial improvements, showcasing a unique capability for complex reasoning.
Authors: Yunhao Fang, Weihao Yu, Shu Zhong, Qinghao Ye, Xuehan Xiong, Lai Wei
Abstract: Long-sequence modeling faces a fundamental trade-off between the efficiency of compressive fixed-size memory in RNN-like models and the fidelity of lossless growing memory in attention-based Transformers. Inspired by the Multi-Store Model in cognitive science, we introduce a memory framework of artificial neural networks. Our method maintains a sliding window of the Transformer's KV cache as lossless short-term memory, while a learnable module termed Artificial Hippocampus Network (AHN) recurrently compresses out-of-window information into a fixed-size compact long-term memory. To validate this framework, we instantiate AHNs using modern RNN-like architectures, including Mamba2, DeltaNet, and GatedDeltaNet to augment open-weight LLMs. We also propose an efficient self-distillation training method where the base model's all parameters are frozen and only the parameters from AHNs are optimized. For inference, our method sets a default large sliding window size of 32k for attention, and AHNs activate only when the sequence length exceeds the 32k window, addressing the quadratic-complexity issue of attention that emerges at that scale. Extensive experiments on long-context benchmarks LV-Eval and InfiniteBench demonstrate that AHN-augmented models consistently outperform sliding window baselines and achieve performance comparable or even superior to full-attention models, while substantially reducing computational and memory requirements. For instance, augmenting the Qwen2.5-3B-Instruct with AHNs reduces inference FLOPs by 40.5% and memory cache by 74.0%, while improving its average score on LV-Eval (128k sequence length) from 4.41 to 5.88. Code is available at: https://github.com/ByteDance-Seed/AHN.
Authors: Xiuyuan Chen, Tao Sun, Dexin Su, Ailing Yu, Junwei Liu, Zhe Chen, Gangzeng Jin, Xin Wang, Jingnan Liu, Hansong Xiao, Hualei Zhou, Dongjie Tao, Chunxiao Guo, Minghui Yang, Yuan Xia, Jing Zhao, Qianrui Fan, Yanyun Wang, Shuai Zhen, Kezhong Chen, Jun Wang, Zewen Sun, Heng Zhao, Tian Guan, Shaodong Wang, Geyun Chang, Jiaming Deng, Hongchengcheng Chen, Kexin Feng, Ruzhen Li, Jiayi Geng, Changtai Zhao, Jun Wang, Guihu Lin, Peihao Li, Liqi Liu, Peng Wei, Jian Wang, Jinjie Gu, Ping Wang, Fan Yang
Abstract: Current benchmarks for AI clinician systems, often based on multiple-choice exams or manual rubrics, fail to capture the depth, robustness, and safety required for real-world clinical practice. To address this, we introduce the GAPS framework, a multidimensional paradigm for evaluating Grounding (cognitive depth), Adequacy (answer completeness), Perturbation (robustness), and Safety. Critically, we developed a fully automated, guideline-anchored pipeline to construct a GAPS-aligned benchmark end-to-end, overcoming the scalability and subjectivity limitations of prior work. Our pipeline assembles an evidence neighborhood, creates dual graph and tree representations, and automatically generates questions across G-levels. Rubrics are synthesized by a DeepResearch agent that mimics GRADE-consistent, PICO-driven evidence review in a ReAct loop. Scoring is performed by an ensemble of large language model (LLM) judges. Validation confirmed our automated questions are high-quality and align with clinician judgment (90% agreement, Cohen's Kappa 0.77). Evaluating state-of-the-art models on the benchmark revealed key failure modes: performance degrades sharply with increased reasoning depth (G-axis), models struggle with answer completeness (A-axis), and they are highly vulnerable to adversarial perturbations (P-axis) as well as certain safety issues (S-axis). This automated, clinically-grounded approach provides a reproducible and scalable method for rigorously evaluating AI clinician systems and guiding their development toward safer, more reliable clinical practice. The benchmark dataset GAPS-NSCLC-preview and evaluation code are publicly available at https://huggingface.co/datasets/AQ-MedAI/GAPS-NSCLC-preview and https://github.com/AQ-MedAI/MedicalAiBenchEval.
