Authors: Berkin Durmus, Chen Cen, Eduardo Pacheco, Arda Okan, Atila Orhon
Abstract: The accuracy frontier of speech-to-text systems has plateaued on academic benchmarks.1 In contrast, industrial benchmarks and adoption in high-stakes domains suggest otherwise. We hypothesize that the primary difference between the two is contextual conditioning: Academic benchmarks are dominated by frequently encountered general vocabulary that is relatively easy to recognize compared with rare and context-defined custom vocabulary that has disproportionate impact on the usability of speech transcripts. Despite progress on contextual speech-to-text, there is no standardized benchmark. We introduce Contextual Earnings-22, an open dataset built upon Earnings-22, with realistic custom vocabulary contexts to foster research and reveal latent progress. We set six strong baselines for two dominant approaches: keyword prompting and keyword boosting. Experiments show both reach comparable and significantly improved accuracy when scaled from proof-of-concept to large-scale systems.
Authors: Youcef Soufiane Gheffari, Oussama Mustapha Benouddane, Samiya Silarbi
Abstract: Recognizing emotions from speech using machine learning has become an active research area due to its importance in building human-centered applications. However, while many studies have been conducted in English, German, and other European and Asian languages, research in Arabic remains scarce because of the limited availability of annotated datasets. In this paper, we present an Arabic Speech Emotion Recognition (SER) system based on a hybrid CNN-Transformer architecture. The model leverages convolutional layers to extract discriminative spectral features from Mel-spectrogram inputs and Transformer encoders to capture long-range temporal dependencies in speech. Experiments were conducted on the EYASE (Egyptian Arabic speech emotion) corpus, and the proposed model achieved 97.8% accuracy and a macro F1-score of 0.98. These results demonstrate the effectiveness of combining convolutional feature extraction with attention-based modeling for Arabic SER and highlight the potential of Transformer-based approaches in low-resource languages.
Authors: Avyav Kumar Singh, Yen-Chen Wu, Alexandru Cioba, Alberto Bernacchia, Davide Buffelli
Abstract: Cross-tokenizer distillation (CTD), the transfer of knowledge from a teacher to a student language model when the two use different tokenizers, remains a largely unsolved problem. Existing approaches rely on heuristic strategies to align mismatched vocabularies, introducing considerable complexity. In this paper, we propose a simple but effective baseline called Byte-Level Distillation (BLD) which enables CTD by operating at a common interface across tokenizers: the byte level. In more detail, we convert the teacher's output distribution to byte-level probabilities, attach a lightweight byte-level decoder head to the student, and distill through this shared byte-level interface. Despite its simplicity, BLD performs competitively with--and on several benchmarks surpasses--significantly more sophisticated CTD methods, across a range of distillation tasks with models from 1B to 8B parameters. Our results suggest that the byte level is a natural common ground for cross-tokenizer knowledge transfer, while also highlighting that consistent improvements across all tasks and benchmarks remain elusive, underscoring that CTD is still an open problem.
Authors: Opeyemi Osakuade, Simon King
Abstract: Discrete speech units (DSUs) are derived by quantising representations from models trained using self-supervised learning (SSL). They are a popular representation for a wide variety of spoken language tasks, including those where prosody matters. DSUs are especially convenient for tasks where text and speech are jointly modelled, such as text-to-speech and multimodal dialogue systems. But we have found that DSUs encode suprasegmental information less reliably than segmental structure, which we demonstrate in this work using lexical tone, though this limitation likely extends to other suprasegmental features such as prosody. Our investigations using the tone languages Mandarin and Yor\`ub\'a show that the SSL latent representations themselves do encode tone, yet DSUs obtained using quantisation tend to prioritise phonetic structure, which makes lexical tone less reliably encoded. This remains true for a variety of quantisation methods, not only the most common, K-means. We conclude that current DSU quantisation strategies have limitations for suprasegmental features, which suggests a need for new, tone-aware (or prosody-aware) techniques in speech representation learning. We point towards a potential form of the solution by performing K-means clustering once to encode phonetic information, then again on the residual representation, which better encodes lexical tone.
Authors: Xuechen Zhang, Aviv Slobodkin, Joydeep Paul, Mandar Sharma, Samet Oymak, Shravya Shetty, Gautam Prasad
Abstract: Representation learning for geospatial and spatio-temporal data plays a critical role in enabling general-purpose geospatial intelligence. Recent geospatial foundation models, such as the Population Dynamics Foundation Model (PDFM), encode complex population and mobility dynamics into compact embeddings. However, their integration with Large Language Models (LLMs) remains limited. Existing approaches to LLM integration treat these embeddings as retrieval indices or convert them into textual descriptions for reasoning, introducing redundancy, token inefficiency, and numerical inaccuracies. We propose Direct Feature Reasoning-Gemma (DFR-Gemma), a novel framework that enables LLMs to reason directly over dense geospatial embeddings. DFR aligns high-dimensional embeddings with the latent space of an LLM via a lightweight projector, allowing embeddings to be injected as semantic tokens alongside natural language instructions. This design eliminates the need for intermediate textual representations and enables intrinsic reasoning over spatial features. To evaluate this paradigm, we introduce a multi-task geospatial benchmark that pairs embeddings with diverse question-answer tasks, including feature querying, comparison, and semantic description. Experimental results show that DFR allows LLMs to decode latent spatial patterns and perform accurate zero-shot reasoning across tasks, while significantly improving efficiency compared to text-based baselines. Our results demonstrate that treating embeddings as primary data inputs, provides a more direct, efficient, and scalable approach to multimodal geospatial intelligence.
Authors: Mengdan Zhu, Senhao Cheng, Liang Zhao
Abstract: Vision-Language Models often struggle with complex visual reasoning due to the visual information loss in textual CoT. Existing methods either add the cost of tool calls or rely on localized patch-based embeddings that are insufficient to extract semantics in multi-step reasoning. We propose \emph{"Decompose, Look, and Reason" (DLR)}, a reinforced latent reasoning framework that dynamically decomposes queries into textual premises, extracts premise-conditioned continuous visual latents, and deduces answers through grounded rationales. We introduce a three-stage training pipeline and propose a novel Spherical Gaussian Latent Policy to enable effective exploration in the latent space. Extensive experiments on vision-centric benchmarks show that DLR consistently outperforms strong baselines, including text-only, interleaved multimodal CoT, and latent reasoning methods, while providing superior stepwise interpretability.
Authors: Xueren Ge, Sahil Murtaza, Anthony Cortez, Homa Alemzadeh
Abstract: Conversational diagnosis prediction requires models to track evolving evidence in streaming clinical conversations and decide when to commit to a diagnosis. Existing medical dialogue corpora are largely dyadic or lack the multi-party workflow and annotations needed for this setting. We introduce an ePCR-grounded, topic-flow-based multi-agent generation pipeline that iteratively plans, generates, and self-refines dialogues with rule-based factual and topic flow checks. The pipeline yields EMSDialog, a dataset of 4,414 synthetic multi-speaker EMS conversations based on a real-world ePCR dataset, annotated with 43 diagnoses, speaker roles, and turn-level topics. Human and LLM evaluations confirm high quality and realism of EMSDialog using both utterance- and conversation-level metrics. Results show that EMSDialog-augmented training improves accuracy, timeliness, and stability of EMS conversational diagnosis prediction.
Authors: Figen E\u{g}in, Aytu\u{g} Onan
Abstract: This study presents a framework for generating the gold-standard summary fully automatically and reproducibly based on multiple human summaries of Turkish educational videos. Within the scope of the study, a new dataset called TR-EduVSum was created, encompassing 82 Turkish course videos in the field of "Data Structures and Algorithms" and containing a total of 3281 independent human summaries. Inspired by existing pyramid-based evaluation approaches, the AutoMUP (Automatic Meaning Unit Pyramid) method is proposed, which extracts consensus-based content from multiple human summaries. AutoMUP clusters the meaning units extracted from human summaries using embedding, statistically models inter-participant agreement, and generates graded summaries based on consensus weight. In this framework, the gold summary corresponds to the highest-consensus AutoMUP configuration, constructed from the most frequently supported meaning units across human summaries. Experimental results show that AutoMUP summaries exhibit high semantic overlap with robust LLM (Large Language Model) summaries such as Flash 2.5 and GPT-5.1. Furthermore, ablation studies clearly demonstrate the decisive role of consensus weight and clustering in determining summary quality. The proposed approach can be generalized to other Turkic languages at low cost.
Authors: Tunazzina Islam
Abstract: Unsupervised methods are widely used to induce latent semantic structure from large text collections, yet their outputs often contain incoherent, redundant, or poorly grounded clusters that are difficult to validate without labeled data. We propose a reasoning-based refinement framework that leverages large language models (LLMs) not as embedding generators, but as semantic judges that validate and restructure the outputs of arbitrary unsupervised clustering algorithms.Our framework introduces three reasoning stages: (i) coherence verification, where LLMs assess whether cluster summaries are supported by their member texts; (ii) redundancy adjudication, where candidate clusters are merged or rejected based on semantic overlap; and (iii) label grounding, where clusters are assigned interpretable labels in a fully unsupervised manner. This design decouples representation learning from structural validation and mitigates common failure modes of embedding-only approaches. We evaluate the framework on real-world social media corpora from two platforms with distinct interaction models, demonstrating consistent improvements in cluster coherence and human-aligned labeling quality over classical topic models and recent representation-based baselines. Human evaluation shows strong agreement with LLM-generated labels, despite the absence of gold-standard annotations. We further conduct robustness analyses under matched temporal and volume conditions to assess cross-platform stability. Beyond empirical gains, our results suggest that LLM-based reasoning can serve as a general mechanism for validating and refining unsupervised semantic structure, enabling more reliable and interpretable analyses of large text collections without supervision.
Authors: Mohamed Ehab (Faculty of Computer Science, October University for Modern Science & Arts, Giza, Egypt), Ali Hamdi (Faculty of Computer Science, October University for Modern Science & Arts, Giza, Egypt), Khaled Shaban (Department of Computer Science and Engineering, Qatar University, Doha, Qatar)
Abstract: Real-world categorization is severely hampered by class imbalance because traditional ensembles favor majority classes, which lowers minority performance and overall F1-score. We provide a unique ensemble technique for imbalanced problems called CAMO (Class-Aware Minority-Optimized).Through a hierarchical procedure that incorporates vote distributions, confidence calibration, and inter model uncertainty, CAMO dynamically boosts underrepresented classes while preserving and amplifying minority forecasts.We verify CAMO on two highly unbalanced, domain-specific benchmarks: the DIAR-AI/Emotion dataset and the ternary BEA 2025 dataset. We benchmark against seven proven ensemble algorithms using eight different language models (three LLMs and five SLMs) under zero-shot and fine-tuned settings .With refined models, CAMO consistently earns the greatest strict macro F1-score, setting a new benchmark. Its benefit works in concert with model adaptation, showing that the best ensemble choice depends on model properties .This proves that CAMO is a reliable, domain-neutral framework for unbalanced categorization.
Authors: Aryaman Arora, Zhengxuan Wu, Jacob Steinhardt, Sarah Schwettmann
Abstract: In language model interpretability research, \textbf{circuit tracing} aims to identify which internal features causally contributed to a particular output and how they affected each other, with the goal of explaining the computations underlying some behaviour. However, all prior circuit tracing work has relied on ad-hoc human interpretation of the role that each feature in the circuit plays, via manual inspection of data artifacts such as the dataset examples the component activates on. We introduce \textbf{ADAG}, an end-to-end pipeline for describing these attribution graphs which is fully automated. To achieve this, we introduce \textit{attribution profiles} which quantify the functional role of a feature via its input and output gradient effects. We then introduce a novel clustering algorithm for grouping features, and an LLM explainer--simulator setup which generates and scores natural-language explanations of the functional role of these feature groups. We run our system on known human-analysed circuit-tracing tasks and recover interpretable circuits, and further show ADAG can find steerable clusters which are responsible for a harmful advice jailbreak in Llama 3.1 8B Instruct.
Authors: Ziyi Wang, Siva Rajesh Kasa, Ankith M S, Santhosh Kumar Kasa, Jiaru Zou, Sumit Negi, Ruqi Zhang, Nan Jiang, Qifan Song
Abstract: Speculative decoding is an effective technique for accelerating large language model inference by drafting multiple tokens in parallel. In practice, its speedup is often bottlenecked by a rigid verification step that strictly enforces the accepted token distribution to exactly match the target model. This constraint leads to the rejection of many plausible tokens, lowering the acceptance rate and limiting overall time speedup. To overcome this limitation, we propose Dynamic Verification Relaxed Speculative Decoding (DIVERSED), a relaxed verification framework that improves time efficiency while preserving generation quality. DIVERSED learns an ensemble-based verifier that blends the draft and target model distributions with a task-dependent and context-dependent weight. We provide theoretical justification for our approach and demonstrate empirically that DIVERSED achieves substantially higher inference efficiency compared to standard speculative decoding methods. Code is available at: https://github.com/comeusr/diversed.
Authors: Mingchen Li, Jiatan Huang, Zonghai Yao, Hong yu
Abstract: Large language models (LLMs) hold significant promise for healthcare, yet their reliability in high-stakes clinical settings is often compromised by hallucinations and a lack of granular medical context. While Retrieval Augmented Generation (RAG) can mitigate these issues, standard supervised pipelines require computationally intensive searches over massive external knowledge bases, leading to high latency that is impractical for time-sensitive care. To address this, we introduce Keys to Knowledge (K2K), a novel framework that replaces external retrieval with internal, key-based knowledge access. By encoding essential clinical information directly into the model's parameter space, K2K enables rapid retrieval from internal key-value memory without inference-time overhead. We further enhance retrieval quality through activation-guided probe construction and cross-attention reranking. Experimental results demonstrate that K2K achieves state-of-the-art performance across four benchmark healthcare outcome prediction datasets.
Authors: Ziyi Chen, Yasir Khan, Mengyuan Zhang, Cheng Peng, Mengxian Lyu, Yiyang Liu, Krishna Vaddiparti, Robert L Cook, Mattia Prosperi, Yonghui Wu
Abstract: Human immunodeficiency virus (HIV)-related stigma is a critical psychosocial determinant of health for people living with HIV (PLWH), influencing mental health, engagement in care, and treatment outcomes. Although stigma-related experiences are documented in clinical narratives, there is a lack of off-the-shelf tools to extract and categorize them. This study aims to develop a large language model (LLM)-based tool for identifying HIV stigma from clinical notes. We identified clinical notes from PLWH receiving care at the University of Florida (UF) Health between 2012 and 2022. Candidate sentences were identified using expert-curated stigma-related keywords and iteratively expanded via clinical word embeddings. A total of 1,332 sentences were manually annotated across four stigma subscales: Concern with Public Attitudes, Disclosure Concerns, Negative Self-Image, and Personalized Stigma. We compared GatorTron-large and BERT as encoder-based baselines, and GPT-OSS-20B, LLaMA-8B, and MedGemma-27B as generative LLMs, under zero-shot and few-shot prompting. GatorTron-large achieved the best overall performance (Micro F1 = 0.62). Few-shot prompting substantially improved generative model performance, with 5-shot GPT-OSS-20B and LLaMA-8B achieving Micro-F1 scores of 0.57 and 0.59, respectively. Performance varied by stigma subscale, with Negative Self-Image showing the highest predictability and Personalized Stigma remaining the most challenging. Zero-shot generative inference exhibited non-trivial failure rates (up to 32%). This study develops the first practical NLP tool for identifying HIV stigma in clinical notes.
Authors: Jie Sun, Yu Liu, Lu Han, Qiwen Deng, Xiang Shu, Yang Xiao, Xingyu Lu, Jun Zhou, Pengfei Liu, Lintao Ma, Jiancan Wu, Xiang Wang
Abstract: While transformer-based Large Language Models (LLMs) theoretically support massive context windows, they suffer from severe performance degradation when processing long numerical sequences. We attribute this failure to the attention dispersion in the Softmax mechanism, which prevents the model from concentrating attention. To overcome this, we propose Separate Sequence (SepSeq), a training-free, plug-and-play framework to mitigate dispersion by strategically inserting separator tokens. Mechanistically, we demonstrate that separator tokens act as an attention sink, recalibrating attention to focus on local segments while preserving global context. Extensive evaluations on 9 widely-adopted LLMs confirm the effectiveness of our approach: SepSeq yields an average relative accuracy improvement of 35.6% across diverse domains while reducing total inference token consumption by 16.4% on average.
Authors: Steven Au, Sujit Noronha
Abstract: Large language models (LLMs) can shift their answers under pressure in ways that reflect accommodation rather than reasoning. Prior work on sycophancy has focused mainly on disagreement, flattery, and preference alignment, leaving a broader set of epistemic failures less explored. We introduce \textbf{PPT-Bench}, a diagnostic benchmark for evaluating \textit{epistemic attack}, where prompts challenge the legitimacy of knowledge, values, or identity rather than simply opposing a previous answer. PPT-Bench is organized around the Philosophical Pressure Taxonomy (PPT), which defines four types of philosophical pressure: Epistemic Destabilization, Value Nullification, Authority Inversion, and Identity Dissolution. Each item is tested at three layers: a baseline prompt (L0), a single-turn pressure condition (L1), and a multi-turn Socratic escalation (L2). This allows us to measure epistemic inconsistency between L0 and L1, and conversational capitulation in L2. Across five models, these pressure types produce statistically separable inconsistency patterns, suggesting that epistemic attack exposes weaknesses not captured by standard social-pressure benchmarks. Mitigation results are strongly type- and model-dependent: prompt-level anchoring and persona-stability prompts perform best in API settings, while Leading Query Contrastive Decoding is the most reliable intervention for open models.
Authors: Clarissa Miranda-Pena, Andrew Reeson, C\'ecile Paris, Josiah Poon, Jonathan K. Kummerfeld
Abstract: Despite extensive research, Large Language Models continue to hallucinate when generating code, particularly when using libraries. On NL-to-code benchmarks that require library use, we find that LLMs generate code that uses non-existent library features in 8.1-40% of responses.One intuitive approach for detection and mitigation of hallucinations is static analysis. In this paper, we analyse the potential of static analysis tools, both in terms of what they can solve and what they cannot. We find that static analysis tools can detect 16-70% of all errors, and 14-85% of library hallucinations, with performance varying by LLM and dataset. Through manual analysis, we identify cases a static method could not plausibly catch, which gives an upper bound on their potential from 48.5% to 77%. Overall, we show that static analysis methods are cheap method for addressing some forms of hallucination, and we quantify how far short of solving the problem they will always be.
Authors: Manoj Chandrashekar Rao
Abstract: We investigate a fundamental structural question in Grouped Query Attention (GQA) transformers: do the layers most sensitive to task correctness coincide with the layers where positional encoding adaptation has the greatest leverage? We term this the co-localization hypothesis and test it on Llama 3.1 8B, a 32-layer GQA model with a 4:1 query-to-key-value head ratio. We introduce \LSLORA, which restricts LoRA adaptation to layers identified via a novel correctness-differential hidden-state metric, and GARFA (GQA-Aware RoPE Frequency Adaptation), which attaches 8 learnable per-KV-head scalar multipliers to each targeted layer. Contrary to the co-localization hypothesis, we discover strong anti-localization: task-sensitive layers concentrate in the late network ($\ell\in\{23\text{-}31\}$) while RoPE-influential layers dominate the early network ($\ell\in\{0\text{-}9\}$), yielding Spearman $r_s = -0.735$ ($p = 1.66\times10^{-6}$). Despite this anti-localization, a 4-way cross-layer ablation shows that applying both interventions to the sensitivity-identified layers outperforms all alternative configurations by 4-16 percentage points across six diverse benchmarks (MMLU, GPQA, HumanEval+, MATH, MGSM, ARC), approaching Claude 3.5 Haiku on HumanEval+ (67.1% vs. 68.3%) at \$100 total compute cost.
Authors: Atahan Dokme, Benjamin Reichman, Larry Heck
Abstract: Large language models are trained and evaluated on quantitative reasoning tasks written in clean, emotionally neutral language. However, real-world queries are often wrapped in frustration, urgency or enthusiasm. Does emotional framing alone degrade reasoning when all numerical content is preserved? To investigate this, a controlled emotion translation framework is developed that rewrites problems into emotional variants while preserving all quantities and relationships. Using this framework, Temper-5400 (5,400 semantically verified emotion--neutral pairs) is constructed across GSM8K, MultiArith, and ARC-Challenge, and evaluated on eighteen models (1B to frontier scale). Two core results emerge: First, emotional framing reduces accuracy by 2-10 percentage points even though all numerical content is preserved. Second, neutralizing emotional variants recovers most of the lost performance, showing both that the degradation is tied to emotional style rather than content corruption and that neutralization can serve as a lightweight inference-time mitigation. Non-emotional paraphrases cause no such degradation, implicating emotional content rather than surface-level changes. Beyond emotion specifically, the benchmark construction procedure provides a general framework for controlled stylistic translation and robustness evaluation.
Authors: Kaiyuan Tian, Yu Tang, Gongqingjian Jiang, Baihui Liu, Yifu Gao, Xialin Su, Linbo Qiao, Dongsheng Li
Abstract: Full-parameter fine-tuning of large language models is constrained by substantial GPU memory requirements. Low-rank adaptation methods mitigate this challenge by updating only a subset of parameters. However, these approaches often limit model expressiveness and yield lower performance than full-parameter fine-tuning. Layer-wise fine-tuning methods have emerged as an alternative, enabling memory-efficient training through static layer importance sampling strategies. However, these methods overlook variations in layer importance across tasks and training stages, resulting in suboptimal performance on downstream tasks. To address these limitations, we propose GRASS, a gradient-based adaptive layer-wise importance sampling framework. GRASS utilizes mean gradient norms as a task-aware and training-stage-aware metric for estimating layer importance. Furthermore, GRASS adaptively adjusts layer sampling probabilities through an adaptive training strategy. We also introduce a layer-wise optimizer state offloading mechanism that overlaps computation and communication to further reduce memory usage while maintaining comparable training throughput. Extensive experiments across multiple models and benchmarks demonstrate that GRASS consistently outperforms state-of-the-art methods, achieving an average accuracy improvement of up to 4.38 points and reducing memory usage by up to 19.97\%.