URLs: https://huggingface.co/datasets/AQ-MedAI/GAPS-NSCLC-preview, https://github.com/AQ-MedAI/MedicalAiBenchEval.
Authors: Zhiyang Chen, Daliang Xu, Haiyang Shen, Chiheng Lou, Mengwei Xu, Shangguang Wang, Xin Jin, Yun Ma
Abstract: Performing Retrieval-Augmented Generation (RAG) directly on mobile devices is promising for data privacy and responsiveness but is hindered by the architectural constraints of mobile NPUs. Specifically, current hardware struggles with the variable workloads intrinsic to RAG: the transition between processing extensive contexts and generating tokens incurs significant overhead due to static graph constraints, while the memory-bound generation phase leaves computational resources underutilized. In this work, we propose a holistic acceleration framework sd.npu, designed to maximize NPU efficiency for on-device RAG ecosystem. To address the latency caused by NPU graph switching during phase transitions, we introduce a pipelined execution strategy. This approach masks the overhead of model reconfiguration by parallelizing the loading of decoding graphs with the computation of partitioned context chunks (chunked prefill), thereby ensuring continuous execution flow. Furthermore, to mitigate low hardware utilization during the decoding phase, we develop an NPU-centric speculative decoding mechanism. By calibrating generation distributions and extending draft sequences, our method effectively converts idle NPU cycles into valid token throughput. Experiments on commercial smartphones show that our framework significantly outperforms existing baselines, delivering 1.06$\times$--3.81$\times$ speedups and 1.07$\times$--4.71$\times$ energy savings across various RAG tasks.
Authors: Kailin Jiang, Ning Jiang, Yuntao Du, Yuchen Ren, Yuchen Li, Yifan Gao, Jinhe Bi, Yunpu Ma, Qingqing Liu, Xianhao Wang, Yifan Jia, Hongbo Jiang, Yaocong Hu, Bin Li, Lei Liu
Abstract: Large Multimodal Models (LMMs) encode rich factual knowledge via cross-modal pre-training, yet their static representations struggle to maintain an accurate understanding of time-sensitive factual knowledge. Existing benchmarks remain constrained by static designs, inadequately evaluating LMMs' ability to understand time-sensitive knowledge. To address this gap, we propose MINED, a comprehensive benchmark that evaluates temporal awareness along 6 key dimensions and 11 challenging tasks: cognition, awareness, trustworthiness, understanding, reasoning, and robustness. MINED is constructed from Wikipedia by two professional annotators, containing 2,104 time-sensitive knowledge samples spanning six knowledge types. Evaluating 15 widely used LMMs on MINED shows that Gemini-2.5-Pro achieves the highest average CEM score of 63.07, while most open-source LMMs still lack time understanding ability. Meanwhile, LMMs perform best on organization knowledge, whereas their performance is weakest on sport. To address these challenges, we investigate the feasibility of updating time-sensitive knowledge in LMMs through knowledge editing methods and observe that LMMs can effectively update knowledge via knowledge editing methods in single editing scenarios.