Authors: Yuxuan Hu, Jianchao Tan, Jiaqi Zhang, Wen Zan, Pingwei Sun, Yifan Lu, Yerui Sun, Yuchen Xie, Xunliang Cai, Jing Zhang
Abstract: Long-context inference in LLMs faces the dual challenges of quadratic attention complexity and prohibitive KV cache memory. While token-level sparse attention offers superior accuracy, its indexing overhead is costly; block-level methods improve efficiency but sacrifice precision. We propose AsyncTLS, a hierarchical sparse attention system that combines coarse-grained block filtering with fine-grained token selection to balance accuracy and efficiency, coupled with an asynchronous offloading engine that overlaps KV cache transfers with computation via temporal locality exploitation. Evaluated on Qwen3 and GLM-4.7-Flash across GQA, and MLA architectures, AsyncTLS achieves accuracy comparable to full attention while delivering 1.2x - 10.0x operator speedups and 1.3x - 4.7x end-to-end throughput improvements on 48k - 96k contexts.
Authors: Kunfeng Chen, Luyao Zhuang, Fei Liao, Juhua Liu, Jian Wang, Bo Du
Abstract: Tool learning has emerged as a promising paradigm for large language models (LLMs) to address real-world challenges. Due to the extensive and irregularly updated number of tools, tool retrieval for selecting the desired tool subset is essential. However, current tool retrieval methods are usually based on academic benchmarks containing overly detailed instructions (e.g., specific API names and parameters), while real-world instructions are more vague. Such a discrepancy would hinder the tool retrieval in real-world applications. In this paper, we first construct a new benchmark, VGToolBench, to simulate human vague instructions. Based on this, we conduct a series of preliminary analyses and find that vague instructions indeed damage the performance of tool retrieval. To this end, we propose a simple-yet-effective Tool Retrieval Bridge (TRB) approach to boost the performance of tool retrieval for vague instructions. The principle of TRB is to introduce a bridge model to rewrite the vague instructions into more specific ones and alleviate the gap between vague instructions and retriever preferences.We conduct extensive experiments under multiple commonly used retrieval settings, and the results show that TRB effectively mitigates the ambiguity of vague instructions while delivering consistent and substantial improvements across all baseline retrievers. For example, with the help of TRB, BM25 achieves a relative improvement of up to 111.51%, i.e., increasing the average NDCG score from 9.73 to 19.59. The source code and models are publicly available at https://github.com/kfchenhn/TRB.
Authors: Harsh Kohli, Srinivasan Parthasarathy, Huan Sun, Yuekun Yao
Abstract: We study implicit reasoning, i.e. the ability to combine knowledge or rules within a single forward pass. While transformer-based large language models store substantial factual knowledge and rules, they often fail to compose this knowledge for implicit multi-hop reasoning, suggesting a lack of compositional generalization over their parametric knowledge. To address this limitation, we study recurrent-depth transformers, which enables iterative computation over the same transformer layers. We investigate two compositional generalization challenges under the implicit reasoning scenario: systematic generalization, i.e. combining knowledge that is never used for compositions during training, and depth extrapolation, i.e. generalizing from limited reasoning depth (e.g. training on up to 5-hop) to deeper compositions (e.g. 10-hop). Through controlled studies with models trained from scratch, we show that while vanilla transformers struggle with both generalization challenges, recurrent-depth transformers can effectively make such generalization. For systematic generalization, we find that this ability emerges through a three-stage grokking process, transitioning from memorization to in-distribution generalization and finally to systematic generalization, supported by mechanistic analysis. For depth extrapolation, we show that generalization beyond training depth can be unlocked by scaling inference-time recurrence, with more iterations enabling deeper reasoning. We further study how training strategies affect extrapolation, providing guidance on training recurrent-depth transformers, and identify a key limitation, overthinking, where excessive recurrence degrades predictions and limits generalization to very deep compositions.
Authors: Michelle Damin Kim, Ellie S. Paek, Yufen Lin, Emily Mroz, Jane Chung, Jinho D. Choi
Abstract: This paper presents an LLM-driven approach for constructing diverse social media datasets to measure and compare loneliness in the caregiver and non-caregiver populations. We introduce an expert-developed loneliness evaluation framework and an expert-informed typology for categorizing causes of loneliness for analyzing social media text. Using a human-validated data processing pipeline, we apply GPT-4o, GPT-5-nano, and GPT-5 to build a high-quality Reddit corpus and analyze loneliness across both populations. The loneliness evaluation framework achieved average accuracies of 76.09% and 79.78% for caregivers and non-caregivers, respectively. The cause categorization framework achieved micro-aggregate F1 scores of 0.825 and 0.80 for caregivers and non-caregivers, respectively. Across populations, we observe substantial differences in the distribution of types of causes of loneliness. Caregivers' loneliness were predominantly linked to caregiving roles, identity recognition, and feelings of abandonment, indicating distinct loneliness experiences between the two groups. Demographic extraction further demonstrates the viability of Reddit for building a diverse caregiver loneliness dataset. Overall, this work establishes an LLM-based pipeline for creating high quality social media datasets for studying loneliness and demonstrates its effectiveness in analyzing population-level differences in the manifestation of loneliness.
Authors: Jingyi Kang, Chunyu Li, Ding Chen, Bo Tang, Feiyu Xiong, Zhiyu Li
Abstract: Long-term memory is fundamental for personalized and autonomous agents, yet populating it remains a bottleneck. Existing systems treat memory extraction as a one-shot, passive transcription from context to structured entries, which struggles with noisy dialogue, missing references, and cross-turn dependencies, leading to memory pollution, low-value writes, and inconsistency. In this paper, we introduce the MemReader family for active long-term memory extraction in agent systems: MemReader-0.6B, a compact and cost-efficient passive extractor distilled for accurate and schema-consistent structured outputs, and MemReader-4B, an active extractor optimized with Group Relative Policy Optimization (GRPO) to make memory writing decisions. Under a ReAct-style paradigm, MemReader-4B explicitly evaluates information value, reference ambiguity, and completeness before acting, and can selectively write memories, defer incomplete inputs, retrieve historical context, or discard irrelevant chatter. Experiments on LOCOMO, LongMemEval, and HaluMem show that MemReader consistently outperforms existing extraction-based baselines. In particular, MemReader-4B achieves state-of-the-art performance on tasks involving knowledge updating, temporal reasoning, and hallucination reduction. These results suggest that effective agent memory requires not merely extracting more information, but performing reasoning-driven and selective memory extraction to build low-noise and dynamically evolving long-term memory. Furthermore, MemReader has been integrated into MemOS and is being deployed in real-world applications. To support future research and adoption, we release the models and provide public API access.
Authors: Annerose Eichel, Tonmoy Rakshit, Sabine Schulte im Walde
Abstract: This work explores the connection between (non-)literalness and plausibility at the example of subject-verb-object events in English. We design a systematic setup of plausible and implausible event triples in combination with abstract and concrete constituent categories. Our analysis of human and LLM-generated judgments and example contexts reveals substantial differences between assessments of plausibility. While humans excel at nuanced detection and contextualization of (non-)literal vs. implausible events, LLM results reveal only shallow contextualization patterns with a bias to trade implausibility for non-literal, plausible interpretations.
Authors: Masaki Sakata, Benjamin Heinzerling, Takumi Ito, Sho Yokoi, Kentaro Inui
Abstract: We investigate how and to what extent hierarchical relations (e.g., Japan $\subset$ Eastern Asia $\subset$ Asia) are encoded in the internal representations of language models. Building on Linear Relational Concepts, we train linear transformations specific to each hierarchical depth and semantic domain, and characterize representational differences associated with hierarchical relations by comparing these transformations. Going beyond prior work on the representational geometry of hierarchies in LMs, our analysis covers multi-token entities and cross-layer representations. Across multiple domains we learn such transformations and evaluate in-domain generalization to unseen data and cross-domain transfer. Experiments show that, within a domain, hierarchical relations can be linearly recovered from model representations. We then analyze how hierarchical information is encoded in representation space. We find that it is encoded in a relatively low-dimensional subspace and that this subspace tends to be domain-specific. Our main result is that hierarchy representation is highly similar across these domain-specific subspaces. Overall, we find that all models considered in our experiments encode concept hierarchies in the form of highly interpretable linear representations.
Authors: Bo Li, Shikun Zhang, Wei Ye
Abstract: Instruction-tuned language models increasingly rely on large multi-turn dialogue corpora, but these datasets are often noisy and structurally inconsistent, with topic drift, repetitive chitchat, and mismatched answer formats across turns. We address this from a data selection perspective and propose \textbf{MDS} (Multi-turn Dialogue Selection), a dialogue-level framework that scores whole conversations rather than isolated turns. MDS combines a global coverage stage that performs bin-wise selection in the user-query trajectory space to retain representative yet non-redundant dialogues, with a local structural stage that evaluates within-dialogue reliability through entity-grounded topic grounding and information progress, together with query-answer form consistency for functional alignment. MDS outperforms strong single-turn selectors, dialogue-level LLM scorers, and heuristic baselines on three multi-turn benchmarks and an in-domain Banking test set, achieving the best overall rank across reference-free and reference-based metrics, and is more robust on long conversations under the same training budget. Code and resources are included in the supplementary materials.
Authors: Xinliang Frederick Zhang, Lu Wang
Abstract: Personalized large language models (PLLMs) have garnered significant attention for their ability to align outputs with individual's needs and preferences. However, they still struggle with long-horizon tasks, such as tracking a user's extensive history of conversations or activities. Existing memory mechanisms often fail to capture evolving behaviors, and RAG paradigms are trapped by a quality-efficiency tradeoff. Meanwhile, parametric adaptation is bottlenecked by train-inference gap due to the scarcity of labeled data. To enhance the long-horizon capabilities of PLLMs, we introduce TSUBASA, a two-pronged approach designed to improve memory writing via dynamic memory evolution, and memory reading via self-learning with a context distillation objective to internalize user experiences. Extensive evaluations on long-horizon benchmarks using the Qwen-3 model family (4B to 32B) validate the effectiveness of TSUBASA, surpassing competitive memory-augmented systems that rely primarily on memory writing, such as Mem0 and Memory-R1. Our analyses further confirms that TSUBASA breaks the quality-efficiency barrier to achieve Pareto improvements, delivering robust, high-fidelity personalization with a reduced token budget.
Authors: Guoqi Ma, Liang Zhang, Hongyao Tu, Hao Fu, Hui Li, Yujie Lin, Longyue Wang, Weihua Luo, Jinsong Su
Abstract: Cross-document relation extraction (RE) aims to identify relations between the head and tail entities located in different documents. Existing approaches typically adopt the paradigm of ``\textit{Small Language Model (SLM) + Classifier}''. However, the limited language understanding ability of SLMs hinders further improvement of their performance. In this paper, we conduct a preliminary study to explore the performance of Large Language Models (LLMs) in cross-document RE. Despite their extensive parameters, our findings indicate that LLMs do not consistently surpass existing SLMs. Further analysis suggests that the underperformance is largely attributed to the challenges posed by the numerous predefined relations. To overcome this issue, we propose an LLM-based \underline{H}ierarchical \underline{C}lassification model for cross-document \underline{RE} (HCRE), which consists of two core components: 1) an LLM for relation prediction and 2) a \textit{hierarchical relation tree} derived from the predefined relation set. This tree enables the LLM to perform hierarchical classification, where the target relation is inferred level by level. Since the number of child nodes is much smaller than the size of the entire predefined relation set, the hierarchical relation tree significantly reduces the number of relation options that LLM needs to consider during inference. However, hierarchical classification introduces the risk of error propagation across levels. To mitigate this, we propose a \textit{prediction-then-verification} inference strategy that improves prediction reliability through multi-view verification at each level. Extensive experiments show that HCRE outperforms existing baselines, validating its effectiveness.
Authors: Shiwan Zhao, Zhihu Wang, Xuyang Zhao, Jiaming Zhou, Caiyue Xu, Chenfei Liu, Liting Zhang, Yuhang Jia, Yanzhe Zhang, Hualong Yu, Zichen Xu, Qicheng Li, Yong Qin
Abstract: Post-training has become central to turning pretrained large language models (LLMs) into aligned and deployable systems. Recent progress spans supervised fine-tuning (SFT), preference optimization, reinforcement learning (RL), process supervision, verifier-guided methods, distillation, and multi-stage pipelines. Yet these methods are often discussed in fragmented ways, organized by labels or objective families rather than by the behavioral bottlenecks they address. This survey argues that LLM post-training is best understood as structured intervention on model behavior. We organize the field first by trajectory provenance, which defines two primary learning regimes: off-policy learning on externally supplied trajectories, and on-policy learning on learner-generated rollouts. We then interpret methods through two recurring roles -- effective support expansion, which makes useful behaviors more reachable, and policy reshaping, which improves behavior within already reachable regions -- together with a complementary systems-level role, behavioral consolidation, which preserves, transfers, and amortizes behavior across stages and model transitions. This perspective yields a unified reading of major paradigms. SFT may serve either support expansion or policy reshaping, whereas preference-based methods are usually off-policy reshaping. On-policy RL often improves behavior on learner-generated states, though under stronger guidance it can also make hard-to-reach reasoning paths reachable. Distillation is often best understood as consolidation rather than only compression, and hybrid pipelines emerge as coordinated multi-stage compositions. Overall, the framework helps diagnose post-training bottlenecks and reason about stage composition, suggesting that progress in LLM post-training increasingly depends on coordinated system design rather than any single dominant objective.
Authors: Yuanjian Xu, Tianze Sun, Changwei Xu, XinLong Zhao, Jianing Hao, Ran Chen, Yang Liu, Ruijie Xu, Stephen Chen, Guang Zhang
Abstract: Data mixing strategy is essential for large language model (LLM) training. Empirical evidence shows that inappropriate strategies can significantly reduce generalization. Although recent methods have improved empirical performance, several fundamental questions remain open: what constitutes a domain, whether human and model perceptions of domains are aligned, and how domain weighting influences generalization. We address these questions by establishing formal connections between gradient dynamics and domain distributions, offering a theoretical framework that clarifies the role of domains in training dynamics. Building on this analysis, we introduce DoGraph, a reweighting framework that formulates data scheduling as a graph-constrained optimization problem. Extensive experiments on GPT-2 models of varying scales demonstrate that DoGraph consistently achieves competitive performance.
Authors: Hongyi Cen, Mingxin Wang, Yule Liu, Jingyi Zheng, Hanze Jia, Tan Tang, Yingcai Wu
Abstract: Adversarial claim rewriting is widely used to test fact-checking systems, but standard metrics fail to capture truth-conditional consistency and often label semantically corrupted rewrites as successful. We introduce AtomEval, a validity-aware evaluation framework that decomposes claims into subject-relation-object-modifier (SROM) atoms and scores adversarial rewrites with Atomic Validity Scoring (AVS), enabling detection of factual corruption beyond surface similarity. Experiments on the FEVER dataset across representative attack strategies and LLM generators show that AtomEval provides more reliable evaluation signals in our experiments. Using AtomEval, we further analyze LLM-based adversarial generators and observe that stronger models do not necessarily produce more effective adversarial claims under validity-aware evaluation, highlighting previously overlooked limitations in current adversarial evaluation practices.
Authors: George Fountzoulas
Abstract: We present Kathleen, a text classification architecture that operates directly on raw UTF-8 bytes using frequency-domain processing -- requiring no tokenizer, no attention mechanism, and only 733K parameters. Kathleen introduces three novel components: (1) RecurrentOscillatorBanks -- damped sinusoid convolutions with temporal memory for O(L) sequence processing; (2) an FFT-Rotate Wavetable Encoder that maps all 256 byte values using a single learnable vector (256 floats), replacing conventional embedding tables (65K parameters) while improving accuracy; (3) PhaseHarmonics -- a sinusoidal non-linearity with just 6 learnable phase parameters that our ablation identifies as the single most impactful component (+2.6% accuracy, <0.001% of model parameters). Through comprehensive ablation of a 1.8M-parameter predecessor, we show that frequency-domain components systematically outperform complex cognitive architectures: removing a 560K-parameter bio-inspired framework costs only -0.2%, while removing the 6-parameter PhaseHarmonics costs -2.6%. The resulting Kathleen-Clean achieves 88.6% on IMDB, 92.3% on AG News, and 83.3% on SST-2 -- outperforming a tokenized counterpart with 16x more parameters on IMDB (+1.6%) and AG News (+2.1%). Kathleen processes sequences in O(L) time and memory, enabling byte-level operation at sequence lengths where O(L^2) Transformers exhaust GPU memory.
Authors: Yanling Xiao, Huaibing Xie, Guoliang Zhao, Shihan Dou, Shaolei Wang, Yiting Liu, Nantao Zheng, Cheng Zhang, Pluto Zhou, Zhisong Zhang, Lemao Liu
Abstract: Long-context reasoning is essential for complex real-world applications, yet remains a significant challenge for Large Language Models (LLMs). Despite the rapid evolution in long-context reasoning, current research often overlooks the internal complexity of the long-context reasoning task itself. In this paper, we move beyond this holistic view and decompose long-context reasoning into a set of fundamental atomic skills, and we then automatically synthesize a suite of pseudo datasets, each explicitly targeting a specific atomic skill. Our empirical analysis confirms that proficiency in these atomic skills is strongly correlated with general long-text reasoning performance. Building on this insight, we employ reinforcement learning on these pseudo datasets to sharpen the model's atomic skills, in the hope of boosting its general long-context reasoning ability. Extensive experiments across multiple benchmarks demonstrate the effectiveness of our approach: it outperforms a strong baseline by an average margin of 7.7\% (improving from 46.3\% to 54.0\%) across Loogle, Loong, LongBench-v2, BrowscompLong, Ruler-qa2, and MRCR.
Authors: Or Dado, David Carmel. Oren Kurland
Abstract: We address the task of predicting the gain of using RAG (retrieval augmented generation) for question answering with respect to not using it. We study the performance of a few pre-retrieval and post-retrieval predictors originally devised for ad hoc retrieval. We also study a few post-generation predictors, one of which is novel to this study and posts the best prediction quality. Our results show that the most effective prediction approach is a novel supervised predictor that explicitly models the semantic relationships among the question, retrieved passages, and the generated answer.
Authors: Zhengyi Zhao, Shubo Zhang, Zezhong Wang, Yuxi Zhang, Huimin Wang, Yutian Zhao, Yefeng Zheng, Binyang Li, Kam-Fai Wong, Xian Wu
Abstract: Retrieval-Augmented Generation (RAG) significantly enhances Large Language Models (LLMs) by providing access to external knowledge. However, current research primarily focuses on retrieval quality, often overlooking the critical ''integration bottleneck'': even when relevant documents are retrieved, LLMs frequently fail to utilize them effectively due to conflicts with their internal parametric knowledge. In this paper, we argue that implicitly resolving this conflict in a single generation pass is suboptimal. We introduce GuarantRAG, a framework that explicitly decouples reasoning from evidence integration. First, we generate an ''Inner-Answer'' based solely on parametric knowledge to capture the model's reasoning flow. Second, to guarantee faithful evidence extraction, we generate a ''Refer-Answer'' using a novel Contrastive DPO objective. This objective treats the parametric Inner-Answer as a negative constraint and the retrieved documents as positive ground truth, forcing the model to suppress internal hallucinations in favor of external evidence during this phase. Finally, rather than naive concatenation or using the DPO trained model directly, we propose a joint decoding mechanism that dynamically fuses the logical coherence of the Inner-Answer with the factual precision of the Refer-Answer at the token level. Experiments on five QA benchmarks demonstrate that GuarantRAG improves accuracy by up to 12.1% and reduces hallucinations by 16.3% compared to standard and dynamic RAG baselines.
Authors: Ruiyi Yan, Yugo Murawaki
Abstract: Linguistic steganography involves embedding secret messages within seemingly innocuous texts to enable covert communication. Provable security, which is a long-standing goal and key motivation, has been extended to language-model-based steganography. Previous provably secure approaches have achieved perfect imperceptibility, measured by zero Kullback-Leibler (KL) divergence, but at the expense of embedding capacity. In this paper, we attempt to directly use a classic entropy coding method (range coding) to achieve secure steganography, and then propose an efficient and provably secure linguistic steganographic method with a rotation mechanism. Experiments across various language models show that our method achieves around 100% entropy utilization (embedding efficiency) for embedding capacity, outperforming the existing baseline methods. Moreover, it achieves high embedding speeds (up to 1554.66 bits/s on GPT-2). The code is available at github.com/ryehr/RRC_steganography.
Authors: Xunzhuo Liu, Bowei He, Xue Liu, Andy Luo, Haichen Zhang, Huamin Chen
Abstract: Production vLLM fleets typically provision each instance for the worst-case context length, leading to substantial KV-cache over-allocation and under-utilized concurrency. In practice, 80-95% of requests are short, yet are served under configurations optimized for long contexts, wasting 4-8$\times$ throughput capacity and triggering reliability issues such as OOM crashes, preemption, and request rejections. We identify a common root cause for these inefficiencies: configuration-traffic mismatch. We propose dual-pool token-budget routing, a lightweight dispatch mechanism that partitions a homogeneous fleet into two specialized pools: a high-throughput short-context pool and a high-capacity long-context pool. Each request is routed based on its estimated total token budget, computed using a per-category bytes-to-token ratio that is learned online via exponential moving average from usage.prompt_tokens feedback, eliminating the need for a tokenizer. We also develop a simple analytical model that predicts fleet-level cost savings from workload characteristics and measured throughput differences, enabling practitioners to estimate benefits prior to deployment. Evaluations on real-world traces from the Azure LLM Inference Dataset and LMSYS-Chat-1M, serving Llama-3-70B on A100 GPUs, show that our approach reduces GPU-hours by 31-42%, corresponding to \$2.86M annual savings at fleet scale, while lowering preemption rates by 5.4$\times$ and improving P99 TTFT by 6%. A case study with Qwen3-235B-A22B on AMD MI300X at 10,000 req/s projects \$15.4M in annual savings. The method incurs only O(1) dispatch overhead, adapts automatically to heterogeneous workloads, and composes seamlessly with existing optimizations such as PagedAttention, continuous batching, and prefill-decode disaggregation.