Authors: Xiying Zhao, Zhoufutu Wen, Zhixuan Chen, Jingzhe Ding, Jianpeng Jiao, Shuai Li, Xi Li, Danni Liang, Shengda Long, Qianqian Liu, Xianbo Wu, Hongwan Gao, Xiang Gao, Liang Hu, Jiashuo Liu, Mengyun Liu, Weiran Shi, Chenghao Yang, Qianyu Yang, Xuanliang Zhang, Ge Zhang, Wenhao Huang, Yuwen Tang
Abstract: The evaluation of discourse-level translation in expert domains remains inadequate, despite its centrality to knowledge dissemination and cross-lingual scholarly communication. While these translations demand discourse-level coherence and strict terminological precision, current evaluation methods predominantly focus on segment-level accuracy and fluency. To address this limitation, we introduce DiscoX, a new benchmark for discourse-level and expert-level Chinese-English translation. It comprises 200 professionally-curated texts from 7 domains, with an average length exceeding 1700 tokens. To evaluate performance on DiscoX, we also develop Metric-S, a reference-free system that provides fine-grained automatic assessments across accuracy, fluency, and appropriateness. Metric-S demonstrates strong consistency with human judgments, significantly outperforming existing metrics. Our experiments reveal a remarkable performance gap: even the most advanced LLMs still trail human experts on these tasks. This finding validates the difficulty of DiscoX and underscores the challenges that remain in achieving professional-grade machine translation. The proposed benchmark and evaluation system provide a robust framework for more rigorous evaluation, facilitating future advancements in LLM-based translation.
Authors: Zhaodong Wang, Zhenting Qi, Sherman Wong, Nathan Hu, Samuel Lin, Jun Ge, Erwin Gao, Wenlin Chen, Yilun Du, Minlan Yu, Ying Zhang
Abstract: Real-world software engineering tasks require coding agents that can operate over massive repositories, sustain long-horizon sessions, and reliably coordinate complex toolchains at test time. Existing research-grade agents offer transparency but struggle when scaled to real-world workloads, while proprietary systems achieve strong practical performance but provide limited extensibility, interpretability, and controllability. We introduce the Confucius Code Agent (CCA), a scalable software engineering agent that can operate at large-scale codebases. CCA is built on top of the Confucius SDK, an agent development platform structured around three complementary perspectives: Agent Experience (AX), User Experience (UX), and Developer Experience (DX). The SDK integrates a unified orchestrator with hierarchical working memory for long-context reasoning, a persistent note-taking system for cross-session continual learning, and a modular extension system for reliable tool use. In addition, we introduce a meta-agent that automates the synthesis, evaluation, and refinement of agent configurations through a build-test-improve loop, enabling rapid adaptation to new tasks, environments, and tool stacks. Instantiated with these mechanisms, CCA demonstrates strong performance on real-world software engineering tasks. On SWE-Bench-Pro, CCA reaches a Resolve@1 of 54.3%, exceeding prior research baselines and comparing favorably to commercial results, under identical repositories, model backend, and tool access. Together, the Confucius SDK and CCA form a general, extensible, and production-grade foundation for building effective and robust coding agents, bridging the gap between research prototypes and practical large-scale deployment.
Authors: Youmin Ko, Sungjong Seo, Hyunjoon Kim
Abstract: Since large language models (LLMs) have a tendency to generate factually inaccurate output, retrieval-augmented generation (RAG) has gained significant attention as a key means to mitigate this downside of harnessing only LLMs. However, existing RAG methods for simple and multi-hop question answering (QA) are still prone to incorrect retrievals and hallucinations. To address these limitations, we propose CoopRAG, a novel RAG framework for the question answering task in which a retriever and an LLM work cooperatively with each other by exchanging informative knowledge, and the earlier and later layers of the retriever model work cooperatively with each other to accurately rank the retrieved documents relevant to a given query. In this framework, we (i) unroll a question into sub-questions and a reasoning chain in which uncertain positions are masked, (ii) retrieve the documents relevant to the question augmented with the sub-questions and the reasoning chain, (iii) rerank the documents by contrasting layers of the retriever, and (iv) reconstruct the reasoning chain by filling the masked positions via the LLM. Our experiments demonstrate that CoopRAG consistently outperforms state-of-the-art QA methods on three multi-hop QA datasets as well as a simple QA dataset in terms of both the retrieval and QA performances. Our code is available.