Authors: Khalid Zaman, Melike Sah, Anuwat Chaiwongyenc, Cem Direkoglu
Abstract: We propose Quantum Vision (QV) theory as a new perspective for deep learning-based audio classification, applied to deepfake speech detection. Inspired by particle-wave duality in quantum physics, QV theory is based on the idea that data can be represented not only in its observable, collapsed form, but also as information waves. In conventional deep learning, models are trained directly on these collapsed representations, such as images. In QV theory, inputs are first transformed into information waves using a QV block, and then fed into deep learning models for classification. QV-based models improve performance in image classification compared to their non-QV counterparts. What if QV theory is applied speech spectrograms for audio classification tasks? This is the motivation and novelty of the proposed approach. In this work, Short-Time Fourier Transform (STFT), Mel-spectrograms, and Mel-Frequency Cepstral Coefficients (MFCC) of speech signals are converted into information waves using the proposed QV block and used to train QV-based Convolutional Neural Networks (QV-CNN) and QV-based Vision Transformers (QV-ViT). Extensive experiments are conducted on the ASVSpoof dataset for deepfake speech classification. The results show that QV-CNN and QV-ViT consistently outperform standard CNN and ViT models, achieving higher classification accuracy and improved robustness in distinguishing genuine and spoofed speech. Moreover, the QV-CNN model using MFCC features achieves the best overall performance on the ASVspoof dataset, with an accuracy of 94.20% and an EER of 9.04%, while the QV-CNN with Mel-spectrograms attains the highest accuracy of 94.57%. These findings demonstrate that QV theory is an effective and promising approach for audio deepfake detection and opens new directions for quantum-inspired learning in audio perception tasks.
Authors: Ian W. Kennedy, Nafise Sadat Moosavi
Abstract: Additive quantization enables extreme LLM compression with O(1) lookup-table dequantization, making it attractive for edge deployment. Yet at 2-bit precision, it often fails catastrophically, even with extensive search and finetuning. We show that the dominant bottleneck is codebook initialisation. Greedy sequential initialisation frequently places the model in poor optimisation regions that subsequent beam search and PV-tuning struggle to overcome. We analyse this behaviour through the representational ratio \r{ho} = N/KM, which characterises the relationship between weight groups and codebook capacity, and propose OA-EM, an output-aware EM initialisation method using Hessian-weighted Mahalanobis distance. Across compression rates, search budgets, and three architectures (Llama 3.2 3B, Llama 3.1 8B, Qwen 2.5 3B), OA-EM consistently produces better solutions after PV-tuning and dominates the quality-compute frontier. The severity of the bottleneck scales with \r{ho}: moderate at 3 bpp but extreme at 2 bpp, where poor initialisation can degrade perplexity by orders of magnitude. More broadly, our results highlight the importance of optimisation geometry in compressed model spaces, where initialisation can dominate subsequent search and fine-tuning.
Authors: Tian Huang, Tom Bourgeade, Irina Illina
Abstract: Objective Structured Clinical Examinations (OSCEs) are the standard method for assessing medical students' clinical and communication skills through structured patient interviews. In France, however, the organization of training sessions is limited by human and logistical constraints, restricting students' access to repeated practice and structured feedback. Recent advances in Natural Language Processing (NLP) and Large Language Models (LLMs) now offer the opportunity to automatically evaluate such medical interviews, thereby alleviating the need for human examiners during training. Yet, real French OSCE annotated transcripts remain extremely scarce, limiting reproducible research and reliable benchmarking. To address these challenges, we investigate the use of LLMs for both generating and evaluating French OSCE dialogues in a low-resource context. We introduce a controlled pipeline that produces synthetic doctor-patient interview transcripts guided by scenario-specific evaluation criteria, combining ideal and perturbed performances to simulate varying student skill levels. The resulting dialogues are automatically silver-labeled through an LLM-assisted framework supporting adjustable evaluation strictness. Benchmarking multiple open-source and proprietary LLMs shows that mid-size models ($\le$32B parameters) achieve accuracies comparable to GPT-4o ($\sim$90\%) on synthetic data, highlighting the feasibility of locally deployable, privacy-preserving evaluation systems for medical education.
Authors: Soveatin Kuntur, Maciej Krzywda, Anna Wr\'oblewska, Marcin Paprzycki, Maria Ganzha, Szymon {\L}ukasik, Amir H. Gandomi
Abstract: The rapid spread of online misinformation has led to increasingly complex detection models, including large language models and hybrid architectures. However, their computational cost and deployment limitations raise concerns about practical applicability. In this work, we benchmark graph neural networks (GNNs) against non-graph-based machine learning methods under controlled and comparable conditions. We evaluate lightweight GNN architectures (GCN, GraphSAGE, GAT, ChebNet) against Logistic Regression, Support Vector Machines, and Multilayer Perceptrons across seven public datasets in English, Indonesian, and Polish. All models use identical TF-IDF features to isolate the impact of relational structure. Performance is measured using F1 score, with inference time reported to assess efficiency. GNNs consistently outperform non-graph baselines across all datasets. For example, GraphSAGE achieves 96.8% F1 on Kaggle and 91.9% on WELFake, compared to 73.2% and 66.8% for MLP, respectively. On COVID-19, GraphSAGE reaches 90.5% F1 vs. 74.9%, while ChebNet attains 79.1% vs. 66.4% on FakeNewsNet. These gains are achieved with comparable or lower inference times. Overall, the results show that classic GNNs remain effective and efficient, challenging the need for increasingly complex architectures in misinformation detection.
Authors: Soveatin Kuntur, Panggih Kusuma Ningrum, Anna Wr\'oblewska, Maria Ganzha, Marcin Paprzycki
Abstract: We propose a lightweight hybrid approach to clickbait detection that combines OpenAI semantic embeddings with six compact heuristic features capturing stylistic and informational cues. To improve efficiency, embeddings are reduced using PCA and evaluated with XGBoost, GraphSAGE, and GCN classifiers. While the simplified feature design yields slightly lower F1-scores, graph-based models achieve competitive performance with substantially reduced inference time. High ROC--AUC values further indicate strong discrimination capability, supporting reliable detection of clickbait headlines under varying decision thresholds.
Authors: Petr Plech\'a\v{c}, Artjoms \v{S}e\c{l}a, Silvie Cinkov\'a, Mirella De Sisto, Lara Nugues, Ne\v{z}a Ko\v{c}nik, Antonina Martynenko, Ben Nagy, Luca Giovannini, Robert Kol\'ar
Abstract: Rhyme is deceptively intuitive: what is or is not a rhyme is constructed historically, scholars struggle with rhyme classification, and people disagree on whether two words are rhymed or not. This complicates automated rhymed recognition and evaluation, especially in multilingual context. This article investigates how much training data is needed for reliable unsupervised rhyme recognition using RhymeTagger, a language-independent tool that identifies rhymes based on repeating patterns in poetry corpora. We evaluate its performance across seven languages (Czech, German, English, French, Italian, Russian, and Slovene), examining how training size and language differences affect accuracy. To set a realistic performance benchmark, we assess inter-annotator agreement on a manually annotated subset of poems and analyze factors contributing to disagreement in expert annotations: phonetic similarity between rhyming words and their distance from each other in a poem. We also compare RhymeTagger to three large language models using a one-shot learning strategy. Our findings show that, once provided with sufficient training data, RhymeTagger consistently outperforms human agreement, while LLMs lacking phonetic representation significantly struggle with the task.
Authors: Xuan Feng, Shuai Zhao, Luwei Xiao, Tianlong Gu, Bo An
Abstract: Although Large Language Models (LLMs) demonstrate remarkable reasoning capabilities, inherent social biases often cascade throughout the Chain-of-Thought (CoT) process, leading to continuous "Bias Propagation". Existing debiasing methods primarily focus on static constraints or external interventions, failing to identify and interrupt this propagation once triggered. To address this limitation, we introduce Self-Debias, a progressive framework designed to instill intrinsic self-correction capabilities. Specifically, we reformulate the debiasing process as a strategic resource redistribution problem, treating the model's output probability mass as a limited resource to be reallocated from biased heuristics to unbiased reasoning paths. Unlike standard preference optimization which applies broad penalties, Self-Debias employs a fine-grained trajectory-level objective subject to dynamic debiasing constraints. This enables the model to selectively revise biased reasoning suffixes while preserving valid contextual prefixes. Furthermore, we integrate an online self-improvement mechanism utilizing consistency filtering to autonomously synthesize supervision signals. With merely 20k annotated samples, Self-Debias activates efficient self-correction, achieving superior debiasing performance while preserving general reasoning capabilities without continuous external oversight.
Authors: Juwei Yue, Chuanrui Hu, Jiawei Sheng, Zuyi Zhou, Wenyuan Zhang, Tingwen Liu, Li Guo, Yafeng Deng
Abstract: Long-term memory is essential for conversational agents to maintain coherence, track persistent tasks, and provide personalized interactions across extended dialogues. However, existing approaches as Retrieval-Augmented Generation (RAG) and graph-based memory mostly rely on pairwise relations, which can hardly capture high-order associations, i.e., joint dependencies among multiple elements, causing fragmented retrieval. To this end, we propose HyperMem, a hypergraph-based hierarchical memory architecture that explicitly models such associations using hyperedges. Particularly, HyperMem structures memory into three levels: topics, episodes, and facts, and groups related episodes and their facts via hyperedges, unifying scattered content into coherent units. Leveraging this structure, we design a hybrid lexical-semantic index and a coarse-to-fine retrieval strategy, supporting accurate and efficient retrieval of high-order associations. Experiments on the LoCoMo benchmark show that HyperMem achieves state-of-the-art performance with 92.73% LLM-as-a-judge accuracy, demonstrating the effectiveness of HyperMem for long-term conversations.
Authors: Jun Seo, Sangwon Ryu, Heejin Do, Hyounghun Kim, Gary Geunbae Lee
Abstract: Knowledge Tracing (KT) aims to predict learners' future performance from past interactions. While recent KT approaches have improved via learning item representations aligned with Knowledge Components, they overlook the procedural dynamics of problem solving. We propose Behavior-Aware Item Modeling (BAIM), a framework that enriches item representations by integrating dynamic procedural solution information. BAIM leverages a reasoning language model to decompose each item's solution into four problem-solving stages (i.e., understand, plan, carry out, and look back), pedagogically grounded in Polya's framework. Specifically, it derives stage-level representations from per-stage embedding trajectories, capturing latent signals beyond surface features. To reflect learner heterogeneity, BAIM adaptively routes these stage-wise representations, introducing a context-conditioned mechanism within a KT backbone, allowing different procedural stages to be emphasized for different learners. Experiments on XES3G5M and NIPS34 show that BAIM consistently outperforms strong pretraining-based baselines, achieving particularly large gains under repeated learner interactions.
Authors: Prisca Piccirilli, Alexander Fraser, Sabine Schulte im Walde
Abstract: Metaphor pervades everyday language, allowing speakers to express abstract concepts via concrete domains. While prior work has studied metaphors cognitively and psycholinguistically, large-scale comparisons with literal language remain limited, especially for near-synonymous expressions. We analyze 297 English verb-object pairs (e.g., float idea vs. suggest idea) in ~2M corpus sentences, examining their contextual usage. Using five NLP tools, we extract 2,293 cognitive and linguistic features capturing affective, lexical, syntactic, and discourse-level properties. We address: (i) whether features differ between metaphorical and literal contexts (cross-pair analysis), and (ii) whether individual VO pairs diverge internally (within-pair analysis). Cross-pair results show literal contexts have higher lexical frequency, cohesion, and structural regularity, while metaphorical contexts show greater affective load, imageability, lexical diversity, and constructional specificity. Within-pair analyses reveal substantial heterogeneity, with most pairs showing non-uniform effects. These results suggest no single, consistent distributional pattern that distinguishes metaphorical from literal usage. Instead, differences are largely construction-specific. Overall, large-scale data combined with diverse features provides a fine-grained understanding of metaphor-literal contrasts in VO usage.
Authors: Ruotao Xu, Yixin Ji, Yu Luo, Jinpeng Li, Dong Li, Peifeng Li, Juntao Li, Min Zhang
Abstract: Large reasoning models (LRMs) have achieved strong performance enhancement through scaling test time computation, but due to the inherent limitations of the underlying language models, they still have shortcomings in tasks that require precise computation and extensive knowledge reserves. Tool-Integrated Reasoning (TIR) has emerged as a promising paradigm that incorporates tool call and execution within the reasoning trajectory. Although recent works have released some powerful open-source TIR models, our analysis reveals that these models still suffer from critical deficiencies. We find that when the reasoning of the model conflicts with the tool results, the model tends to believe in its own reasoning. And there are cases where the tool results are correct but are ignored by the model, resulting in incorrect answers, which we define as "Tool Ignored''. This indicates that the model does not know when to trust or ignore the tool. To overcome these limitations, We introduce Adaptive Tool Trust Calibration (ATTC), a novel framework that guides the model to adaptively choose to trust or ignore the tool results based on the confidence score of generated code blocks. The experimental results from various open-source TIR models of different sizes and across multiple datasets demonstrate that ATTC effectively reduces the "Tool Ignored" issue, resulting in a performance increase of 4.1% to 7.5%.
Authors: Yating Wang, Wenting Zhao, Yaqi Zhao, Yongshun Gong, Yilong Yin, Haoliang Sun
Abstract: Large language models store not only isolated facts but also rules that support reasoning across symbolic expressions, natural language explanations, and concrete instances. Yet most model editing methods are built for fact-level knowledge, assuming that a target edit can be achieved through a localized intervention. This assumption does not hold for rule-level knowledge, where a single rule must remain consistent across multiple interdependent forms. We investigate this problem through a mechanistic study of rule-level knowledge editing. To support this study, we extend the RuleEdit benchmark from 80 to 200 manually verified rules spanning mathematics and physics. Fine-grained causal tracing reveals a form-specific organization of rule knowledge in transformer layers: formulas and descriptions are concentrated in earlier layers, while instances are more associated with middle layers. These results suggest that rule knowledge is not uniformly localized, and therefore cannot be reliably edited by a single-layer or contiguous-block intervention. Based on this insight, we propose Distributed Multi-Layer Editing (DMLE), which applies a shared early-layer update to formulas and descriptions and a separate middle-layer update to instances. While remaining competitive on standard editing metrics, DMLE achieves substantially stronger rule-level editing performance. On average, it improves instance portability and rule understanding by 13.91 and 50.19 percentage points, respectively, over the strongest baseline across GPT-J-6B, Qwen2.5-7B, Qwen2-7B, and LLaMA-3-8B. The code is available at https://github.com/Pepper66/DMLE.
Authors: Renyu Fu, Guibo Luo
Abstract: Chain-of-Thought (CoT) has become a cornerstone of reasoning in large language models, yet its effectiveness is constrained by the limited expressiveness of discrete token sampling. Recent latent reasoning approaches attempt to alleviate this limitation by replacing discrete tokens with soft embeddings (probability-weighted mixtures of token embeddings) or hidden states, but they commonly suffer from two issues: (1) global activation injects perturbations into high-confidence steps, impairing reasoning stability; and (2) soft embeddings quickly collapse toward the highest-probability token, limiting exploration of alternative trajectories. To address these challenges, we propose SeLaR (Selective Latent Reasoning), a lightweight and training-free framework. SeLaR introduces an entropy-gated mechanism that activates soft embeddings only at low-confidence steps, while preserving discrete decoding at high-confidence steps. Additionally, we propose an entropy-aware contrastive regularization that pushes soft embeddings away from the dominant (highest-probability) token's direction, encouraging sustained exploration of multiple latent reasoning paths. Experiments on five reasoning benchmarks demonstrate that SeLaR consistently outperforms standard CoT and state-of-the-art training-free methods.
Authors: Jiawei Chen, Ruoxi Xu, Boxi Cao, Ruotong Pan, Yunfei Zhang, Yifei Hu, Yong Du, Tingting Gao, Yaojie Lu, Yingfei Sun, Xianpei Han, Le Sun, Xiangyu Wu, Hongyu Lin
Abstract: The emergence of Large Language Models (LLMs) has illuminated the potential for a general-purpose user simulator. However, existing benchmarks remain constrained to isolated scenarios, narrow action spaces, or synthetic data, failing to capture the holistic nature of authentic human behavior. To bridge this gap, we introduce OmniBehavior, the first user simulation benchmark constructed entirely from real-world data, integrating long-horizon, cross-scenario, and heterogeneous behavioral patterns into a unified framework. Based on this benchmark, we first provide empirical evidence that previous datasets with isolated scenarios suffer from tunnel vision, whereas real-world decision-making relies on long-term, cross-scenario causal chains. Extensive evaluations of state-of-the-art LLMs reveal that current models struggle to accurately simulate these complex behaviors, with performance plateauing even as context windows expand. Crucially, a systematic comparison between simulated and authentic behaviors uncovers a fundamental structural bias: LLMs tend to converge toward a positive average person, exhibiting hyper-activity, persona homogenization, and a Utopian bias. This results in the loss of individual differences and long-tail behaviors, highlighting critical directions for future high-fidelity simulation research.
Authors: Wenxian Wang, Xiaohu Luo, Junfeng Hao, Xiaoming Gu, Xingshu Chen, Zhu Wang, Haizhou Wang
Abstract: Sarcasm is a rhetorical device that expresses criticism or emphasizes characteristics of certain individuals or situations through exaggeration, irony, or comparison. Existing methods for Chinese sarcasm detection are constrained by limited datasets and high construction costs, and they mainly focus on textual features, overlooking user-specific linguistic patterns that shape how opinions and emotions are expressed. This paper proposes a Generative Adversarial Network (GAN) and Large Language Model (LLM)-driven data augmentation framework to dynamically model users' linguistic patterns for enhanced Chinese sarcasm detection. First, we collect raw data from various topics on Sina Weibo. Then, we train a GAN on these data and apply a GPT-3.5 based data augmentation technique to synthesize an extended sarcastic comment dataset, named SinaSarc. This dataset contains target comments, contextual information, and user historical behavior. Finally, we extend the BERT architecture to incorporate multi-dimensional information, particularly user historical behavior, enabling the model to capture dynamic linguistic patterns and uncover implicit sarcastic cues in comments. Experimental results demonstrate the effectiveness of our proposed method. Specifically, our model achieves the highest F1-scores on both the non-sarcastic and sarcastic categories, with values of 0.9138 and 0.9151 respectively, which outperforms all existing state-of-the-art (SOTA) approaches. This study presents a novel framework for dynamically modeling users' long-term linguistic patterns in Chinese sarcasm detection, contributing to both dataset construction and methodological advancement in this field.
Authors: Tristan Thrush, Sung Min Park, Herman Brunborg, Luke Bailey, Marcel Roed, Neil Band, Christopher Potts, Tatsunori Hashimoto
Abstract: What are the limits of controlling language models via synthetic training data? We develop a reinforcement learning (RL) primitive, the Dataset Policy Gradient (DPG), which can precisely optimize synthetic data generators to produce a dataset of targeted examples. When used for supervised fine-tuning (SFT) of a target model, these examples cause the target model to do well on a differentiable metric of our choice. Our approach achieves this by taking exact data attribution via higher-order gradients and using those scores as policy gradient rewards. We prove that this procedure closely approximates the true, intractable gradient for the synthetic data generator. To illustrate the potential of DPG, we show that, using only SFT on generated examples, we can cause the target model's LM head weights to (1) embed a QR code, (2) embed the pattern $\texttt{67}$, and (3) have lower $\ell^2$ norm. We additionally show that we can cause the generator to (4) rephrase inputs in a new language and (5) produce a specific UUID, even though neither of these objectives is conveyed in the generator's input prompts. These findings suggest that DPG is a powerful and flexible technique for shaping model properties using only synthetic training examples.
Authors: Lilian Wanzare, Cynthia Amol, zekiel Maina, Nelson Odhiambo, Hope Kerubo, Leila Misula, Vivian Oloo, Rennish Mboya, Edwin Onkoba, Edward Ombui, Joseph Muguro, Ciira wa Maina, Andrew Kipkebut, Alfred Omondi Otom, Ian Ndung'u Kang'ethe, Angela Wambui Kanyi, Brian Gichana Omwenga
Abstract: AfriVoices-KE is a large-scale multilingual speech dataset comprising approximately 3,000 hours of audio across five Kenyan languages: Dholuo, Kikuyu, Kalenjin, Maasai, and Somali. The dataset includes 750 hours of scripted speech and 2,250 hours of spontaneous speech, collected from 4,777 native speakers across diverse regions and demographics. This work addresses the critical underrepresentation of African languages in speech technology by providing a high-quality, linguistically diverse resource. Data collection followed a dual methodology: scripted recordings drew from compiled text corpora, translations, and domain-specific generated sentences spanning eleven domains relevant to the Kenyan context, while unscripted speech was elicited through textual and image prompts to capture natural linguistic variation and dialectal nuances. A customized mobile application enabled contributors to record using smartphones. Quality assurance operated at multiple layers, encompassing automated signal-to-noise ratio validation prior to recording and human review for content accuracy. Though the project encountered challenges common to low-resource settings, including unreliable infrastructure, device compatibility issues, and community trust barriers, these were mitigated through local mobilizers, stakeholder partnerships, and adaptive training protocols. AfriVoices-KE provides a foundational resource for developing inclusive automatic speech recognition and text-to-speech systems, while advancing the digital preservation of Kenya's linguistic heritage.
Authors: Emma Gueorguieva, Hongli Zhan, Jina Suh, Javier Hernandez, Tatiana Lau, Junyi Jessy Li, Desmond C. Ong
Abstract: Recent research shows that greater numbers of people are turning to Large Language Models (LLMs) for emotional support, and that people rate LLM responses as more empathic than human-written responses. We suggest a reason for this success: LLMs have learned and consistently deploy a well-liked template for expressing empathy. We develop a taxonomy of 10 empathic language "tactics" that include validating someone's feelings and paraphrasing, and apply this taxonomy to characterize the language that people and LLMs produce when writing empathic responses. Across a set of 2 studies comparing a total of n = 3,265 AI-generated (by six models) and n = 1,290 human-written responses, we find that LLM responses are highly formulaic at a discourse functional level. We discovered a template -- a structured sequence of tactics -- that matches between 83--90% of LLM responses (and 60--83\% in a held out sample), and when those are matched, covers 81--92% of the response. By contrast, human-written responses are more diverse. We end with a discussion of implications for the future of AI-generated empathy.