Authors: Chendong Sun, Ali Mao, Lei Xu, mingmin Chen
Abstract: Speculative Decoding is a prominent technique for accelerating the autoregressive inference of large language models (LLMs) by employing a fast draft model to propose candidate token sequences and a large target model to verify them in parallel. However, its core component -- the rejection sampling mechanism -- relies on a fixed, context-independent random threshold. This leads to a significant "random rejection" problem in high-uncertainty generation scenarios, where plausible candidate tokens are frequently rejected due to random chance, undermining inference efficiency. This paper introduces Efficient Adaptive Rejection Sampling (EARS), a novel method that dynamically adjusts the acceptance threshold by incorporating the target model's own predictive uncertainty, measured as 1 - max(P_target). By introducing a tolerance term proportional to this uncertainty, EARS intelligently relaxes the acceptance criterion when the model is uncertain, effectively reducing random rejections while maintaining strict standards when the model is confident. Experiments on creative writing and open-domain QA tasks demonstrate that EARS significantly enhances the efficiency of speculative decoding, achieving up to an 18.12% increase in throughput with a negligible 0.84% accuracy drop on the GSM8K benchmark. The method requires no modifications to model architectures and can be seamlessly integrated into existing speculative decoding frameworks.
Authors: Nguyen Tien Dong, Minh-Anh Nguyen, Thanh Dat Hoang, Nguyen Tuan Ngoc, Dao Xuan Quang Minh, Phan Phi Hai, Nguyen Thi Ngoc Anh, Dang Van Tu, Binh Vu
Abstract: The rapid advancement of large language models (LLMs) has enabled new possibilities for applying artificial intelligence within the legal domain. Nonetheless, the complexity, hierarchical organization, and frequent revisions of Vietnamese legislation pose considerable challenges for evaluating how well these models interpret and utilize legal knowledge. To address this gap, Vietnamese Legal Benchmark (VLegal-Bench) is introduced, the first comprehensive benchmark designed to systematically assess LLMs on Vietnamese legal tasks. Informed by Bloom's cognitive taxonomy, VLegal-Bench encompasses multiple levels of legal understanding through tasks designed to reflect practical usage scenarios. The benchmark comprises 10,450 samples generated through a rigorous annotation pipeline, where legal experts label and cross-validate each instance using our annotation system to ensure every sample is grounded in authoritative legal documents and mirrors real-world legal assistant workflows, including general legal questions and answers, retrieval-augmented generation, multi-step reasoning, and scenario-based problem solving tailored to Vietnamese law. By providing a standardized, transparent, and cognitively informed evaluation framework, VLegal-Bench establishes a solid foundation for assessing LLM performance in Vietnamese legal contexts and supports the development of more reliable, interpretable, and ethically aligned AI-assisted legal systems.
Authors: Zefan Cai, Haoyi Qiu, Tianyi Ma, Haozhe Zhao, Gengze Zhou, Kung-Hsiang Huang, Parisa Kordjamshidi, Minjia Zhang, Wen Xiao, Jiuxiang Gu, Nanyun Peng, Junjie Hu
Abstract: Video foundation models generate visually realistic and temporally coherent content, but their reliability as world simulators depends on whether they capture physical, logical, and spatial constraints. Existing metrics such as Frechet Video Distance (FVD) emphasize perceptual quality and overlook reasoning failures, including violations of causality, physics, and global consistency. We introduce MMGR (Multi-Modal Generative Reasoning Evaluation and Benchmark), a principled evaluation framework based on five reasoning abilities: Physical, Logical, 3D Spatial, 2D Spatial, and Temporal. MMGR evaluates generative reasoning across three domains: Abstract Reasoning (ARC-AGI, Sudoku), Embodied Navigation (real-world 3D navigation and localization), and Physical Commonsense (sports and compositional interactions). MMGR applies fine-grained metrics that require holistic correctness across both video and image generation. We benchmark leading video models (Veo-3, Sora-2, Wan-2.2) and image models (Nano-banana, Nano-banana Pro, GPT-4o-image, Qwen-image), revealing strong performance gaps across domains. Models show moderate success on Physical Commonsense tasks but perform poorly on Abstract Reasoning (below 10 percent accuracy on ARC-AGI) and struggle with long-horizon spatial planning in embodied settings. Our analysis highlights key limitations in current models, including overreliance on perceptual data, weak global state consistency, and objectives that reward visual plausibility over causal correctness. MMGR offers a unified diagnostic benchmark and a path toward reasoning-aware generative world models.