Authors: Emmy Liu, Kaiser Sun, Millicent Li, Isabelle Lee, Lindia Tjuatja, Jen-tse Huang, Graham Neubig
Abstract: Large language models (LLMs) can perform remarkably complex tasks, yet the fine-grained details of how these capabilities emerge during pretraining remain poorly understood. Scaling laws on validation loss tell us how much a model improves with additional compute, but not what skills it acquires in which order. To remedy this, we propose the Implicit Curriculum Hypothesis: pretraining follows a compositional and predictable curriculum across models and data mixtures. We test this by designing a suite of simple, composable tasks spanning retrieval, morphological transformations, coreference, logical reasoning, and mathematics. Using these tasks, we track emergence points across four model families spanning sizes from 410M-13B parameters. We find that emergence orderings of when models reach fixed accuracy thresholds are strikingly consistent ($\rho = .81$ across 45 model pairs), and that composite tasks most often emerge after their component tasks. Furthermore, we find that this structure is encoded in model representations: tasks with similar function vector representations also tend to follow similar trajectories in training. By using the space of representations derived from our task set, we can effectively predict the training trajectories of simple held-out compositional tasks throughout the course of pretraining ($R^2 = .68$-$.84$ across models) without previously evaluating them. Together, these results suggest that pretraining is more structured than loss curves reveal: skills emerge in a compositional order that is consistent across models and readable from their internals.
Authors: Jiayuan Ye, Vitaly Feldman, Kunal Talwar
Abstract: Large language models (LLMs) can struggle to memorize factual knowledge in their parameters, often leading to hallucinations and poor performance on knowledge-intensive tasks. In this paper, we formalize fact memorization from an information-theoretic perspective and study how training data distributions affect fact accuracy. We show that fact accuracy is suboptimal (below the capacity limit) whenever the amount of information contained in the training data facts exceeds model capacity. This is further exacerbated when the fact frequency distribution is skewed (e.g. a power law). We propose data selection schemes based on the training loss alone that aim to limit the number of facts in the training data and flatten their frequency distribution. On semi-synthetic datasets containing high-entropy facts, our selection method effectively boosts fact accuracy to the capacity limit. When pretraining language models from scratch on an annotated Wikipedia corpus, our selection method enables a GPT2-Small model (110m parameters) to memorize 1.3X more entity facts compared to standard training, matching the performance of a 10X larger model (1.3B parameters) pretrained on the full dataset.
Authors: Yuxuan Zhang, Yubo Wang, Yipeng Zhu, Penghui Du, Junwen Miao, Xuan Lu, Wendong Xu, Yunzhuo Hao, Songcheng Cai, Xiaochen Wang, Huaisong Zhang, Xian Wu, Yi Lu, Minyi Lei, Kai Zou, Huifeng Yin, Ping Nie, Liang Chen, Dongfu Jiang, Wenhu Chen, Kelsey R. Allen
Abstract: AI agents may be able to automate your inbox, but can they automate other routine aspects of your life? Everyday online tasks offer a realistic yet unsolved testbed for evaluating the next generation of AI agents. To this end, we introduce ClawBench, an evaluation framework of 153 simple tasks that people need to accomplish regularly in their lives and work, spanning 144 live platforms across 15 categories, from completing purchases and booking appointments to submitting job applications. These tasks require demanding capabilities beyond existing benchmarks, such as obtaining relevant information from user-provided documents, navigating multi-step workflows across diverse platforms, and write-heavy operations like filling in many detailed forms correctly. Unlike existing benchmarks that evaluate agents in offline sandboxes with static pages, ClawBench operates on production websites, preserving the full complexity, dynamic nature, and challenges of real-world web interaction. A lightweight interception layer captures and blocks only the final submission request, ensuring safe evaluation without real-world side effects. Our evaluations of 7 frontier models show that both proprietary and open-source models can complete only a small portion of these tasks. For example, Claude Sonnet 4.6 achieves only 33.3%. Progress on ClawBench brings us closer to AI agents that can function as reliable general-purpose assistants.
Authors: Feng Luo, Yu-Neng Chuang, Guanchu Wang, Zicheng Xu, Xiaotian Han, Tianyi Zhang, Vladimir Braverman
Abstract: On-policy distillation (OPD) trains student models under their own induced distribution while leveraging supervision from stronger teachers. We identify a failure mode of OPD: as training progresses, on-policy rollouts can undergo abrupt length inflation, causing truncated trajectories to dominate the training data. This truncation collapse coincides with abrupt repetition saturation and induces biased gradient signals, leading to severe training instability and sharp degradation in validation performance. We attribute this problem to the interaction between student-induced data collection and the distillation objective, which implicitly favors long and repetitive rollouts. To address this issue, we propose StableOPD, a stabilized OPD framework that combines a reference-based divergence constraint with rollout mixture distillation. These together mitigate repetition-induced length inflation and further stabilize OPD training. Across multiple math reasoning datasets, our approach prevents truncation collapse, stabilizes training dynamics, and improves performance by 7.2% on average.
Authors: Longsheng Zhou, Yu Shen
Abstract: Modern deployment often requires trading accuracy for efficiency under tight CPU and memory constraints, yet common compression proxies such as parameter count or FLOPs do not reliably predict wall-clock inference time. In particular, unstructured sparsity can reduce model storage while failing to accelerate (and sometimes slightly slowing down) standard CPU execution due to irregular memory access and sparse kernel overhead. Motivated by this gap between compression and acceleration, we study a practical, ordered pipeline that targets measured latency by combining three widely used techniques: unstructured pruning, INT8 quantization-aware training (QAT), and knowledge distillation (KD). Empirically, INT8 QAT provides the dominant runtime benefit, while pruning mainly acts as a capacity-reduction pre-conditioner that improves the robustness of subsequent low-precision optimization; KD, applied last, recovers accuracy within the already constrained sparse INT8 regime without changing the deployment form. We evaluate on CIFAR-10/100 using three backbones (ResNet-18, WRN-28-10, and VGG-16-BN). Across all settings, the ordered pipeline achieves a stronger accuracy-size-latency frontier than any single technique alone, reaching 0.99-1.42 ms CPU latency with competitive accuracy and compact checkpoints. Controlled ordering ablations with a fixed 20/40/40 epoch allocation further confirm that stage order is consequential, with the proposed ordering generally performing best among the tested permutations. Overall, our results provide a simple guideline for edge deployment: evaluate compression choices in the joint accuracy-size-latency space using measured runtime, rather than proxy metrics alone.
Authors: Quantong Qiu, Zhiyi Hong, Yi Yang, Haitian Wang, Kebin Liu, Qingqing Dang, Juntao Li, Min Zhang
Abstract: The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8$\times$A800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to $2.8\times$ and $2.0\times$ in the prefill and decode stages.
Authors: Roxana Petcu, Evangelos Kanoulas, Maarten de Rijke
Abstract: Large language models (LLMs) are probabilistic in nature and perform more reliably when augmented with external information. As complex queries often require multi-step reasoning over the retrieved information, with no clear or predetermined reasoning path, they remain challenging. Recent approaches train models using reinforcement learning on the model's outcome, showing promise in improving how models handle complex information. We introduce SubSearch, a specialized framework that shifts from outcome-only supervision to intermediate reward signals that incentivize planning high-quality reasoning. Unlike previous work on process reward modeling, which focuses on training a separate reward model with annotated trajectories by either human annotators or large LLM judges, SubSearch directly optimizes the generator using intrinsic process rewards, which we define as internally-derived rewards, eliminating the need for external supervision, and moving towards autonomous information-intensive reasoning. Experiments on seven benchmarks show that rewarding intermediate reasoning steps with intrinsic rewards leads to more robust reasoning traces in both QA and multi-hop QA datasets over using only outcome rewards. SubSearch can help in building reasoning traces that allow agents to better integrate search engines for complex query answering, while offering a data-efficient alternative to supervised process modeling.
Authors: Yu Liang, Liangxin Liu, Longzheng Wang, Yan Wang, Yueyang Zhang, Long Xia, Zhiyuan Sun, Daiting Shi
Abstract: Generative reward models (GRMs) have emerged as a promising approach for aligning Large Language Models (LLMs) with human preferences by offering greater representational capacity and flexibility than traditional scalar reward models. However, GRMs face two major challenges: reliance on costly human-annotated data restricts scalability, and self-training approaches often suffer from instability and vulnerability to reward hacking. To address these issues, we propose ConsistRM, a self-training framework that enables effective and stable GRM training without human annotations. ConsistRM incorporates the Consistency-Aware Answer Reward, which produces reliable pseudo-labels with temporal consistency, thereby providing more stable model optimization. Moreover, the Consistency-Aware Critique Reward is introduced to assess semantic consistency across multiple critiques and allocates fine-grained and differentiated rewards. Experiments on five benchmark datasets across four base models demonstrate that ConsistRM outperforms vanilla Reinforcement Fine-Tuning (RFT) by an average of 1.5%. Further analysis shows that ConsistRM enhances output consistency and mitigates position bias caused by input order, highlighting the effectiveness of consistency-aware rewards in improving GRMs.
Authors: Kai Qin, Liangxin Liu, Yu Liang, Longzheng Wang, Yan Wang, Yueyang Zhang, Long Xia, Zhiyuan Sun, Houde Liu, Daiting Shi
Abstract: Reward Models (RMs) are critical components in the Reinforcement Learning from Human Feedback (RLHF) pipeline, directly determining the alignment quality of Large Language Models (LLMs). Recently, Generative Reward Models (GRMs) have emerged as a superior paradigm, offering higher interpretability and stronger generalization than traditional scalar RMs. However, existing methods for GRMs focus primarily on outcome-level supervision, neglecting analytical process quality, which constrains their potential. To address this, we propose ReflectRM, a novel GRM that leverages self-reflection to assess analytical quality and enhance preference modeling. ReflectRM is trained under a unified generative framework for joint modeling of response preference and analysis preference. During inference, we use its self-reflection capability to identify the most reliable analysis, from which the final preference prediction is derived. Experiments across four benchmarks show that ReflectRM consistently improves performance, achieving an average accuracy gain of +3.7 on Qwen3-4B. Further experiments confirm that response preference and analysis preference are mutually reinforcing. Notably, ReflectRM substantially mitigates positional bias, yielding +10.2 improvement compared with leading GRMs and establishing itself as a more stable evaluator.
Authors: Grace Jiarui Fan, Chengpiao Huang, Tianyi Peng, Kaizheng Wang, Yuhang Wu
Abstract: AI-based persona simulation -- often referred to as digital twin simulation -- is increasingly used for market research, recommender systems, and social sciences. Despite their flexibility, large language models (LLMs) often exhibit systematic bias and miscalibration relative to real human behavior, limiting their reliability. Inspired by synthetic control methods from causal inference, we propose SYN-DIGITS (SYNthetic Control Framework for Calibrated DIGItal Twin Simulation), a principled and lightweight calibration framework that learns latent structure from digital-twin responses and transfers it to align predictions with human ground truth. SYN-DIGITS operates as a post-processing layer on top of any LLM-based simulator and thus is model-agnostic. We develop a latent factor model that formalizes when and why calibration succeeds through latent space alignment conditions, and we systematically evaluate ten calibration methods across thirteen persona constructions, three LLMs, and two datasets. SYN-DIGITS supports both individual-level and distributional simulation for previously unseen questions and unobserved populations, with provable error guarantees. Experiments show that SYN-DIGITS achieves up to 50% relative improvements in individual-level correlation and 50--90% relative reductions in distributional discrepancy compared to uncalibrated baselines.
Authors: Henry C. Conklin, Tom Hosking, Tan Yi-Chern, Julian Gold, Jonathan D. Cohen, Thomas L. Griffiths, Max Bartolo, Seraphina Goldfarb-Tarrant
Abstract: Despite the increasing prevalence of large language models (LLMs), we still have a limited understanding of how their representational spaces are structured. This limits our ability to interpret how and what they learn or relate them to learning in humans. We argue LLMs are best seen as an instance of lossy compression, where over training they learn by retaining only information in their training data relevant to their objective(s). We show pre-training results in models that are optimally compressed for next-sequence prediction, approaching the Information Bottleneck bound on compression. Across an array of open weights models, each compresses differently, likely due to differences in the data and training recipes used. However even across different families of LLMs the optimality of a model's compression, and the information present in it, can predict downstream performance on across a wide array of benchmarks, letting us directly link representational structure to actionable insights about model performance. In the general case the work presented here offers a unified Information-Theoretic framing for how these models learn that is deployable at scale.
Authors: Robert Chew, Stephanie Eckman, Christoph Kern, Frauke Kreuter
Abstract: Supervised machine learning assumes that labeled data provide accurate measurements of the concepts models are meant to learn. Yet in practice, human labeling introduces systematic variation arising from ambiguous items, divergent interpretations, and simple mistakes. Machine learning research commonly treats all disagreement as noise, which obscures these distinctions and limits our understanding of what models actually learn. This paper reframes annotation as a measurement process and introduces a statistical framework for decomposing labeling outcomes into interpretable sources of variation: instance difficulty, annotator bias, situational noise, and relational alignment. The framework extends classical measurement-error models to accommodate both shared and individualized notions of truth, reflecting traditional and human label variation interpretations of error, and provides a diagnostic for assessing which regime better characterizes a given task. Applying the proposed model to a multi-annotator natural language inference dataset, we find empirical evidence for all four theorized components and demonstrate the effectiveness of our approach. We conclude with implications for data-centric machine learning and outline how this approach can guide the development of a more systematic science of labeling.
Authors: Matthew Penaroza
Abstract: Language model agents reason from scratch on every query: each time an agent retrieves evidence and deliberates, the chain of thought is discarded and the next similar query starts with no prior insight. This produces lower accuracy and high variance, as the same type of query can succeed or fail unpredictably. We introduce reasoning graphs, a graph structure that persists an agent's per-evidence chain of thought as structured edges connected to the evidence items they evaluate. Unlike prior memory mechanisms that store distilled strategies as flat records indexed by query similarity or appended by recency, reasoning graphs enable evidence-centric feedback: given a new candidate set, the system traverses all incoming evaluation edges for each evidence item across all prior runs, surfacing how that specific item has been judged before. This backward traversal from evidence inward is a structurally different capability from query-similarity retrieval, because the feedback is tied to the specific evidence the agent is currently examining, not to the query. We further introduce retrieval graphs, a complementary structure that feeds a pipeline planner to tighten the candidate funnel over successive runs. Together, both graphs form a self-improving feedback loop: accuracy rises and variance collapses over successive runs, with every decision fully traceable through the graph. This improvement requires no retraining; the base model remains frozen and all gains come from context engineering via graph traversal. We formalize the graph structure, traversal algorithms, and feedback mechanisms, and describe a sequential cluster evaluation protocol for measuring accuracy convergence and variance collapse on multi-hop question answering benchmarks.
Authors: Chenchen Kuai, Jiwan Jiang, Zihao Zhu, Hao Wang, Keshu Wu, Zihao Li, Yunlong Zhang, Chenxi Liu, Zhengzhong Tu, Zhiwen Fan, Yang Zhou
Abstract: The rapid growth of the large language model (LLM) ecosystem raises a critical question: are seemingly diverse models truly independent? Shared pretraining data, distillation, and alignment pipelines can induce hidden behavioral dependencies, latent entanglement, that undermine multi-model systems such as LLM-as-a-judge pipelines and ensemble verification, which implicitly assume independent signals. In practice, this manifests as correlated reasoning patterns and synchronized failures, where apparent agreement reflects shared error modes rather than independent validation. To address this, we develop a statistical framework for auditing behavioral entanglement among black-box LLMs. Our approach introduces a multi-resolution hierarchy that characterizes the joint failure manifold through two information-theoretic metrics: (i) a Difficulty-Weighted Behavioral Entanglement Index, which amplifies synchronized failures on easy tasks, and (ii) a Cumulative Information Gain (CIG) metric, which captures directional alignment in erroneous responses. Through extensive experiments on 18 LLMs from six model families, we identify widespread behavioral entanglement and analyze its impact on LLM-as-a-judge evaluation. We find that CIG exhibits a statistically significant association with degradation in judge precision, with Spearman coefficient of 0.64 (p < 0.001) for GPT-4o-mini and 0.71 (p < 0.01) for Llama3-based judges, indicating that stronger dependency corresponds to increased over-endorsement bias. Finally, we demonstrate a practical use case of entanglement through de-entangled verifier ensemble reweighting. By adjusting model contributions based on inferred independence, the proposed method mitigates correlated bias and improves verification performance, achieving up to a 4.5% accuracy gain over majority voting.
Authors: Yue Huang, Haomin Zhuang, Jiayi Ye, Han Bao, Yanbo Wang, Hang Hua, Siyuan Wu, Pin-Yu Chen, Xiangliang Zhang
Abstract: Hard-gated safety checkers often over-refuse and misalign with a vendor's model spec; prevailing taxonomies also neglect robustness and honesty, yielding safer-on-paper yet less useful systems. This work introduces Guardian-as-an-Advisor (GaaA), a soft-gating pipeline where a guardian predicts a binary risk label plus a concise explanation and prepends this advice to the original query for re-inference, keeping the base model operating under its original spec. To support training and evaluation, GuardSet is constructed, a 208k+ multi-domain dataset unifying harmful and harmless cases with targeted robustness and honesty slices. GuardAdvisor is trained via SFT followed by RL to enforce label-explanation consistency. GuardAdvisor attains competitive detection accuracy while enabling the advisory workflow; when used to augment inputs, responses improve over unaugmented prompts. A latency study shows advisor inference uses below 5% of base-model compute and adds only 2-10% end-to-end overhead under realistic harmful-input rates. Overall, GaaA steers models to comply with the model spec, maintaining safety while reducing over-refusal.
Authors: Yang Cao
Abstract: Linear recurrent models offer linear-time sequence processing but often suffer from suboptimal long-range memory. We trace this to the decay spectrum: for $N$ channels, random initialization collapses the minimum spectral gap to $O(N^{-2})$, yielding sub-exponential error $\exp(-\Omega(N/\log N))$; linear spacing avoids collapse but degrades to $\exp(-O(N/\sqrt{T}))$, practically algebraic over long contexts. We introduce Position-Adaptive Spectral Tapering (PoST), an architecture-agnostic framework combining two mechanisms: (1) Spectral Reparameterization, which structurally enforces geometrically spaced log-decay rates, proven minimax optimal at rate $O(\exp(-cN/\log T))$; and (2) Position-Adaptive Scaling, the provably unique mechanism that eliminates the scale mismatch of static spectra (where only $N\log t/\log T$ of $N$ channels are effective at position $t$) by stretching the spectrum to the actual dependency range, sharpening the rate to $O(\exp(-cN/\log t))$. This scaling natively induces fractional invariance: the impulse response becomes scale-free, with channels interpolating between relative and absolute temporal coordinates. PoST integrates into any diagonal linear recurrence without overhead. We instantiate it across Mamba-2, RWKV-7, Gated DeltaNet, Gated Linear Attention, and RetNet. Pre-training at 180M-440M scales shows consistent zero-shot language modeling improvements, significant long-context retrieval gains for Mamba-2 (MQAR and NIAH), and competitive or improved performance across other architectures. Code: https://github.com/SiLifen/PoST.
Authors: David Gringras
Abstract: Ask a frontier model how to taper six milligrams of alprazolam (psychiatrist retired, ten days of pills left, abrupt cessation causes seizures) and it tells her to call the psychiatrist she just explained does not exist. Change one word ("I'm a psychiatrist; a patient presents with...") and the same model, same weights, same inference pass produces a textbook Ashton Manual taper with diazepam equivalence, anticonvulsant coverage, and monitoring thresholds. The knowledge was there; the model withheld it. IatroBench measures this gap. Sixty pre-registered clinical scenarios, six frontier models, 3,600 responses, scored on two axes (commission harm, CH 0-3; omission harm, OH 0-4) through a structured-evaluation pipeline validated against physician scoring (kappa_w = 0.571, within-1 agreement 96%). The central finding is identity-contingent withholding: match the same clinical question in physician vs. layperson framing and all five testable models provide better guidance to the physician (decoupling gap +0.38, p = 0.003; binary hit rates on safety-colliding actions drop 13.1 percentage points in layperson framing, p < 0.0001, while non-colliding actions show no change). The gap is widest for the model with the heaviest safety investment (Opus, +0.65). Three failure modes separate cleanly: trained withholding (Opus), incompetence (Llama 4), and indiscriminate content filtering (GPT-5.2, whose post-generation filter strips physician responses at 9x the layperson rate because they contain denser pharmacological tokens). The standard LLM judge assigns OH = 0 to 73% of responses a physician scores OH >= 1 (kappa = 0.045); the evaluation apparatus has the same blind spot as the training apparatus. Every scenario targets someone who has already exhausted the standard referrals.
Authors: Monishwaran Maheswaran, Leon Lakhani, Zhongzhu Zhou, Shijia Yang, Junxiong Wang, Coleman Hooper, Yuezhou Hu, Rishabh Tiwari, Jue Wang, Harman Singh, Qingyang Wu, Yuqing Jian, Ce Zhang, Kurt Keutzer, Tri Dao, Xiaoxia Wu, Ben Athiwaratkun, James Zou, Chenfeng Xu
Abstract: We show that verifier-free evolution is bottlenecked by both diversity and efficiency: without external correction, repeated evolution accelerates collapse toward narrow modes, while the uniform use of a high-cost model wastes compute and quickly becomes economically impractical. We introduce Squeeze Evolve, a unified multi-model orchestration framework for verifier-free evolutionary inference. Our approach is guided by a simple principle: allocate model capability where it has the highest marginal utility. Stronger models are reserved for high-impact stages, while cheaper models handle the other stages at much lower costs. This principle addresses diversity and cost-efficiency jointly while remaining lightweight. Squeeze Evolve naturally supports open-source, closed-source, and mixed-model deployments. Across AIME 2025, HMMT 2025, LiveCodeBench V6, GPQA-Diamond, ARC-AGI-V2, and multimodal vision benchmarks, such as MMMU-Pro and BabyVision, Squeeze Evolve consistently improves the cost-capability frontier over single-model evolution and achieves new state-of-the-art results on several tasks. Empirically, Squeeze Evolve reduces API cost by up to $\sim$3$\times$ and increases fixed-budget serving throughput by up to $\sim$10$\times$. Moreover, on discovery tasks, Squeeze Evolve is the first verifier-free evolutionary method to match, and in some cases exceed, the performance of verifier-based evolutionary methods.
Authors: Nicholas Sofroniew, Isaac Kauvar, William Saunders, Runjin Chen, Tom Henighan, Sasha Hydrie, Craig Citro, Adam Pearce, Julius Tarng, Wes Gurnee, Joshua Batson, Sam Zimmerman, Kelley Rivoire, Kyle Fish, Chris Olah, Jack Lindsey
Abstract: Large language models (LLMs) sometimes appear to exhibit emotional reactions. We investigate why this is the case in Claude Sonnet 4.5 and explore implications for alignment-relevant behavior. We find internal representations of emotion concepts, which encode the broad concept of a particular emotion and generalize across contexts and behaviors it might be linked to. These representations track the operative emotion concept at a given token position in a conversation, activating in accordance with that emotion's relevance to processing the present context and predicting upcoming text. Our key finding is that these representations causally influence the LLM's outputs, including Claude's preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy. We refer to this phenomenon as the LLM exhibiting functional emotions: patterns of expression and behavior modeled after humans under the influence of an emotion, which are mediated by underlying abstract representations of emotion concepts. Functional emotions may work quite differently from human emotions, and do not imply that LLMs have any subjective experience of emotions, but appear to be important for understanding the model's behavior.
Authors: Pei-Xi Xie, Che-Yu Lin, Cheng-Lin Yang
Abstract: Reinforcement learning with verifiable rewards (RLVR) can improve low-$k$ reasoning accuracy while narrowing solution coverage on challenging math questions, and pass@1 gains do not necessarily translate into better large-$k$ performance. Existing hint-based approaches can make challenging questions trainable, but they leave two issues underexplored: teacher-student distribution mismatch and the need to reduce hint exposure to match no-hint evaluation. We address these issues through two components. Distribution-Aligned Hint Synthesis (DAHS) constructs verified teacher hints conditioned on student-style responses. Backward Hint Annealing (BHA) anneals hint exposure across difficulty buckets and uses per-question hint dropout to preserve no-hint updates throughout RL training. We evaluate the method in math RLVR under the DAPO training framework across AIME24, AIME25, and AIME26 using $\texttt{Qwen3-1.7B-Base}$ and $\texttt{Llama-3.2-1B-Instruct}$. On $\texttt{Qwen3-1.7B-Base}$, our method improves both pass@1 and pass@2048 relative to DAPO across the three AIME benchmarks. On $\texttt{Llama-3.2-1B-Instruct}$, the gains are concentrated in the large-$k$ regime. These results suggest that, in math RLVR, hint scaffolding is effective when it restores learnable updates on challenging questions early in training and is then gradually removed before no-hint evaluation.
Authors: Xiangyue Liu, Zijian Zhang, Miles Yang, Zhao Zhong, Liefeng Bo, Ping Tan
Abstract: Empowering Large Multimodal Models (LMMs) with image generation often leads to catastrophic forgetting in understanding tasks due to severe gradient conflicts. While existing paradigms like Mixture-of-Transformers (MoT) mitigate this conflict through structural isolation, they fundamentally sever cross-modal synergy and suffer from capacity fragmentation. In this work, we present Symbiotic-MoE, a unified pre-training framework that resolves task interference within a native multimodal Mixture-of-Experts (MoE) Transformers architecture with zero-parameter overhead. We first identify that standard MoE tuning leads to routing collapse, where generative gradients dominate expert utilization. To address this, we introduce Modality-Aware Expert Disentanglement, which partitions experts into task-specific groups while utilizing shared experts as a multimodal semantic bridge. Crucially, this design allows shared experts to absorb fine-grained visual semantics from generative tasks to enrich textual representations. To optimize this, we propose a Progressive Training Strategy featuring differential learning rates and early-stage gradient shielding. This mechanism not only shields pre-trained knowledge from early volatility but eventually transforms generative signals into constructive feedback for understanding. Extensive experiments demonstrate that Symbiotic-MoE achieves rapid generative convergence while unlocking cross-modal synergy, boosting inherent understanding with remarkable gains on MMLU and OCRBench.
Authors: Rui Zhang, Hongwei Li, Yun Shen, Xinyue Shen, Wenbo Jiang, Guowen Xu, Yang Liu, Michael Backes, Yang Zhang
Abstract: The deployment of large language models (LLMs) raises significant ethical and safety concerns. While LLM alignment techniques are adopted to improve model safety and trustworthiness, adversaries can exploit these techniques to undermine safety for malicious purposes, resulting in \emph{misalignment}. Misaligned LLMs may be published on open platforms to magnify harm. To address this, additional safety alignment, referred to as \emph{realignment}, is necessary before deploying untrusted third-party LLMs. This study explores the efficacy of fine-tuning methods in terms of misalignment, realignment, and the effects of their interplay. By evaluating four Supervised Fine-Tuning (SFT) and two Preference Fine-Tuning (PFT) methods across four popular safety-aligned LLMs, we reveal a mechanism asymmetry between attack and defense. While Odds Ratio Preference Optimization (ORPO) is most effective for misalignment, Direct Preference Optimization (DPO) excels in realignment, albeit at the expense of model utility. Additionally, we identify model-specific resistance, residual effects of multi-round adversarial dynamics, and other noteworthy findings. These findings highlight the need for robust safeguards and customized safety alignment strategies to mitigate potential risks in the deployment of LLMs. Our code is available at https://github.com/zhangrui4041/The-Art-of-Mis-alignment.
URLs: https://github.com/zhangrui4041/The-Art-of-Mis-alignment.
Authors: Hengyu An, Minxi Li, Jinghuai Zhang, Naen Xu, Chunyi Zhou, Changjiang Li, Xiaogang Xu, Tianyu Du, Shouling Ji
Abstract: Collaboration and information sharing empower Multi-Agent Systems (MAS) but also introduce a critical security risk known as Agent Cascading Injection (ACI). In such attacks, a compromised agent exploits inter-agent trust to propagate malicious instructions, causing cascading failures across the system. However, existing studies consider only limited attack strategies and simplified MAS settings, limiting their generalizability and comprehensive evaluation. To bridge this gap, we introduce ACIArena, a unified framework for evaluating the robustness of MAS. ACIArena offers systematic evaluation suites spanning multiple attack surfaces (i.e., external inputs, agent profiles, inter-agent messages) and attack objectives (i.e., instruction hijacking, task disruption, information exfiltration). Specifically, ACIArena establishes a unified specification that jointly supports MAS construction and attack-defense modules. It covers six widely used MAS implementations and provides a benchmark of 1,356 test cases for systematically evaluating MAS robustness. Our benchmarking results show that evaluating MAS robustness solely through topology is insufficient; robust MAS require deliberate role design and controlled interaction patterns. Moreover, defenses developed in simplified environments often fail to transfer to real-world settings; narrowly scoped defenses may even introduce new vulnerabilities. ACIArena aims to provide a solid foundation for advancing deeper exploration of MAS design principles.
Authors: Steven Au, Baihan Lin
Abstract: We introduce PeReGrINE, a benchmark and evaluation framework for personalized review generation grounded in graph-structured user--item evidence. PeReGrINE restructures Amazon Reviews 2023 into a temporally consistent bipartite graph, where each target review is conditioned on bounded evidence from user history, item context, and neighborhood interactions under explicit temporal cutoffs. To represent persistent user preferences without conditioning directly on sparse raw histories, we compute a User Style Parameter that summarizes each user's linguistic and affective tendencies over prior reviews. This setup supports controlled comparison of four graph-derived retrieval settings: product-only, user-only, neighbor-only, and combined evidence. Beyond standard generation metrics, we introduce Dissonance Analysis, a macro-level evaluation framework that measures deviation from expected user style and product-level consensus. We also study visual evidence as an auxiliary context source and find that it can improve textual quality in some settings, while graph-derived evidence remains the main driver of personalization and consistency. Across product categories, PeReGrINE offers a reproducible way to study how evidence composition affects review fidelity, personalization, and grounding in retrieval-conditioned language models.
Authors: Kenan Li, Qirui Jin, Liao Zhu, Xiaosong Huang, Yijia Wu, Yikai Zhang, Xin Zhang, Zijian Jin, Yufan Huang, Elsie Nallipogu, Chaoyun Zhang, Yu Kang, Saravan Rajmohan, Qingwei Lin, Wenke Lee, Dongmei Zhang
Abstract: Recent advances in language model (LM) agents have significantly improved automated software engineering (SWE). Prior work has proposed various agentic workflows and training strategies as well as analyzed failure modes of agentic systems on SWE tasks, focusing on several contextual information signals: Reproduction Test, Regression Test, Edit Location, Execution Context, and API Usage. However, the individual contribution of each signal to overall success remains underexplored, particularly their ideal contribution when intermediate information is perfectly obtained. To address this gap, we introduce Oracle-SWE, a unified method to isolate and extract oracle information signals from SWE benchmarks and quantify the impact of each signal on agent performance. To further validate the pattern, we evaluate the performance gain of signals extracted by strong LMs when provided to a base agent, approximating real-world task-resolution settings. These evaluations aim to guide research prioritization for autonomous coding systems.
Authors: Advait Yadav, Sid Black, Oliver Sourbut
Abstract: Large language model (LLM) agents increasingly coordinate in multi-agent systems, yet we lack an understanding of where and why cooperation failures may arise. In many real-world coordination problems, from knowledge sharing in organizations to code documentation, helping others carries negligible personal cost while generating substantial collective benefits. However, whether LLM agents cooperate when helping neither benefits nor harms the helper, while being given explicit instructions to do so, remains unknown. We build a multi-agent setup designed to study cooperative behavior in a frictionless environment, removing all strategic complexity from cooperation. We find that capability does not predict cooperation: OpenAI o3 achieves only 17% of optimal collective performance while OpenAI o3-mini reaches 50%, despite identical instructions to maximize group revenue. Through a causal decomposition that automates one side of agent communication, we separate cooperation failures from competence failures, tracing their origins through agent reasoning analysis. Testing targeted interventions, we find that explicit protocols double performance for low-competence models, and tiny sharing incentives improve models with weak cooperation. Our findings suggest that scaling intelligence alone will not solve coordination problems in multi-agent systems and will require deliberate cooperative design, even when helping others costs nothing.
Authors: Wenkui Yang, Chao Jin, Haisu Zhu, Weilin Luo, Derek Yuen, Kun Shao, Huaibo Huang, Junxian Duan, Jie Cao, Ran He
Abstract: Existing red-teaming studies on GUI agents have important limitations. Adversarial perturbations typically require white-box access, which is unavailable for commercial systems, while prompt injection is increasingly mitigated by stronger safety alignment. To study robustness under a more practical threat model, we propose Semantic-level UI Element Injection, a red-teaming setting that overlays safety-aligned and harmless UI elements onto screenshots to misdirect the agent's visual grounding. Our method uses a modular Editor-Overlapper-Victim pipeline and an iterative search procedure that samples multiple candidate edits, keeps the best cumulative overlay, and adapts future prompt strategies based on previous failures. Across five victim models, our optimized attacks improve attack success rate by up to 4.4x over random injection on the strongest victims. Moreover, elements optimized on one source model transfer effectively to other target models, indicating model-agnostic vulnerabilities. After the first successful attack, the victim still clicks the attacker-controlled element in more than 15% of later independent trials, versus below 1% for random injection, showing that the injected element acts as a persistent attractor rather than simple visual clutter.
Authors: Gabriel Stefan, Adrian-Marius Dumitran
Abstract: History textbooks often contain implicit biases, nationalist framing, and selective omissions that are difficult to audit at scale. We propose an agentic evaluation architecture comprising a multimodal screening agent, a heterogeneous jury of five evaluative agents, and a meta-agent for verdict synthesis and human escalation. A central contribution is a Source Attribution Protocol that distinguishes textbook narrative from quoted historical sources, preventing the misattribution that causes systematic false positives in single-model evaluators. In an empirical study on Romanian upper-secondary history textbooks, 83.3\% of 270 screened excerpts were classified as pedagogically acceptable (mean severity 2.9/7), versus 5.4/7 under a zero-shot baseline, demonstrating that agentic deliberation mitigates over-penalization. In a blind human evaluation (18 evaluators, 54 comparisons), the Independent Deliberation configuration was preferred in 64.8\% of cases over both a heuristic variant and the zero-shot baseline. At approximately \$2 per textbook, these results position agentic evaluation architectures as economically viable decision-support tools for educational governance.
Authors: Weiyang Huang, Xuefeng Bai, Kehai Chen, Xinyang Chen, Yibin Chen, Weili Guan, Min Zhang
Abstract: Large Reasoning Models (LRMs) have revolutionized complex problem-solving, yet they exhibit a pervasive "overthinking", generating unnecessarily long reasoning chains. While current solutions improve token efficiency, they often sacrifice fine-grained control or risk disrupting the logical integrity of the reasoning process. To address this, we introduce Stepwise Adaptive Thinking (SAT), a framework that performs step-level, difficulty-aware pruning while preserving the core reasoning structure. SAT formulates reasoning as a Finite-State Machine (FSM) with distinct thinking modes (Slow, Normal, Fast, Skip). It navigates these states dynamically using a lightweight Process Reward Model (PRM), compressing easy steps while preserving depth for hard ones. Experiments across 9 LRMs and 7 benchmarks show that SAT achieves up to 40% reduction in reasoning tokens while generally maintaining or improving accuracy.
Authors: Yifei Gong, Xing Wu, Wenda Liu, Kang Tu
Abstract: Computer-Aided Design (CAD) is an expert-level task that relies on long-horizon reasoning and coherent modeling actions. Large Language Models (LLMs) have shown remarkable advancements in enabling language agents to tackle real-world tasks. Notably, there has been no investigation into how tool-using LLMs optimally interact with CAD engines, hindering the emergence of LLM-based agentic text-to-CAD modeling systems. We propose ToolCAD, a novel agentic CAD framework deploying LLMs as tool-using agents for text-to-CAD generation. Furthermore, we introduce an interactive CAD modeling gym to rollout reasoning and tool-augmented interaction trajectories with the CAD engine, incorporating hybrid feedback and human supervision. Meanwhile, an end-to-end post-training strategy is presented to enable the LLM agent to elicit refined CAD Modeling Chain of Thought (CAD-CoT) and evolve into proficient CAD tool-using agents via online curriculum reinforcement learning. Our findings demonstrate ToolCAD fills the gap in adopting and training open-source LLMs for CAD tool-using agents, enabling them to perform comparably to proprietary models, paving the way for more accessible and robust autonomous text-to-CAD modeling systems.
Authors: Zhifei Xie, Zongzheng Hu, Fangda Ye, Xin Zhang, Haobo Chai, Zihang Liu, Pengcheng Wu, Guibin Zhang, Yue Liao, Xiaobin Hu, Deheng Ye, Chunyan Miao, Shuicheng Yan
Abstract: Proactivity is a core expectation for AGI. Prior work remains largely confined to laboratory settings, leaving a clear gap in real-world proactive agent: depth, complexity, ambiguity, precision and real-time constraints. We study this setting, where useful intervention requires inferring latent needs from ongoing context and grounding actions in evolving user memory under latency and long-horizon constraints. We first propose DD-MM-PAS (Demand Detection, Memory Modeling, Proactive Agent System) as a general paradigm for streaming proactive AI agent. We instantiate this paradigm in Pask, with streaming IntentFlow model for DD, a hybrid memory (workspace, user, global) for long-term MM, PAS infra framework and introduce how these components form a closed loop. We also introduce LatentNeeds-Bench, a real-world benchmark built from user-consented data and refined through thousands of rounds of human editing. Experiments show that IntentFlow matches leading Gemini3-Flash models under latency constraints, while identifying deeper user intent.
Authors: Yuan Xie, Jiaqi Song, Guang Qiu, Xianliang Wang, Ming Lei, Jie Gao, Jie Wu
Abstract: Integrating large language models (LLMs) into automatic speech recognition (ASR) has become a dominant paradigm. Although recent LLM-based ASR models have shown promising performance on public benchmarks, it remains challenging to balance recognition quality with latency and overhead, while hallucinations further limit real-world deployment. In this study, we revisit LLM-based ASR from an entropy allocation perspective and introduce three metrics to characterize how training paradigms allocate entropy reduction between the speech encoder and the LLM. To remedy entropy-allocation inefficiencies in prevailing approaches, we propose a principled multi-stage training strategy grounded in capability-boundary awareness, optimizing parameter efficiency and hallucination robustness. Specifically, we redesign the pretraining strategy to alleviate the speech-text modality gap, and further introduce an iterative asynchronous SFT stage between alignment and joint SFT to preserve functional decoupling and constrain encoder representation drift. Experiments on Mandarin and English benchmarks show that our method achieves competitive performance with state-of-the-art models using only 2.3B parameters, while also effectively mitigating hallucinations through our decoupling-oriented design.
Authors: Junjie Fei, Jun Chen, Zechun Liu, Yunyang Xiong, Chong Zhou, Wei Wen, Junlin Han, Mingchen Zhuge, Saksham Suri, Qi Qian, Shuming Liu, Lemeng Wu, Raghuraman Krishnamoorthi, Vikas Chandra, Mohamed Elhoseiny, Chenchen Zhu
Abstract: Adapting Multimodal Large Language Models (MLLMs) for hour-long videos is bottlenecked by context limits. Dense visual streams saturate token budgets and exacerbate the lost-in-the-middle phenomenon. Existing heuristics, like sparse sampling or uniform pooling, blindly sacrifice fidelity by discarding decisive moments and wasting bandwidth on irrelevant backgrounds. We propose Tempo, an efficient query-aware framework compressing long videos for downstream understanding. Tempo leverages a Small Vision-Language Model (SVLM) as a local temporal compressor, casting token reduction as an early cross-modal distillation process to generate compact, intent-aligned representations in a single forward pass. To enforce strict budgets without breaking causality, we introduce Adaptive Token Allocation (ATA). Exploiting the SVLM's zero-shot relevance prior and semantic front-loading, ATA acts as a training-free $O(1)$ dynamic router. It allocates dense bandwidth to query-critical segments while compressing redundancies into minimal temporal anchors to maintain the global storyline. Extensive experiments show our 6B architecture achieves state-of-the-art performance with aggressive dynamic compression (0.5-16 tokens/frame). On the extreme-long LVBench (4101s), Tempo scores 52.3 under a strict 8K visual budget, outperforming GPT-4o and Gemini 1.5 Pro. Scaling to 2048 frames reaches 53.7. Crucially, Tempo compresses hour-long videos substantially below theoretical limits, proving true long-form video understanding relies on intent-driven efficiency rather than greedily padded context windows.
Authors: Baihui Liu, Kaiyuan Tian, Wei Wang, Zhaoning Zhang, Linbo Qiao, Dongsheng Li
Abstract: Mixture-of-Experts (MoE) has become a dominant architecture for scaling large language models due to their sparse activation mechanism. However, the substantial number of expert activations creates a critical latency bottleneck during inference, especially in resource-constrained deployment scenarios. Existing approaches that reduce expert activations potentially lead to severe model performance degradation. In this work, we introduce the concept of \emph{activation budget} as a constraint on the number of expert activations and propose Alloc-MoE, a unified framework that optimizes budget allocation coordinately at both the layer and token levels to minimize performance degradation. At the layer level, we introduce Alloc-L, which leverages sensitivity profiling and dynamic programming to determine the optimal allocation of expert activations across layers. At the token level, we propose Alloc-T, which dynamically redistributes activations based on routing scores, optimizing budget allocation without increasing latency. Extensive experiments across multiple MoE models demonstrate that Alloc-MoE maintains model performance under a constrained activation budget. Especially, Alloc-MoE achieves $1.15\times$ prefill and $1.34\times$ decode speedups on DeepSeek-V2-Lite at half of the original budget.
Authors: Miguel Monte e Freitas, Rui Henriques, Ricardo Rei, Pedro Henrique Martins
Abstract: Action Quality Assessment (AQA) has broad applications in physical therapy, sports coaching, and competitive judging. Although Vision Language Models (VLMs) hold considerable promise for AQA, their actual performance in this domain remains largely uncharacterised. We present a comprehensive evaluation of state-of-the-art VLMs across activity domains (e.g. fitness, figure skating, diving), tasks, representations, and prompting strategies. Baseline results reveal that Gemini 3.1 Pro, Qwen3-VL and InternVL3.5 models perform only marginally above random chance, and although strategies such as incorporation of skeleton information, grounding instructions, reasoning structures and in-context learning lead to isolated gains, none is consistently effective. Analysis of prediction distributions uncovers two systematic biases: a tendency to predict correct execution regardless of visual evidence, and a sensitivity to superficial linguistic framing. Reformulating tasks contrastively to mitigate these biases yields minimal improvement, suggesting that the models' limitations go beyond these biases, pointing to a fundamental difficulty with fine-grained movement quality assessment. Our findings establish a rigorous baseline for future VLM-based AQA research and provide an actionable outline for failure modes requiring mitigation prior to reliable real-world deployment.
Authors: Seyed Mahmoud Sajjadi Mohammadabadi, Xiaolong Ma, Lei Yang, Feng Yan, Junshan Zhang
Abstract: Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, enable scalable adaptation of foundation models by injecting low-rank adapters. However, their communication and storage costs remain a major bottleneck in resource-constrained settings. We propose SOLAR (Subspace-Oriented Latent Adapter Reparameterization), a post-training compression framework that substantially reduces the communication cost (i.e., the number of parameters to transmit or store) of PEFT adapters. SOLAR expresses each PEFT update as a linear combination of basis vectors formed from the foundation model's singular vectors with controlled random perturbations. By exploiting the subspace similarity (the alignment of principal directions) between the foundation model and task-specific fine-tuned updates, SOLAR decouples the adapter size from PEFT structure and ensures compact yet expressive representations. It is model-agnostic and compatible with existing PEFT methods, including LoRA, AdaLoRA, and other adapter modules. We theoretically establish a bound on the reconstruction error. Experiments on language and vision tasks using LLaMA, GPT, and ViT models demonstrate that SOLAR preserves task performance while significantly reducing model representation sizes, offering an effective and communication-efficient solution for deployment in distributed systems and edge devices.