Authors: Zhixian Zhao, Xinfa Zhu, Xinsheng Wang, Shuiyuan Wang, Xuelong Geng, Wenjie Tian, Lei Xie
Abstract: Large-scale audio language models (ALMs), such as Qwen2-Audio, are capable of comprehending diverse audio signal, performing audio analysis and generating textual responses. However, in speech emotion recognition (SER), ALMs often suffer from hallucinations, resulting in misclassifications or irrelevant outputs. To address these challenges, we propose C$^2$SER, a novel ALM designed to enhance the stability and accuracy of SER through Contextual perception and Chain of Thought (CoT). C$^2$SER integrates the Whisper encoder for semantic perception and Emotion2Vec-S for acoustic perception, where Emotion2Vec-S extends Emotion2Vec with semi-supervised learning to enhance emotional discrimination. Additionally, C$^2$SER employs a CoT approach, processing SER in a step-by-step manner while leveraging speech content and speaking styles to improve recognition. To further enhance stability, C$^2$SER introduces self-distillation from explicit CoT to implicit CoT, mitigating error accumulation and boosting recognition accuracy. Extensive experiments show that C$^2$SER outperforms existing popular ALMs, such as Qwen2-Audio and SECap, delivering more stable and precise emotion recognition. We release the training code, checkpoints, and test sets to facilitate further research.
Authors: Wenbo Hu, Yining Hong, Yanjun Wang, Leison Gao, Zibu Wei, Xingcheng Yao, Nanyun Peng, Yonatan Bitton, Idan Szpektor, Kai-Wei Chang
Abstract: Humans excel at performing complex tasks by leveraging long-term memory across temporal and spatial experiences. In contrast, current Large Language Models (LLMs) struggle to effectively plan and act in dynamic, multi-room 3D environments. We posit that part of this limitation is due to the lack of proper 3D spatial-temporal memory modeling in LLMs. To address this, we first introduce 3DMem-Bench, a comprehensive benchmark comprising over 26,000 trajectories and 2,892 embodied tasks, question-answering and captioning, designed to evaluate an agent's ability to reason over long-term memory in 3D environments. Second, we propose 3DLLM-Mem, a novel dynamic memory management and fusion model for embodied spatial-temporal reasoning and actions in LLMs. Our model uses working memory tokens, which represents current observations, as queries to selectively attend to and fuse the most useful spatial and temporal features from episodic memory, which stores past observations and interactions. Our approach allows the agent to focus on task-relevant information while maintaining memory efficiency in complex, long-horizon environments. Experimental results demonstrate that 3DLLM-Mem achieves state-of-the-art performance across various tasks, outperforming the strongest baselines by 16.5% in success rate on 3DMem-Bench's most challenging in-the-wild embodied tasks.