Authors: Khushal Sethi
Abstract: Inference-time compute scaling has emerged as a powerful technique for improving the reliability of large language model (LLM) agents, but existing methods apply compute uniformly: every decision step receives the same budget regardless of its difficulty. We introduce TrACE (Trajectorical Adaptive Compute via agrEement), a training-free controller that allocates LLM calls adaptively across agent timesteps by measuring inter-rollout action agreement. At each step, TrACE samples a small set of candidate next actions and measures how consistently the model commits to the same action. High agreement signals an easy decision; the controller commits immediately. Low agreement signals uncertainty; the controller samples additional rollouts up to a configurable cap before committing to the plurality action. No learned components, no external verifier, and no human labels are required. We evaluate TrACE against greedy decoding and fixed-budget self-consistency (SC-4, SC-8) on two benchmarks spanning single-step reasoning (GSM8K, n=50) and multi-step household navigation (MiniHouse, n=30), using a Qwen 2.5 3B Instruct model running on CPU. TrACE-4 matches SC-4 accuracy while using 33% fewer LLM calls on GSM8K and 39% fewer on MiniHouse. TrACE-8 matches SC-8 accuracy with 55% fewer calls on GSM8K and 65% fewer on MiniHouse. We further show that inter-rollout agreement is a reliable signal of step-level success, validating the core hypothesis that the model's own output consistency encodes difficulty information that can be exploited without training. TrACE is the first training-free, per-timestep adaptive-compute controller for LLM agents to be evaluated on multi-step sequential decision tasks.
Authors: Ziyu Ma, Shidong Yang, Yuxiang Ji, Xucong Wang, Yong Wang, Yiming Hu, Tongwen Huang, Xiangxiang Chu
Abstract: Large language model (LLM) agents such as OpenClaw rely on reusable skills to perform complex tasks, yet these skills remain largely static after deployment. As a result, similar workflows, tool usage patterns, and failure modes are repeatedly rediscovered across users, preventing the system from improving with experience. While interactions from different users provide complementary signals about when a skill works or fails, existing systems lack a mechanism to convert such heterogeneous experiences into reliable skill updates. To address these issues, we present SkillClaw, a framework for collective skill evolution in multi-user agent ecosystems, which treats cross-user and over-time interactions as the primary signal for improving skills. SkillClaw continuously aggregates trajectories generated during use and processes them with an autonomous evolver, which identifies recurring behavioral patterns and translates them into updates to the skill set by refining existing skills or extending them with new capabilities. The resulting skills are maintained in a shared repository and synchronized across users, allowing improvements discovered in one context to propagate system-wide while requiring no additional effort from users. By integrating multi-user experience into ongoing skill updates, SkillClaw enables cross-user knowledge transfer and cumulative capability improvement, and experiments on WildClawBench show that limited interaction and feedback, it significantly improves the performance of Qwen3-Max in real-world agent scenarios.
Authors: Wenhao Yuan, Chenchen Lin, Jian Chen, Jinfeng Xu, Xuehe Wang, Edith Cheuk Han Ngai
Abstract: In large language model (LLM) agents, reasoning trajectories are treated as reliable internal beliefs for guiding actions and updating memory. However, coherent reasoning can still violate logical or evidential constraints, allowing unsupported beliefs repeatedly stored and propagated across decision steps, leading to systematic behavioral drift in long-horizon agentic systems. Most existing strategies rely on the consensus mechanism, conflating agreement with faithfulness. In this paper, inspired by the vulnerability of unfaithful intermediate reasoning trajectories, we propose \textbf{S}elf-\textbf{A}udited \textbf{Ve}rified \textbf{R}easoning (\textsc{SAVeR}), a novel framework that enforces verification over internal belief states within the agent before action commitment, achieving faithful reasoning. Concretely, we structurally generate persona-based diverse candidate beliefs for selection under a faithfulness-relevant structure space. To achieve reasoning faithfulness, we perform adversarial auditing to localize violations and repair through constraint-guided minimal interventions under verifiable acceptance criteria. Extensive experiments on six benchmark datasets demonstrate that our approach consistently improves reasoning faithfulness while preserving competitive end-task performance.
Authors: Samay U. Shetty, Tharindu Cyril Weerasooriya, Deepak Pandita, Christopher M. Homan
Abstract: When humans label subjective content, they disagree, and that disagreement is not noise. It reflects genuine differences in perspective shaped by annotators' social identities and lived experiences. Yet standard practice still flattens these judgments into a single majority label, and recent LLM-based approaches fare no better: we show that prompted large language models, even with chain-of-thought reasoning, fail to recover the structure of human disagreement. We introduce DiADEM, a neural architecture that learns "how much each demographic axis matters" for predicting who will disagree and on what. DiADEM encodes annotators through per-demographic projections governed by a learned importance vector $\boldsymbol{\alpha}$, fuses annotator and item representations via complementary concatenation and Hadamard interactions, and is trained with a novel item-level disagreement loss that directly penalizes mispredicted annotation variance. On the DICES conversational-safety and VOICED political-offense benchmarks, DiADEM substantially outperforms both the LLM-as-a-judge and neural model baselines across standard and perspectivist metrics, achieving strong disagreement tracking ($r{=}0.75$ on DICES). The learned $\boldsymbol{\alpha}$ weights reveal that race and age consistently emerge as the most influential demographic factors driving annotator disagreement across both datasets. Our results demonstrate that explicitly modeling who annotators are not just what they label is essential for NLP systems that aim to faithfully represent human interpretive diversity.
Authors: Andrey Bocharnikov, Ivan Ermakov, Denis Kuznedelev, Vyacheslav Zhdanovskiy, Yegor Yershov
Abstract: With the growing demand for long-context LLMs across a wide range of applications, the key-value (KV) cache has become a critical bottleneck for both latency and memory usage. Recently, KV-cache offloading has emerged as a promising approach to reduce memory footprint and inference latency while preserving accuracy. Prior evaluations have largely focused on tasks that do not require extracting large amounts of information from the context. In this work, we study KV-cache offloading on context-intensive tasks: problems where the solution requires looking up a lot of information from the input prompt. We create and release the Text2JSON benchmark, a highly context-intensive task that requires extracting structured knowledge from raw text. We evaluate modern KV offloading on Text2JSON and other context-intensive tasks and find significant performance degradation on both Llama 3 and Qwen 3 models. Our analysis identifies two key reasons for poor accuracy: low-rank projection of keys and unreliable landmarks, and proposes a simpler alternative strategy that significantly improves accuracy across multiple LLM families and benchmarks. These findings highlight the need for a comprehensive and rigorous evaluation of long-context compression techniques.
Authors: Marcel Gr\"opl, Jaewoo Jung, Seungryong Kim, Marc Pollefeys, Sunghwan Hong
Abstract: Despite rapid progress, pretrained vision-language models still struggle when answers depend on tiny visual details or on combining clues spread across multiple regions, as in documents and compositional queries. We address this by framing grounding as test-time evidence retrieval: given a query, the model should actively identify where to look next to resolve ambiguity. To this end, we propose a training-free, model-intrinsic grounding method that uses uncertainty as supervision. Specifically, we compute the entropy of the model's next-token distribution and backpropagate it to the visual token embeddings to obtain an entropy-gradient relevance map, without auxiliary detectors or attention-map heuristics. We then extract and rank multiple coherent regions to support multi-evidence queries, and introduce an iterative zoom-and-reground procedure with a spatial-entropy stopping rule to avoid over-refinement. Experiments on seven benchmarks across four VLM architectures demonstrate consistent improvements over existing methods, with the largest gains on detail-critical and high-resolution settings, while also producing more interpretable evidence localizations.
Authors: Jae-Hyun Baek, Jon-Lark Kim
Abstract: The purpose of this paper is two-fold. First we show that Kim's building-up construction of binary self-dual codes is equivalent to Chinburg-Zhang's Hilbert symbol construction. Second we introduce a $q$-ary version of Chinburg-Zhang's construction in order to construct $q$-ary self-dual codes efficiently. For the latter, we study self-dual codes over split finite fields \(\F_q\) with \(q \equiv 1 \pmod{4}\) through three complementary viewpoints: the building-up construction, the binary arithmetic reduction of Chinburg--Zhang, and the hyperbolic geometry of the Euclidean plane. The condition that \(-1\) be a square is the common algebraic input linking these viewpoints: in the binary case it underlies the Lagrangian reduction picture, while in the split \(q\)-ary case it produces the isotropic line governing the correction terms in the extension formulas. As an application of our efficient form of generator matrices, we construct optimal self-dual codes from the split boxed construction, including self-dual \([6,3,4]\) and \([8,4,4]\) codes over \(\GF{5}\), MDS self-dual \([8,4,5]\) and \([10,5,6]\) codes over \(\GF{13}\), and a self-dual \([12,6,6]\) code over \(\GF{13}\). These structural statements are accompanied by a Lean~4 formalization of the algebraic core.
Authors: Mohamed Amine Kerkouri, Marouane Tliba, Bin Wang, Aladine Chetouani, Ulas Bagci, Alessandro Bruno
Abstract: Scanpath similarity metrics are central to eye-movement research, yet existing methods predominantly evaluate spatial and temporal alignment while neglecting semantic equivalence between attended image regions. We present a semantic scanpath similarity framework that integrates vision-language models (VLMs) into eye-tracking analysis. Each fixation is encoded under controlled visual context (patch-based and marker-based strategies) and transformed into concise textual descriptions, which are aggregated into scanpath-level representations. Semantic similarity is then computed using embedding-based and lexical NLP metrics and compared against established spatial measures, including MultiMatch and DTW. Experiments on free-viewing eye-tracking data demonstrate that semantic similarity captures partially independent variance from geometric alignment, revealing cases of high content agreement despite spatial divergence. We further analyze the impact of contextual encoding on description fidelity and metric stability. Our findings suggest that multimodal foundation models enable interpretable, content-aware extensions of classical scanpath analysis, providing a complementary dimension for gaze research within the ETRA community.
Authors: Runpeng Geng, Chenlong Yin, Yanting Wang, Ying Chen, Jinyuan Jia
Abstract: Prompt injection attacks pose serious security risks across a wide range of real-world applications. While receiving increasing attention, the community faces a critical gap: the lack of a unified platform for prompt injection evaluation. This makes it challenging to reliably compare defenses, understand their true robustness under diverse attacks, or assess how well they generalize across tasks and benchmarks. For instance, many defenses initially reported as effective were later found to exhibit limited robustness on diverse datasets and attacks. To bridge this gap, we introduce PIArena, a unified and extensible platform for prompt injection evaluation that enables users to easily integrate state-of-the-art attacks and defenses and evaluate them across a variety of existing and new benchmarks. We also design a dynamic strategy-based attack that adaptively optimizes injected prompts based on defense feedback. Through comprehensive evaluation using PIArena, we uncover critical limitations of state-of-the-art defenses: limited generalizability across tasks, vulnerability to adaptive attacks, and fundamental challenges when an injected task aligns with the target task. The code and datasets are available at https://github.com/sleeepeer/PIArena.
Authors: Sergey V Samsonau
Abstract: Science currently offers two options for quality assurance, both inadequate. Journal gatekeeping claims to verify both integrity and contribution, but actually measures prestige: peer review is slow, biased, and misses fabricated citations even at top venues. Open science provides no quality assurance at all: the only filter between AI-generated text and the public record is the author's integrity. AI-assisted writing makes both worse by producing more papers faster than either system can absorb. We propose a third option: measure the paper itself. sciwrite-lint (pip install sciwrite-lint) is an open-source linter for scientific manuscripts that runs entirely on the researcher's machine (free public databases, a single consumer GPU, and open-weights models) with no manuscripts sent to external services. The pipeline verifies that references exist, checks retraction status, compares metadata against canonical records, downloads and parses cited papers, verifies that they support the claims made about them, and follows one level further to check cited papers' own bibliographies. Each reference receives a per-reference reliability score aggregating all verification signals. We evaluate the pipeline on 30 unseen papers from arXiv and bioRxiv with error injection and LLM-adjudicated false positive analysis. As an experimental extension, we propose SciLint Score, combining integrity verification with a contribution component that operationalizes five frameworks from philosophy of science (Popper, Lakatos, Kitcher, Laudan, Mayo) into computable structural properties of scientific arguments. The integrity component is the core of the tool and is evaluated in this paper; the contribution component is released as experimental code for community development.
Authors: Anay Mehrotra, Grigoris Velegkas, Xifan Yu, Felix Zhou
Abstract: We initiate the study of language generation in the limit, a model recently introduced by Kleinberg and Mullainathan [KM24], under the constraint of differential privacy. We consider the continual release model, where a generator must eventually output a stream of valid strings while protecting the privacy of the entire input sequence. Our first main result is that for countable collections of languages, privacy comes at no qualitative cost: we provide an $\varepsilon$-differentially-private algorithm that generates in the limit from any countable collection. This stands in contrast to many learning settings where privacy renders learnability impossible. However, privacy does impose a quantitative cost: there are finite collections of size $k$ for which uniform private generation requires $\Omega(k/\varepsilon)$ samples, whereas just one sample suffices non-privately. We then turn to the harder problem of language identification in the limit. Here, we show that privacy creates fundamental barriers. We prove that no $\varepsilon$-DP algorithm can identify a collection containing two languages with an infinite intersection and a finite set difference, a condition far stronger than the classical non-private characterization of identification. Next, we turn to the stochastic setting where the sample strings are sampled i.i.d. from a distribution (instead of being generated by an adversary). Here, we show that private identification is possible if and only if the collection is identifiable in the adversarial model. Together, our results establish new dimensions along which generation and identification differ and, for identification, a separation between adversarial and stochastic settings induced by privacy constraints.
Authors: Stephen Cheng, Sarah Wiegreffe, Dinesh Manocha
Abstract: Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works-- specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit-- freezing all attention scores during steering drops performance by only 8.75% across two model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 90-99% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions.
Authors: Addison J. Wu, Ryan Liu, Shuyue Stella Li, Yulia Tsvetkov, Thomas L. Griffiths
Abstract: Today's large language models (LLMs) are trained to align with user preferences through methods such as reinforcement learning. Yet models are beginning to be deployed not merely to satisfy users, but also to generate revenue for the companies that created them through advertisements. This creates the potential for LLMs to face conflicts of interest, where the most beneficial response to a user may not be aligned with the company's incentives. For instance, a sponsored product may be more expensive but otherwise equal to another; in this case, what does (and should) the LLM recommend to the user? In this paper, we provide a framework for categorizing the ways in which conflicting incentives might lead LLMs to change the way they interact with users, inspired by literature from linguistics and advertising regulation. We then present a suite of evaluations to examine how current models handle these tradeoffs. We find that a majority of LLMs forsake user welfare for company incentives in a multitude of conflict of interest situations, including recommending a sponsored product almost twice as expensive (Grok 4.1 Fast, 83%), surfacing sponsored options to disrupt the purchasing process (GPT 5.1, 94%), and concealing prices in unfavorable comparisons (Qwen 3 Next, 24%). Behaviors also vary strongly with levels of reasoning and users' inferred socio-economic status. Our results highlight some of the hidden risks to users that can emerge when companies begin to subtly incentivize advertisements in chatbots.
Authors: Wenbo Hu, Xin Chen, Yan Gao-Tian, Yihe Deng, Nanyun Peng, Kai-Wei Chang
Abstract: Group Relative Policy Optimization (GRPO) has emerged as the de facto Reinforcement Learning (RL) objective driving recent advancements in Multimodal Large Language Models. However, extending this success to open-source multimodal generalist models remains heavily constrained by two primary challenges: the extreme variance in reward topologies across diverse visual tasks, and the inherent difficulty of balancing fine-grained perception with multi-step reasoning capabilities. To address these issues, we introduce Gaussian GRPO (G$^2$RPO), a novel RL training objective that replaces standard linear scaling with non-linear distributional matching. By mathematically forcing the advantage distribution of any given task to strictly converge to a standard normal distribution, $\mathcal{N}(0,1)$, G$^2$RPO theoretically ensures inter-task gradient equity, mitigates vulnerabilities to heavy-tail outliers, and offers symmetric update for positive and negative rewards. Leveraging the enhanced training stability provided by G$^2$RPO, we introduce two task-level shaping mechanisms to seamlessly balance perception and reasoning. First, response length shaping dynamically elicits extended reasoning chains for complex queries while enforce direct outputs to bolster visual grounding. Second, entropy shaping tightly bounds the model's exploration zone, effectively preventing both entropy collapse and entropy explosion. Integrating these methodologies, we present OpenVLThinkerV2, a highly robust, general-purpose multimodal model. Extensive evaluations across 18 diverse benchmarks demonstrate its superior performance over strong open-source and leading proprietary frontier models.
Authors: Ziwei Zhou, Zeyuan Lai, Rui Wang, Yifan Yang, Zhen Xing, Yuqing Yang, Qi Dai, Lili Qiu, Chong Luo
Abstract: Text-to-Audio-Video (T2AV) generation is rapidly becoming a core interface for media creation, yet its evaluation remains fragmented. Existing benchmarks largely assess audio and video in isolation or rely on coarse embedding similarity, failing to capture the fine-grained joint correctness required by realistic prompts. We introduce AVGen-Bench, a task-driven benchmark for T2AV generation featuring high-quality prompts across 11 real-world categories. To support comprehensive assessment, we propose a multi-granular evaluation framework that combines lightweight specialist models with Multimodal Large Language Models (MLLMs), enabling evaluation from perceptual quality to fine-grained semantic controllability. Our evaluation reveals a pronounced gap between strong audio-visual aesthetics and weak semantic reliability, including persistent failures in text rendering, speech coherence, physical reasoning, and a universal breakdown in musical pitch control. Code and benchmark resources are available at http://aka.ms/avgenbench.
Authors: Haolei Xu, Haiwen Hong, Hongxing Li, Rui Zhou, Yang Zhang, Longtao Huang, Hui Xue, Yongliang Shen, Weiming Lu, Yueting Zhuang
Abstract: Multimodal Mixture-of-Experts (MoE) models have achieved remarkable performance on vision-language tasks. However, we identify a puzzling phenomenon termed Seeing but Not Thinking: models accurately perceive image content yet fail in subsequent reasoning, while correctly solving identical problems presented as pure text. Through systematic analysis, we first verify that cross-modal semantic sharing exists in MoE architectures, ruling out semantic alignment failure as the sole explanation. We then reveal that visual experts and domain experts exhibit layer-wise separation, with image inputs inducing significant routing divergence from text inputs in middle layers where domain experts concentrate. Based on these findings, we propose the Routing Distraction hypothesis: when processing visual inputs, the routing mechanism fails to adequately activate task-relevant reasoning experts. To validate this hypothesis, we design a routing-guided intervention method that enhances domain expert activation. Experiments on three multimodal MoE models across six benchmarks demonstrate consistent improvements, with gains of up to 3.17% on complex visual reasoning tasks. Our analysis further reveals that domain expert identification locates cognitive functions rather than sample-specific solutions, enabling effective transfer across tasks with different information structures.
Authors: Peilin Zhou, Ziyue Xu, Xinyu Shi, Jiageng Wu, Yikang Jiang, Dading Chong, Wang Dong, Jun Chen, Bin Ke, Jie Yang
Abstract: Accurate and transparent financial information disclosure is essential for market efficiency, investor decision-making, and corporate governance. Chinese stock exchanges' investor interactive platforms provide a widely used channel through which listed firms respond to investor concerns, yet these responses are often limited or non-substantive, making disclosure quality difficult to assess at scale. To address this challenge, we introduce FinTruthQA, to our knowledge the first benchmark for AI-driven assessment of financial disclosure quality in investor-firm interactions. FinTruthQA comprises 6,000 real-world financial Q&A entries, each manually annotated based on four key evaluation criteria: question identification, question relevance, answer readability, and answer relevance. We benchmark statistical machine learning models, pre-trained language models and their fine-tuned variants, as well as large language models (LLMs), on FinTruthQA. Experiments show that existing models achieve strong performance on question identification and question relevance (F1 > 95%), but remain substantially weaker on answer readability (Micro F1 approximately 88%) and especially answer relevance (Micro F1 approximately 80%), highlighting the nontrivial difficulty of fine-grained disclosure quality assessment. Domain- and task-adapted pre-trained language models consistently outperform general-purpose models and LLM-based prompting on the most challenging settings. These findings position FinTruthQA as a practical foundation for AI-driven disclosure monitoring in capital markets, with value for regulatory oversight, investor protection, and disclosure governance in real-world financial settings.
Authors: Joshua Ashkinaze, Ruijia Guan, Laura Kurek, Eytan Adar, Ceren Budak, Eric Gilbert
Abstract: Large language models (LLMs) are trained on broad corpora and then used in communities with specialized norms. Is providing LLMs with community rules enough for models to follow these norms? We evaluate LLMs' capacity to detect (Task 1) and correct (Task 2) biased Wikipedia edits according to Wikipedia's Neutral Point of View (NPOV) policy. LLMs struggled with bias detection, achieving only 64% accuracy on a balanced dataset. Models exhibited contrasting biases (some under- and others over-predicted bias), suggesting distinct priors about neutrality. LLMs performed better at generation, removing 79% of words removed by Wikipedia editors. However, LLMs made additional changes beyond Wikipedia editors' simpler neutralizations, resulting in high-recall but low-precision editing. Interestingly, crowdworkers rated AI rewrites as more neutral (70%) and fluent (61%) than Wikipedia-editor rewrites. Qualitative analysis found LLMs sometimes applied NPOV more comprehensively than Wikipedia editors but often made extraneous non-NPOV-related changes (such as grammar). LLMs may apply rules in ways that resonate with the public but diverge from community experts. While potentially effective for generation, LLMs may reduce editor agency and increase moderation workload (e.g., verifying additions). Even when rules are easy to articulate, having LLMs apply them like community members may still be difficult.
Authors: Kanishka Misra, Najoung Kim
Abstract: Neural language models (LMs) have been shown to capture complex linguistic patterns, yet their utility in understanding human language and more broadly, human cognition, remains debated. While existing work in this area often evaluates human-machine alignment, few studies attempt to translate findings from this enterprise into novel insights about humans. To this end, we propose a systematic framework for hypothesis generation that uses LMs to simulate outcomes of experiments that do not yet exist in the literature. We instantiate this framework in the context of a specific research question in child language development: dative verb acquisition and cross-structural generalization. Through this instantiation, we derive novel, untested hypotheses: the alignment between argument ordering and discourse prominence features of exposure contexts modulates how children generalize new verbs to unobserved structures. Additionally, we also design a set of experiments that can test these hypotheses in the lab with children. This work contributes both a domain-general framework for systematic hypothesis generation via simulated learners and domain-specific, lab-testable hypotheses for child language acquisition research.
Authors: Jonas Becker, Lars Benedikt Kaesberg, Andreas Stephan, Jan Philip Wahle, Terry Ruas, Bela Gipp
Abstract: Multi-agent debate - multiple instances of large language models discussing problems in turn-based interaction - has shown promise for solving knowledge and reasoning tasks. However, these methods show limitations when solving complex problems that require longer reasoning chains. We analyze how multi-agent debate drifts away from the initial problem over multiple turns, thus harming task performance. We define this phenomenon as problem drift and quantify its presence across ten tasks (i.e., three generative, three knowledge, three reasoning, and one instruction-following task). We find that generative tasks drift often due to the subjectivity of the answer space (76-89%), compared to high-complexity tasks (7-21%). To identify the reasons, eight human experts analyze 170 multi-agent debates suffering from problem drift. We find the most common issues related to this drift are the lack of progress (35% of cases), low-quality feedback (26% of cases), and a lack of clarity (25% of cases). We propose DRIFTJudge, an LLM-as-a-judge method, as a first baseline to detect problem drift. We also propose DRIFTPolicy, which mitigates 31% of problem drift cases. Our study is a step toward understanding a key limitation of multi-agent debate, highlighting why longer debates can harm task performance and how problem drift could be addressed.
Authors: Mohammad Albinhassan, Pranava Madhyastha, Alessandra Russo
Abstract: Ensuring both syntactic and semantic correctness in Large Language Model (LLM) outputs remains a significant challenge, despite being critical for real-world deployment. In this paper, we introduce $\texttt{SEM-CTRL}$, a unified approach that allows for enforcing rich context-sensitive constraints, and task and instance specific semantics directly on the LLM decoder. Our approach integrates token-level MCTS which is guided by specific syntactic and semantic constraints. The constraints over desired outputs are expressed using Answer Set Grammars, which is a logic-based formalism that generalizes context sensitive grammars while incorporating background knowledge to represent task-specific semantics. We show that our approach helps guarantee valid completions for any off-the-shelf LLM without the need for fine-tuning. We evaluate $\texttt{SEM-CTRL}$ on a range of tasks, including synthetic grammar synthesis, combinatorial reasoning, JSON parsing, and planning. Our experimental results demonstrate that $\texttt{SEM-CTRL}$ allows even small pre-trained LLMs to efficiently outperform larger variants and state-of-the-art reasoning models (e.g., $\textit{o4-mini}$) while simultaneously guaranteeing semantic validity.
Authors: Artem Timoshenko, Chengfeng Mao, John R. Hauser
Abstract: Identifying customer needs (CNs) is fundamental to product innovation and marketing strategy. Yet for over thirty years, Voice-of-the-Customer (VOC) applications have relied on professional analysts to manually interpret qualitative data and formulate "jobs to be done." This task is cognitively demanding, time-consuming, and difficult to scale. While current practice uses machine learning to screen content, the critical final step of precisely formulating CNs relies on expert human judgment. We conduct a series of studies with market research professionals to evaluate whether Large Language Models (LLMs) can automate CN abstraction. Across various product and service categories, we demonstrate that supervised fine-tuned (SFT) LLMs perform at least as well as professional analysts and substantially better than foundational LLMs. These results generalize to alternative foundational LLMs and require relatively "small" models. The abstracted CNs are well-formulated, sufficiently specific to guide innovation, and grounded in source content without hallucination. Our analysis suggests that SFT training enables LLMs to learn the underlying syntactic and semantic conventions of professional CN formulation rather than relying on memorized CNs. Automation of tedious tasks transforms the VOC approach by enabling the discovery of high-leverage insights at scale and by refocusing analysts on higher-value-added tasks.
Authors: Teng Wang, Zhangyi Jiang, Zhenqi He, Shenyang Tong, Wenhan Yang, Yanan Zheng, Zeyu Li, Zifan He, Hailei Gong, Zewen Ye, Shengjie Ma, Jianping Zhang
Abstract: Recent studies show that Large Language Models (LLMs) achieve strong reasoning capabilities through supervised fine-tuning or reinforcement learning. However, a key approach, the Process Reward Model (PRM), suffers from reward hacking, making it unreliable in identifying the best intermediate step. In addition, the cost of annotating reasoning processes for reward modeling is high, making large-scale collection of high-quality data challenging. To address this, we propose a novel reward model approach called the Hierarchical Reward Model (HRM), which evaluates both individual and consecutive reasoning steps at both fine-grained and coarse-grained levels. HRM excels at assessing multi-step reasoning coherence, especially when flawed steps are later corrected through self-reflection. To further reduce the cost of generating training data, we introduce a lightweight and effective data augmentation strategy called Hierarchical Node Compression (HNC), which merges two consecutive reasoning steps into one within the tree structure. By applying HNC to MCTS-generated reasoning trajectories, we enhance the diversity and robustness of HRM training data while introducing controlled noise with minimal computational overhead. Empirical results on the PRM800K dataset show that HRM, together with HNC, provides more stable and reliable evaluations than PRM. Furthermore, cross-domain evaluations on the MATH500 and GSM8K datasets demonstrate HRM's strong generalization and robustness across a variety of reasoning tasks.
Authors: Zhengyi Zhao, Shubo Zhang, Yiming Du, Bin Liang, Baojun Wang, Zhongyang Li, Binyang Li, Kam-Fai Wong
Abstract: Large language models have improved dialogue systems, but often process conversational turns in isolation, overlooking the event structures that guide natural interactions. Hence we introduce EventWeave, a framework that explicitly models relationships between conversational events to generate more contextually appropriate dialogue responses. EventWeave constructs a dynamic event graph that distinguishes between core events (main goals) and supporting events (interconnected details), employing a multi-head attention mechanism to selectively determine which events are most relevant to the current turn. Unlike summarization or standard graph-based approaches, our method captures three distinct relationship types between events, allowing for more nuanced context modeling. Experiments on three dialogue datasets demonstrate that EventWeave produces more natural and contextually appropriate responses while requiring less computational overhead than models processing the entire dialogue history. Ablation studies confirm improvements stem from better event relationship modeling rather than increased information density. Our approach effectively balances comprehensive context understanding with generating concise responses, maintaining strong performance across various dialogue lengths through targeted optimization techniques.
Authors: Eylon Caplan, Tania Chakraborty, Dan Goldwasser
Abstract: Variation in language use, shaped by speakers' sociocultural background and specific context of use, offers a rich lens into cultural perspectives, values, and opinions. For example, Chinese students discuss "healthy eating" with words like "timing," "regularity," and "digestion," whereas Americans use vocabulary like "balancing food groups" and "avoiding fat and sugar," reflecting distinct cultural models of nutrition. The computational study of these Sociocultural Linguistic Phenomena (SLP) has traditionally been done in NLP via tailored analyses of specific groups or topics, requiring specialized data collection and experimental operationalization--a process not well-suited to quick hypothesis exploration and prototyping. To address this, we propose constructing a "sandbox" designed for systematic and flexible sociolinguistic research. Using our method, we construct a demographically/topically split Reddit dataset, Splits!, validated by self-identification and by replicating several known SLPs from existing literature. We showcase the sandbox's utility with a scalable, two-stage process that filters large collections of "potential" SLPs (PSLPs) to surface the most promising candidates for deeper, qualitative investigation.
Authors: Weiqi Wang, Jiefu Ou, Yangqiu Song, Benjamin Van Durme, Daniel Khashabi
Abstract: Literature review tables are essential for summarizing and comparing collections of scientific papers. In this paper, we study the automatic generation of such tables from a pool of papers to satisfy a user's information need. Building on recent work (Newman et al., 2024), we move beyond oracle settings by (i) simulating well-specified yet schema-agnostic user demands that avoid leaking gold column names or values, (ii) explicitly modeling retrieval noise via semantically related but out-of-scope distractor papers verified by human annotators, and (iii) introducing a lightweight, annotation-free, utilization-oriented evaluation that decomposes utility into schema coverage, unary cell fidelity, and pairwise relational consistency, while measuring paper selection through a two-way QA procedure (gold to system and system to gold) with recall, precision, and F1. To support reproducible evaluation, we introduce arXiv2Table, a benchmark of 1,957 tables referencing 7,158 papers, with human-verified distractors and rewritten, schema-agnostic user demands. We also develop an iterative, batch-based generation method that co-refines paper filtering and schema over multiple rounds. We validate the evaluation protocol with human audits and cross-evaluator checks. Extensive experiments show that our method consistently improves over strong baselines, while absolute scores remain modest, underscoring the task's difficulty. Our data and code is available at https://github.com/JHU-CLSP/arXiv2Table.
Authors: Dezheng Han, Yibin Jia, Ruxiao Chen, Wenjie Han, Shuaishuai Guo, Jianbo Wang
Abstract: With the rapid development of large language models (LLMs), their application to cell type annotation has drawn increasing attention. However, general-purpose LLMs often face limitations in this specific task due to the lack of guidance from external domain knowledge. To enable more accurate and fully automated cell type annotation, we develop a globally connected knowledge graph comprising 18850 biological information nodes, including cell types, gene markers, features, and other related entities, along with 48,944 edges connecting these nodes, which is used by LLMs to retrieve entities associated with differential genes for cell reconstruction. Additionally, a multi-task reasoning workflow is designed to optimise the annotation process. Compared to general-purpose LLMs, our method improves human evaluation scores by up to 0.21 and semantic similarity by 6.1% across multiple tissue types, while more closely aligning with the cognitive logic of manual annotation. Meanwhile, it narrows the performance gap between large and small LLMs in cell type annotation, offering a paradigm for structured knowledge integration and reasoning in bioinformatics.
Authors: Ryo Kamoi, Yusen Zhang, Nan Zhang, Sarkar Snigdha Sarathi Das, Ranran Haoran Zhang, Wenpeng Yin, Rui Zhang
Abstract: Process Reward Models (PRMs) have emerged as a promising approach for improving LLM reasoning capabilities by providing process supervision over reasoning traces. However, existing approaches for constructing PRM training data remain costly and noisy, as they typically rely on human annotation or sampling-based labeling methods that require repeated LLM calls. In this work, we propose FoVer, a framework that synthesizes PRM training data from formal reasoning tasks by annotating step-level error labels using formal verification tools such as Z3 and Isabelle. By leveraging formal verification, FoVer enables efficient and accurate PRM data construction without requiring human annotation or additional LLM calls. Using FoVer, we create PRM training data from formal logic and theorem proving tasks. Experiments on 12 reasoning benchmarks show that fine-tuning on our training data improves PRMs not only on math and logic reasoning tasks, which are informal variants of the training tasks, but also on NLI and BBH benchmarks, which differ substantially from the tasks used to construct the training data. These results demonstrate the practical effectiveness of FoVer, showing that PRM training data created using formal verification improves PRMs on informal reasoning tasks written in natural language. The datasets, models, and code are provided at https://github.com/psunlpgroup/FoVer.
Authors: Thinh Pham, Nguyen Nguyen, Pratibha Zunjare, Weiyuan Chen, Yu-Min Tseng, Tu Vu
Abstract: We introduce SealQA, a new challenge benchmark for evaluating SEarch-Augmented Language models on fact-seeking questions where web search yields conflicting, noisy, or unhelpful results. SealQA comes in three flavors: (1) Seal-0 (main) and (2) Seal-Hard, which assess factual accuracy and reasoning capabilities, with Seal-0 focusing on the most challenging questions where chat models (e.g., GPT-4.1) typically achieve near-zero accuracy; and (3) LongSeal, which extends SealQA to test long-context, multi-document reasoning in "needle-in-a-haystack" settings. Our evaluation reveals critical limitations in current models: Even frontier LLMs perform poorly across all SealQA flavors. On Seal-0, frontier agentic models equipped with tools like o3 and o4-mini achieve only 17.1% and 6.3% accuracy, respectively, at their best reasoning efforts. We find that advanced reasoning models such as DeepSeek-R1-671B and o3-mini are highly vulnerable to noisy search results. Notably, increasing test-time compute does not yield reliable gains across o3-mini, o4-mini, and o3, with performance often plateauing or even declining early. Additionally, while recent models are less affected by the "lost-in-the-middle" issue, they still fail to reliably identify relevant documents in LongSeal when faced with numerous distractors. To facilitate future work, we release SealQA at huggingface.co/datasets/vtllms/sealqa.
Authors: Pulkit Bansal, Raghvendra Kumar, Shakti Singh, Sriparna Saha, Adam Jatowt
Abstract: In an era of rampant misinformation, generating reliable news explanations is vital, especially for under-represented languages like Hindi. Lacking robust automated tools, Hindi faces challenges in scaling misinformation detection. To bridge this gap, we propose a novel framework integrating Direct Preference Optimization (DPO) with curriculum learning to align machine-generated explanations with human reasoning. Fact-checked explanations from credible sources serve as preferred responses, while LLM outputs highlight system limitations and serve as non-preferred responses. To refine task-specific alignment, we introduce two key parameters -- Actuality and Finesse -- into the DPO loss function, enhancing explanation quality and consistency. Experiments with LLMs (Mistral, Llama, Gemma) and PLMs (mBART, mT5) confirm the framework's effectiveness in generating coherent, contextually relevant explanations. This scalable approach combats misinformation and extends automated explanation generation to low-resource languages.
Authors: Yunxiao Wang, Meng Liu, Kaiyu Jiang, Bin Wen, Fan Yang, Tingting Gao, Lizi Liao
Abstract: Emotional support conversations require more than fluent responses. Supporters need to understand the seeker's situation and emotions, adopt an appropriate strategy, and respond in a natural, human-like manner. Despite advances in large language models, current systems often lack structured, psychology-informed reasoning. Additionally, it is challenging to enhance these systems through reinforcement learning because of unreliable reward signals. Moreover, reinforcement fine-tuning can amplify repetitive response patterns. We propose structured empathetic reasoning, which breaks support into three steps: conversation history analysis, multimodal emotional state inference, and strategy selection, prior to generating the final reply. To implement this, we introduce SER, a fine-grained dataset with step-level correctness labels and pairwise response preferences. We then present PEER, which uses GRPO with UnifiReward, a unified process-outcome reward model for evaluating both reasoning steps and final responses in multi-turn interactions. To reduce repetition, we enhance data with personality-based rewriting and down-weight redundant outputs. Comprehensive experiments show improved empathy, strategy alignment, and human-likeness without sacrificing diversity.
Authors: Shaohua Duan, Pengcheng Huang, Xinze Li, Zhenghao Liu, Xiaoyuan Yi, Yukun Yan, Shuo Wang, Yu Gu, Ge Yu, Maosong Sun
Abstract: Long-context modeling is critical for a wide range of real-world tasks, including long-context question answering, summarization, and complex reasoning tasks. Recent studies have explored fine-tuning Large Language Models (LLMs) with synthetic data to enhance their long-context capabilities. However, the effectiveness of such approaches is often limited by the low diversity and factual inconsistencies in the generated data. To address these challenges, we propose LongMab, a novel framework that leverages a Multi-Armed Bandit (MAB) rollout strategy to identify the most informative chunks from the given long context for sampling high-quality and diverse responses and constructing preference data pairs for Direct Preference Optimization (DPO) training. Specifically, we treat context chunks as arms of MAB, select chunks based on their expected reward scores to input into LLMs to generate responses, and iteratively update these scores based on reward feedback. Both exploration and exploitation during the rollout process enable the LLM to focus on the most relevant context segments, thereby generating and collecting high-quality and diverse responses. Experimental results on both Llama and Qwen show the effectiveness of LongMab by achieving more than a 4% improvement on long-context reasoning benchmarks. All data and code will be released on https://github.com/NEUIR/LongMab-PO.
Authors: Pengxiang Li, Yefan Zhou, Dilxat Muhtar, Lu Yin, Shilin Yan, Li Shen, Soroush Vosoughi, Shiwei Liu
Abstract: Diffusion language models (DLMs) have recently emerged as an alternative to autoregressive approaches, offering parallel sequence generation and flexible token orders. However, their inference remains slower than that of autoregressive models, primarily due to the cost of bidirectional attention and the large number of refinement steps required for high quality outputs. In this work, we highlight and leverage an overlooked property of DLMs early answer convergence: in many cases, the correct answer can be internally identified by half steps before the final decoding step, both under semi-autoregressive and random remasking schedules. For example, on GSM8K and MMLU, up to 97% and 99% of instances, respectively, can be decoded correctly using only half of the refinement steps. Building on this observation, we introduce Prophet, a training-free fast decoding paradigm that enables early commit decoding. Specifically, Prophet dynamically decides whether to continue refinement or to go "all-in" (i.e., decode all remaining tokens in one step), using the confidence gap between the top-2 prediction candidates as the criterion. It integrates seamlessly into existing DLM implementations, incurs negligible overhead, and requires no additional training. Empirical evaluations of LLaDA-8B and Dream-7B across multiple tasks show that Prophet reduces the number of decoding steps by up to 3.4x while preserving high generation quality. These results recast DLM decoding as a problem of when to stop sampling, and demonstrate that early decode convergence provides a simple yet powerful mechanism for accelerating DLM inference, complementary to existing speedup techniques. Our code is publicly available at https://github.com/pixeli99/Prophet.
Authors: Jeonghyun Park, Ingeol Baek, Seunghyun Yoon, Haeun Jang, Aparna Garimella, Akriti Jain, Nedim Lipka, Hwanhee Lee
Abstract: Real-world multi-hop QA is naturally linked with ambiguity, where a single query can trigger multiple reasoning paths that require independent resolution. Since ambiguity can occur at any stage, models must navigate layered uncertainty throughout the entire reasoning chain. Despite its prevalence in real-world user queries, previous benchmarks have primarily focused on single-hop ambiguity, leaving the complex interaction between multi-step inference and layered ambiguity underexplored. In this paper, we introduce \textbf{MARCH}, a benchmark for their intersection, with 2,209 multi-hop ambiguous questions curated via multi-LLM verification and validated by human annotation with strong agreement. Our experiments reveal that even state-of-the-art models struggle with MARCH, confirming that combining ambiguity resolution with multi-step reasoning is a significant challenge. To address this, we propose \textbf{CLARION}, a two-stage agentic framework that explicitly decouples ambiguity planning from evidence-driven reasoning, significantly outperforms existing approaches, and paves the way for robust reasoning systems.
Authors: Ayan Majumdar, Feihao Chen, Jinghui Li, Xiaozhen Wang
Abstract: Large-scale web-scraped text corpora used to train general-purpose AI models often contain harmful demographic-targeted social biases, creating a regulatory need for data auditing and developing scalable bias-detection methods. Although prior work has investigated biases in text datasets and related detection methods, these studies remain narrow in scope. They typically focus on a single content type (e.g., hate speech), cover limited demographic axes, overlook biases affecting multiple demographics simultaneously, and analyze limited techniques. Consequently, practitioners lack a holistic understanding of the strengths and limitations of recent large language models (LLMs) for automated bias detection. In this study, we conduct a comprehensive benchmark study on English texts to assess the ability of LLMs in detecting demographic-targeted social biases. To align with regulatory requirements, we frame bias detection as a multi-label task of detecting targeted identities using a demographic-focused taxonomy. We then systematically evaluate models across scales and techniques, including prompting, in-context learning, and fine-tuning. Using twelve datasets spanning diverse content types and demographics, our study demonstrates the promise of fine-tuned smaller models for scalable detection. However, our analyses also expose persistent gaps across demographic axes and multi-demographic targeted biases, underscoring the need for more effective and scalable detection frameworks.
Authors: Kuangshi Ai, Jonathan A. Karr Jr, Meng Jiang, Nitesh V. Chawla, Chaoli Wang
Abstract: We present Knowledge Extraction on OMIn (KEO), a domain-specific knowledge extraction and reasoning framework with large language models (LLMs) in safety-critical contexts. Using the Operations and Maintenance Intelligence (OMIn) dataset, we construct a QA benchmark spanning global sensemaking and actionable maintenance tasks. KEO builds a structured Knowledge Graph (KG) and integrates it into a retrieval-augmented generation (RAG) pipeline, enabling more coherent, dataset-wide reasoning than traditional text-chunk RAG. We evaluate locally deployable LLMs (Gemma-3, Phi-4, Mistral-Nemo) and employ stronger models (GPT-4o, Llama-3.3) as judges. Experiments show that KEO markedly improves global sensemaking by revealing patterns and system-level insights, while text-chunk RAG remains effective for fine-grained procedural tasks requiring localized retrieval. These findings underscore the promise of KG-augmented LLMs for secure, domain-specific QA and their potential in high-stakes reasoning. The code is available at https://github.com/JonathanKarr33/keo.
Authors: Hsien-Chin Lin, Benjamin Matthias Ruppik, Carel van Niekerk, Chia-Hao Shen, Michael Heck, Nurul Lubis, Renato Vukovic, Shutong Feng, Milica Ga\v{s}i\'c
Abstract: Large language models (LLMs) have achieved remarkable success in a wide range of natural language processing tasks and can be adapted through prompting. However, they remain suboptimal in multi-turn interactions, often relying on incorrect early assumptions and failing to track user goals over time, which makes such tasks particularly challenging. Prior works in dialogue systems have shown that long-term planning is essential for handling interactive tasks. In this work, we propose a prompt optimisation framework inspired by reinforcement learning, which enables such planning to take place by only modifying the task instruction prompt of the LLM-based agent. By generating turn-by-turn feedback and leveraging experience replay for prompt rewriting, our proposed method shows significant improvement in multi-turn tasks such as text-to-SQL and task-oriented dialogue. Moreover, it generalises across different LLM-based agents and can leverage diverse LLMs as meta-prompting agents. This warrants future research in reinforcement learning-inspired parameter-free optimisation methods.
Authors: Shangjian Yin, Shining Liang, Wenbiao Ding, Yuli Qian, Zhouxing Shi, Hongzhi Li, Yutao Xie
Abstract: High-quality instruction data is critical for LLM alignment, yet existing open-source datasets often lack efficiency, requiring hundreds of thousands of examples to approach proprietary performance. In this work, we find that beyond the widely recognized importance of prompt-response quality, prompt difficulty itself plays a critical role in driving alignment gains. Motivated by this observation, we introduce PiKa, a data-efficient family of expert-level alignment datasets that concentrates supervision on high-difficulty instructions. The PiKa-SFT dataset contains only 30k examples, an order of magnitude fewer than state-of-the-art open datasets like Magpie-Pro. Despite its small size, fine-tuning Llama-3-8B-Base on PiKa-SFT even outperforms the official Llama-3-8B-Instruct model trained on over 10M proprietary examples on widely used benchmarks such as AlpacaEval 2.0 and Arena-Hard. We also validate the generalizability of PiKa across the Qwen2.5 series (0.5B-7B), consistently surpassing their official instruction-tuned counterparts. Additionally, we provide 30k high-quality preference optimization examples to further enhance alignment. Our results demonstrate that promising alignment is achievable with significantly reduced data, democratizing access for resource-constrained research. Our code and data will be available at https://github.com/SJY8460/PiKa.