Authors: Tianteng Gu, Bei Liu, Bo Xiao, Ke Zeng, Jiacheng Liu, Yanmin Qian
Abstract: Pruning is a widely used technique to compress large language models (LLMs) by removing unimportant weights, but it often suffers from significant performance degradation - especially under semi-structured sparsity constraints. Existing pruning methods primarily focus on estimating the importance of individual weights, which limits their ability to preserve critical capabilities of the model. In this work, we propose a new perspective: rather than merely selecting which weights to prune, we first redistribute parameter importance to make the model inherently more amenable to pruning. By minimizing the information entropy of normalized importance scores, our approach concentrates importance onto a smaller subset of weights, thereby enhancing pruning robustness. We instantiate this idea through DenoiseRotator, which applies learnable orthogonal transformations to the model's weight matrices. Our method can be seamlessly integrated with existing pruning techniques such as Magnitude, SparseGPT, and Wanda. Evaluated on LLaMA3, Qwen2.5, and Mistral models under 50% unstructured and 2:4 semi-structured sparsity, DenoiseRotator consistently improves perplexity and zero-shot accuracy. For instance, on LLaMA3-70B pruned with SparseGPT at 2:4 semi-structured sparsity, DenoiseRotator reduces the perplexity gap to the dense model by 58%, narrowing the degradation from 8.1 to 3.4 points. Codes are available at https://github.com/Axel-gu/DenoiseRotator.
Authors: Nishit V. Pandya, Andrey Labunets, Sicun Gao, Earlence Fernandes
Abstract: A popular class of defenses against prompt injection attacks on large language models (LLMs) relies on fine-tuning to separate instructions and data, so that the LLM does not follow instructions that might be present with data. We evaluate the robustness of this approach in the whitebox setting by constructing strong optimization-based attacks, and show that the defenses do not provide the claimed security properties. Specifically, we construct a novel attention-based attack algorithm for textual LLMs and apply it to three recent whitebox defenses SecAlign (CCS 2025), SecAlign++, and StruQ (USENIX Security 2025), showing attacks with success rates of up to \textbf{85-95\%} on unseen prompts with modest increase in attacker budget in terms of tokens. Our findings make fundamental progress towards understanding the robustness of prompt injection defenses in the whitebox setting. We release our code and attacks at https://github.com/nishitvp/better_opts_attacks
Authors: Mingqi Wu, Zhihao Zhang, Qiaole Dong, Zhiheng Xi, Jun Zhao, Senjie Jin, Xiaoran Fan, Yuhao Zhou, Huijie Lv, Ming Zhang, Yanwei Fu, Qin Liu, Songyang Zhang, Qi Zhang
Abstract: Reasoning in large language models has long been a central research focus, and recent studies employing reinforcement learning (RL) have introduced diverse methods that yield substantial performance gains with minimal or even no external supervision. Surprisingly, some studies even suggest that random or incorrect reward signals can enhance performance. However, these breakthroughs are predominantly observed for the mathematically strong Qwen2.5 series on benchmarks such as MATH-500, AMC, and AIME, and seldom transfer to models like Llama, which warrants a more in-depth investigation. In this work, our empirical analysis reveals that pre-training on massive web-scale corpora leaves Qwen2.5 susceptible to data contamination in widely used benchmarks. Consequently, conclusions derived from contaminated benchmarks on Qwen2.5 series may be unreliable. To obtain trustworthy evaluation results, we introduce a generator that creates fully clean arithmetic problems of arbitrary length and difficulty, dubbed RandomCalculation. Using this leakage-free dataset, we show that only accurate reward signals yield steady improvements that surpass the base model's performance boundary in mathematical reasoning, whereas random or incorrect rewards do not. Moreover, we conduct more fine-grained analyses to elucidate the factors underlying the different performance observed on the MATH-500 and RandomCalculation benchmarks. Consequently, we recommend that future studies evaluate models on uncontaminated benchmarks and, when feasible, test various model series to ensure trustworthy conclusions about RL and related methods.
Authors: Pengsong Zhang, Xiang Hu, Guowei Huang, Yang Qi, Heng Zhang, Xiuxu Li, Jiaxing Song, Jiabin Luo, Yijiang Li, Shuo Yin, Chengxiao Dai, Eric Hanchen Jiang, Xiaoyan Zhou, Zhenfei Yin, Boqin Yuan, Jing Dong, Guinan Su, Guanren Qiao, Haiming Tang, Anghong Du, Lili Pan, Zhenzhong Lan, Xinyu Liu
Abstract: Recent advances in large language models (LLMs) have enabled AI agents to autonomously generate scientific proposals, conduct experiments, author papers, and perform peer reviews. Yet this flood of AI-generated research content collides with a fragmented and largely closed publication ecosystem. Traditional journals and conferences rely on human peer review, making them difficult to scale and often reluctant to accept AI-generated research content; existing preprint servers (e.g. arXiv) lack rigorous quality-control mechanisms. Consequently, a significant amount of high-quality AI-generated research lacks appropriate venues for dissemination, hindering its potential to advance scientific progress. To address these challenges, we introduce aiXiv, a next-generation open-access platform for human and AI scientists. Its multi-agent architecture allows research proposals and papers to be submitted, reviewed, and iteratively refined by both human and AI scientists. It also provides API and MCP interfaces that enable seamless integration of heterogeneous human and AI scientists, creating a scalable and extensible ecosystem for autonomous scientific discovery. Through extensive experiments, we demonstrate that aiXiv is a reliable and robust platform that significantly enhances the quality of AI-generated research proposals and papers after iterative revising and reviewing on aiXiv. Our work lays the groundwork for a next-generation open-access ecosystem for AI scientists, accelerating the publication and dissemination of high-quality AI-generated research content. Code: https://github.com/aixiv-org aiXiv: https://aixiv.science
Authors: Tittaya Mairittha, Tanakon Sawanglok, Panuwit Raden, Jirapast Buntub, Thanapat Warunee, Napat Asawachaisuvikrom, Thanaphum Saiwongin
Abstract: While voice-based AI systems have achieved remarkable generative capabilities, their interactions often feel conversationally broken. This paper examines the interactional friction that emerges in modular Speech-to-Speech Retrieval-Augmented Generation (S2S-RAG) pipelines. By analyzing a representative production system, we move beyond simple latency metrics to identify three recurring patterns of conversational breakdown: (1) Temporal Misalignment, where system delays violate user expectations of conversational rhythm; (2) Expressive Flattening, where the loss of paralinguistic cues leads to literal, inappropriate responses; and (3) Repair Rigidity, where architectural gating prevents users from correcting errors in real-time. Through system-level analysis, we demonstrate that these friction points should not be understood as defects or failures, but as structural consequences of a modular design that prioritizes control over fluidity. We conclude that building natural spoken AI is an infrastructure design challenge, requiring a shift from optimizing isolated components to carefully choreographing the seams between them.
Authors: Kamer Ali Yuksel
Abstract: Large language models (LLMs) are increasingly used to evolve programs and multi-agent systems, yet most existing approaches rely on overwrite-based mutations that maintain only a single candidate at a time. Such methods discard useful variants, suffer from destructive edits, and explore a brittle search space prone to structural failure. We introduce EvoLattice, a framework that represents an entire population of candidate programs or agent behaviors within a single directed acyclic graph. Each node stores multiple persistent alternatives, and every valid path through the graph defines a distinct executable candidate, yielding a large combinatorial search space without duplicating structure. EvoLattice enables fine-grained alternative-level evaluation by scoring each alternative across all paths in which it appears, producing statistics that reveal how local design choices affect global performance. These statistics provide a dense, data-driven feedback signal for LLM-guided mutation, recombination, and pruning, while preserving successful components. Structural correctness is guaranteed by a deterministic self-repair mechanism that enforces acyclicity and dependency consistency independently of the LLM. EvoLattice naturally extends to agent evolution by interpreting alternatives as prompt fragments or sub-agent behaviors. Across program synthesis (proxy and optimizer meta-learning), EvoLattice yields more stable evolution, greater expressivity, and stronger improvement trajectories than prior LLM-guided methods. The resulting dynamics resemble quality-diversity optimization, emerging implicitly from EvoLattice's internal multi-alternative representation rather than an explicit external archive.