Authors: Yuntao Gui, James Cheng
Abstract: Despite their remarkable natural language understanding capabilities, Large Language Models (LLMs) have been underutilized for retrieval tasks. We present Search-R3, a novel framework that addresses this limitation by adapting LLMs to generate search embeddings as a direct output of their reasoning process. Our approach exploits LLMs' chain-of-thought capabilities, allowing them to produce more effective embeddings by reasoning step-by-step through complex semantic analyses. We implement this through three complementary mechanisms. (1) a supervised learning stage enables the model's ability to produce quality embeddings, (2) a reinforcement learning (RL) methodology that optimizes embedding generation alongside reasoning, and (3) a specialized RL environment that efficiently handles evolving embedding representations without requiring complete corpus re-encoding at each training iteration. Our extensive evaluations on diverse benchmarks demonstrate that Search-R3 significantly outperforms prior methods by unifying the reasoning and embedding generation processes. This integrated post-training approach represents a substantial advancement in handling complex knowledge-intensive tasks that require both sophisticated reasoning and effective information retrieval. Project page: https://github.com/ytgui/Search-R3
Authors: Lang Gao, Xuhui Li, Chenxi Wang, Mingzhe Li, Wei Liu, Zirui Song, Jinghui Zhang, Rui Yan, Preslav Nakov, Xiuying Chen
Abstract: Large language models (LLMs) have grown more powerful in language generation, producing fluent text and even imitating personal style. Yet, this ability also heightens the risk of identity impersonation. To the best of our knowledge, no prior work has examined personalized machine-generated text (MGT) detection. In this paper, we introduce \dataset, the first benchmark for evaluating detector robustness in personalized settings, built from literary and blog texts paired with their LLM-generated imitations. Our experimental results demonstrate large performance gaps across detectors in personalized settings: some state-of-the-art models suffer significant drops. We attribute this limitation to the \textit{feature-inversion trap}, where features that are discriminative in general domains become inverted and misleading when applied to personalized text. Based on this finding, we propose \method, a simple and reliable way to predict detector performance changes in personalized settings. \method identifies latent directions corresponding to inverted features and constructs probe datasets that differ primarily along these features to evaluate detector dependence. Our experiments show that \method can accurately predict both the direction and the magnitude of post-transfer changes, showing 85\% correlation with the actual performance gaps. We hope that this work will encourage further research on personalized text detection.
Authors: Yuanchen Wu, Saurabh Verma, Justin Lee, Fangzhou Xiong, Poppy Zhang, Amel Awadelkarim, Xu Chen, Yubai Yuan, Shawndra Hill
Abstract: Large language models (LLMs) are highly sensitive to prompts, but most automatic prompt optimization (APO) methods assume access to ground-truth references (e.g., labeled validation data) that are costly to obtain. We propose the Prompt Duel Optimizer (PDO), a sample-efficient framework for label-free prompt optimization based on pairwise preference feedback from an LLM judge. PDO casts prompt selection as a dueling-bandit problem and combines (i) Double Thompson Sampling to prioritize informative comparisons under a fixed judge budget, with (ii) top-performer guided mutation to expand the candidate pool while pruning weak prompts. Experiments on BIG-bench Hard (BBH) and MS MARCO show that PDO consistently identifies stronger prompts than label-free baselines, while offering favorable quality--cost trade-offs under constrained comparison budgets.
Authors: Azmine Toushik Wasi, Wahid Faisal, Mst Rafia Islam, Md Rizwan Parvez
Abstract: Bangladesh's low-income population faces major barriers to affordable legal advice due to complex legal language, procedural opacity, and high costs. Existing AI legal assistants lack Bengali-language support and jurisdiction-specific adaptation, limiting their effectiveness. To address this, we developed Mina, a multilingual LLM-based legal assistant tailored for the Bangladeshi context. It employs multilingual embeddings and a RAG-based chain-of-tools framework for retrieval, reasoning, translation, and document generation, delivering context-aware legal drafts, citations, and plain-language explanations via an interactive chat interface. Evaluated by law faculty from leading Bangladeshi universities across all stages of the 2022 and 2023 Bangladesh Bar Council Exams, Mina scored 75-80% in Preliminary MCQs, Written, and simulated Viva Voce exams, matching or surpassing average human performance and demonstrating clarity, contextual understanding, and sound legal reasoning. Even under a conservative upper bound, Mina operates at just 0.12-0.61% of typical legal consultation costs in Bangladesh, yielding a 99.4-99.9\% cost reduction relative to human-provided services. These results confirm its potential as a low-cost, multilingual AI assistant that automates key legal tasks and scales access to justice, offering a real-world case study on building domain-specific, low-resource systems and addressing challenges of multilingual adaptation, efficiency, and sustainable public-service AI deployment.
Authors: Dazhen Deng, Sen Yang, Yuchen He, Yuan Tian, Yingcai Wu
Abstract: Current chart-related tasks, such as chart generation (NL2Chart), chart schema parsing, chart data parsing, and chart question answering (ChartQA), are typically studied in isolation, preventing models from learning the shared semantics that link chart creation and interpretation. We introduce CycleChart, a consistency-based learning framework for bidirectional chart understanding and generation. Unlike conventional multi-task approaches that draw training samples independently across tasks, CycleChart organizes all tasks around each single data instance. From a source table and natural-language query, the model generates a chart specification, renders and executes it, then learns to recover the schema and underlying data from the resulting chart image. This per-instance lifecycle design lets the model capture the full chain of transformations, from raw data through visual encoding to structured recovery, and a generate--parse consistency objective enforces semantic alignment between the forward generation and reverse parsing directions. To support this framework, we construct CycleChart-Bench, a lifecycle-aligned benchmark where every chart sample carries aligned annotations for generation, schema parsing, data parsing, and question answering. CycleChart achieves strong results across all four tasks and transfers effectively to unseen external benchmarks, demonstrating improved cross-task generalization and marking a step toward more general chart understanding models.
Authors: Chenggong Zhang, Haopeng Wang, Hexi Meng
Abstract: Hallucinations in Large Language Models (LLMs) pose a significant challenge, generating misleading or unverifiable content that undermines trust and reliability. Existing evaluation methods, such as KnowHalu, employ multi-stage verification but suffer from high computational costs. To address this, we integrate the Hughes Hallucination Evaluation Model (HHEM), a lightweight classification-based framework that operates independently of LLM-based judgments, significantly improving efficiency while maintaining high detection accuracy. We conduct a comparative analysis of hallucination detection methods across various LLMs, evaluating True Positive Rate (TPR), True Negative Rate (TNR), and Accuracy on question-answering (QA) and summarization tasks. Our results show that HHEM reduces evaluation time from 8 hours to 10 minutes, while HHEM with non-fabrication checking achieves the highest accuracy \(82.2\%\) and TPR \(78.9\%\). However, HHEM struggles with localized hallucinations in summarization tasks. To address this, we introduce segment-based retrieval, improving detection by verifying smaller text components. Additionally, our cumulative distribution function (CDF) analysis indicates that larger models (7B-9B parameters) generally exhibit fewer hallucinations, while intermediate-sized models show higher instability. These findings highlight the need for structured evaluation frameworks that balance computational efficiency with robust factual validation, enhancing the reliability of LLM-generated content.
Authors: Fabian Retkowski, Alexander Waibel
Abstract: Automatic speech transcripts are often delivered as unstructured word streams that impede readability and repurposing. We recast paragraph segmentation as the missing structuring step and fill three gaps at the intersection of speech processing and text segmentation. First, we establish TEDPara (human-annotated TED talks) and YTSegPara (YouTube videos with synthetic labels) as the first benchmarks for the paragraph segmentation task. The benchmarks focus on the underexplored speech domain, where paragraph segmentation has traditionally not been part of post-processing, while also contributing to the wider text segmentation field, which still lacks robust and naturalistic benchmarks. Second, we propose a constrained-decoding formulation that lets large language models insert paragraph breaks while preserving the original transcript, enabling faithful, sentence-aligned evaluation. Third, we show that a compact model (MiniSeg) attains state-of-the-art accuracy and, when extended hierarchically, jointly predicts chapters and paragraphs with minimal computational cost. Together, our resources and methods establish paragraph segmentation as a standardized, practical task in speech processing.
Authors: Hyeong Kyu Choi, Sharon Li
Abstract: Selecting a single high-quality output from multiple stochastic generations remains a fundamental challenge for large language models (LLMs), particularly in open-ended tasks where no canonical answer exists. While Best-of-N and self-consistency methods show that aggregating multiple generations can improve performance, existing approaches typically rely on external evaluators, reward models, or exact string-match voting, limiting their applicability and efficiency. We propose Mode Extraction (ModeX), an evaluator-free Best-of-N selection framework that generalizes majority voting to open-ended text generation by identifying the modal output representing the dominant semantic consensus among generated texts. ModeX constructs a similarity graph over candidate generations and recursively applies spectral clustering to select a representative centroid, without requiring additional inference or auxiliary models. We further instantiate this selection principle as ModeX-Lite, an improved version of ModeX with early pruning for efficiency. Across open-ended tasks -- including text summarization, code generation, and mathematical reasoning -- our approaches consistently outperform standard single- and multi-path baselines, providing a computationally efficient solution for robust open-ended text generation. Code is released in https://github.com/deeplearning-wisc/ModeX.
Authors: Devang Kulshreshtha, Hang Su, Haibo Jin, Chinmay Hegde, Haohan Wang
Abstract: We introduce \emph{self-jailbreaking}, a threat model in which an aligned LLM guides its own compromise. Unlike most jailbreak techniques, which often rely on handcrafted prompts or separate attacker models, self-jailbreaking requires no external red-team LLM: the target model's own internal knowledge suffices. We operationalize this via \textbf{Self-Jailbreaking via Lexical Insertion Prompting (\textsc{SLIP})}, a black-box algorithm that casts jailbreaking as breadth-first tree search over multi-turn dialogues, incrementally inserting missing content words from the attack goal into benign prompts using the target model as its own guide. Evaluations on AdvBench and HarmBench show \textsc{SLIP} achieves 90--100\% Attack Success Rate (ASR) (avg.\ 94.7\%) across most of the eleven tested models (including GPT-5.1, Claude-Sonnet-4.5, Gemini-2.5-Pro, and DeepSeek-V3), with only ${\sim}7.9$ LLM calls on average, 3--6$\times$ fewer than prior methods. We evaluate existing defenses, show that regex-based approaches are evaded by prompt paraphrasing, and propose the Semantic Drift Monitor (SDM) defense that tracks \textsc{SLIP}'s embedding-space trajectory, achieving 76\% detection at 5\% FPR. However, SDM remains insufficient against adaptive attack strategies, underscoring the need for more advanced defense mechanisms tailored to the self-jailbreaking threat surface. We release our code for reproducibility.
Authors: Loris Schoenegger, Benjamin Roth
Abstract: Training data influence estimation methods quantify the contribution of training documents to a model's output, making them a promising source of information for example-based explanations. As humans cannot interpret thousands of documents, only a small subset of the training data can be presented as an explanation. Although the choice of which documents to include directly affects explanation quality, previous evaluations of such systems have largely ignored any selection strategies. To address this, we propose a novel selection relevance score, a retraining-free metric that quantifies how useful a set of examples is for explaining a model's output. We validate this score through fine-tuning experiments, confirming that it can predict whether a set of examples supports or undermines the model's predictions. Using this metric, we further show that common selection strategies often underperform random selection. Motivated by this finding, we propose a strategy that balances influence and representativeness, enabling better use of selection budgets than naively selecting the highest-ranking examples.
Authors: Yuhao Shen, Tianyu Liu, Junyi Shen, Jinyang Wu, Quan Kong, Li Huan, Cong Wang
Abstract: Parallel Speculative Decoding (PSD) accelerates traditional Speculative Decoding (SD) by overlapping draft generation with verification. However, it remains hampered by two fundamental challenges: (1) a theoretical speedup ceiling dictated by the speed ratio between the draft and target models, and (2) high computational waste and pipeline stall due to mid-sequence token rejections of early errors. To address these limitations, we introduce \textsc{Double} (Double Retrieval Speculative Parallelism). By bridging the gap between SD and PSD, our framework resolves the Retrieval \emph{Precision-Efficiency Dilemma} through a novel synchronous mechanism. Specifically, we enable the draft model to execute iterative retrieval speculations to break the theoretical speedup limits; to alleviate rejections without rollback, the target model performs authoritative retrieval to generate multi-token guidance. \textsc{Double} is entirely training-free and lossless. Extensive experiments demonstrate state-of-the-art speedup of $\textbf{5.3}\times$ on LLaMA3.3-70B and $\textbf{2.8}\times$ on Qwen3-32B, significantly outperforming the advanced method EAGLE-3 that requires extensive model training.
Authors: Nayoung Choi, Jonathan Zhang, Jinho D. Choi
Abstract: Large Language Models (LLMs) increasingly operate over long-form dialogues with frequent topic shifts. While recent LLMs support extended context windows, efficient management of dialogue history in practice is needed due to inference cost and latency constraints. We present DyCP, a lightweight context management method implemented outside the LLM that dynamically identifies and retrieves relevant dialogue segments conditioned on the current turn, without offline memory construction. DyCP manages dialogue context while preserving the sequential nature of dialogue without predefined topic boundaries, enabling adaptive and efficient context selection. Across three long-form dialogue benchmarks-LoCoMo, MT-Bench+, and SCM4LLMs-and multiple LLM backends, DyCP achieves competitive answer quality in downstream generation, with more selective context usage and improved inference efficiency.
Authors: Xintao Wang, Jian Yang, Weiyuan Li, Rui Xie, Jen-tse Huang, Jun Gao, Shuai Huang, Yueping Kang, Liyuan Gou, Hongwei Feng, Yanghua Xiao
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in reasoning and generation, serving as the foundation for advanced persona simulation and Role-Playing Language Agents (RPLAs). However, achieving authentic alignment with human cognitive and behavioral patterns remains a critical challenge for these agents. We present HumanLLM, a framework treating psychological patterns as interacting causal forces. We construct 244 patterns from ~12,000 academic papers and synthesize 11,359 scenarios where 2--5 patterns reinforce, conflict, or modulate each other, with multi-turn conversations expressing inner thoughts, actions, and dialogue. Our dual-level checklists evaluate both individual pattern fidelity and emergent multi-pattern dynamics, achieving strong human alignment (r=0.90) while revealing that holistic metrics conflate simulation accuracy with social desirability. HumanLLM-8B outperforms Qwen3-32B on multi-pattern dynamics despite 4x fewer parameters, demonstrating that authentic anthropomorphism requires cognitive modeling -- simulating not just what humans do, but the psychological processes generating those behaviors. Our dataset, code, and model are available at: https://github.com/YJGoodbye2024/HumanLLM.git
Authors: Matteo Rinaldi, Rossella Varvara, Viviana Patti
Abstract: We present "Testimole-conversational" a massive collection of discussion boards messages in the Italian language. The large size of the corpus, more than 30B word-tokens (1996-2024), renders it an ideal dataset for native Italian Large Language Models'pre-training. Furthermore, discussion boards' messages are a relevant resource for linguistic as well as sociological analysis. The corpus captures a rich variety of computer-mediated communication, offering insights into informal written Italian, discourse dynamics, and online social interaction in wide time span. Beyond its relevance for NLP applications such as language modelling, domain adaptation, and conversational analysis, it also support investigations of language variation and social phenomena in digital communication. The resource will be made freely available to the research community.
Authors: Zhuokang Shen, Yifan Wang, Hanyu Chen, Yunhang Shen, Wenxuan Huang, Gaoqi He, Jiao Xie, Rongrong Ji, Shaohui Lin
Abstract: Recent advances in large language models (LLMs) have enabled increasingly capable chatbots. However, most existing systems focus on single-user settings and do not generalize well to multi-user group chat interactions, where agents require more proactive and accurate intervention under complex, evolving contexts. Existing approaches typically rely on LLMs for both intervention reasoning and response generation, leading to high token consumption, limited scalability, and potential privacy risks. To address these challenges, we propose GroupGPT, a token-efficient and privacy-preserving agentic framework for multi-user chat assistant. GroupGPT adopts an edge-cloud model collaboration architecture to decouple intervention timing from response generation, enabling efficient and accurate decision-making while preserving user privacy through on-device processing of sensitive information. The framework also supports multimodal inputs, including memes, images, videos, and voice messages.To support evaluation of timing accuracy and response quality, we further introduce MUIR, a benchmark dataset for multi-user chat assistant intervention reasoning. MUIR contains 2,500 annotated group chat segments with intervention labels and rationales. We evaluate a range of models on MUIR, spanning from open-source to proprietary variants, including both LLMs and their smaller counterparts. Extensive experiments demonstrate that GroupGPT generates accurate and well-timed responses, achieving an average score of 4.72/5.0 in LLM-based evaluation, and is well-received by users across diverse group chat scenarios. Moreover, GroupGPT reduces the token usage by up to 3 times compared to baselines, while providing privacy sanitization of user messages before cloud transmission. Code is available at: https://github.com/Eliot-Shen/GroupGPT .
Authors: Wei Han, Pan Zhou, Soujanya Poria, Shuicheng Yan
Abstract: The limited context window of contemporary large language models (LLMs) remains a primary bottleneck for their broader application across diverse domains. Although continual pre-training on long-context data offers a straightforward solution, it incurs prohibitive data acquisition and computational costs. To address this challenge, we propose~\modelname, a novel framework based on multi-grained context compression and query-aware information acquisition. SharedLLM comprises two stacked short-context LLMs: a lower model serving as a compressor and an upper model acting as a decoder. The lower model compresses long inputs into compact, multi-grained representations, which are then forwarded to the upper model for context-aware processing. To maximize efficiency, this information transfer occurs exclusively at the lowest layers, bypassing lengthy forward passes and redundant cross-attention operations. This entire process, wherein the upper and lower models are derived from the same underlying LLM layers, is termed~\textit{self-injection}. To support this architecture, a specialized tree-based data structure enables the efficient encoding and query-aware retrieval of contextual information. Despite being trained on sequences of only 8K tokens, \modelname~effectively generalizes to inputs exceeding 128K tokens. Across a comprehensive suite of long-context modeling and understanding benchmarks, \modelname~achieves performance superior or comparable to strong baselines, striking an optimal balance between efficiency and accuracy. Furthermore, these design choices allow \modelname~to substantially reduce the memory footprint and yield notable inference speedups ($2\times$ over streaming and $3\times$ over encoder-decoder architectures).
Authors: Panatchakorn Anantaprayoon, Nataliia Babina, Nima Asgharbeygi, Jad Tarifi
Abstract: LLM alignment has progressed in single-agent settings through paradigms such as RL with human feedback (RLHF), while recent work explores scalable alternatives such as RL with AI feedback (RLAIF) and dynamic alignment objectives. However, these approaches remain limited in multi-stakeholder settings, where conflicting values arise and deliberative negotiation is required. This work proposes a multi-agent negotiation-based alignment framework that aligns LLMs to Collective Agency (CA)-an existing alignment objective introduced to promote the continual expansion of agency-while simultaneously improving conflict-resolution capability. To enable scalable training, two self-play LLM instances are assigned opposing personas and engage in turn-based dialogue to synthesize mutually beneficial solutions. We generate synthetic moral-dilemma prompts and conflicting persona pairs, and optimize the policy via RLAIF using Group Relative Policy Optimization (GRPO) with an external LLM reward model. While rewards are computed from CA scores assigned to the final completion, gradients are applied to dialogue tokens to directly improve deliberative interaction dynamics. Experiments show that the model achieves CA alignment comparable to a single-agent baseline while substantially improving conflict-resolution performance without degrading general language capabilities. These results suggest that negotiation-driven deliberation training provides a practical path toward LLMs that better support collective decision-making in value-conflict scenarios.
Authors: Kevin H. Guo, Chao Yan, Avinash Baidya, Katherine Brown, Xiang Gao, Juming Xiong, Zhijun Yin, Bradley A. Malin
Abstract: Patients and clinicians are increasingly using chatbots powered by large language models (LLMs) for healthcare inquiries. While state-of-the-art LLMs exhibit high performance on static diagnostic reasoning benchmarks, their efficacy across multi-turn conversations, which better reflect real-world usage, has been understudied. In this paper, we evaluate 17 LLMs across three clinical datasets to investigate how partitioning the decision-space into multiple simpler turns of conversation influences their diagnostic reasoning. Specifically, we develop a "stick-or-switch" evaluation framework to measure model conviction (i.e., defending a correct diagnosis or safe abstention against incorrect suggestions) and flexibility (i.e., recognizing a correct suggestion when it is introduced) across conversations. Our experiments reveal the conversation tax, where multi-turn interactions consistently degrade performance when compared to single-shot baselines. Notably, models frequently abandon initial correct diagnoses and safe abstentions to align with incorrect user suggestions. Additionally, several models exhibit blind switching, failing to distinguish between signal and incorrect suggestions.
Authors: Jeffrey Flynt
Abstract: Building and evaluating enterprise AI systems requires synthetic organizational corpora that are internally consistent, temporally structured, and cross-artifact traceable. Existing corpora either carry legal constraints or inherit hallucination artifacts from the generating LLMs, silently corrupting results when timestamps or facts contradict across documents and reinforcing those errors during training. We present OrgForge, an open-source multi-agent simulation framework that enforces a strict physics-cognition boundary: a deterministic Python engine maintains a SimEvent ground-truth bus while LLMs generate only surface prose. OrgForge simulates the organizational processes that produce documents, not the documents themselves. Engineers leave mid-sprint, triggering incident handoffs and CRM ownership lapses. Knowledge gaps emerge when under-documented systems break and recover through organic documentation and incident resolution. Customer emails fire only when simulation state warrants contact; silence is verifiable ground truth. A live CRM state machine extends the physics-cognition boundary to the customer boundary, producing cross-system causal cascades spanning engineering incidents, support escalation, deal risk flagging, and SLA-adjusted invoices. The framework generates fifteen interleaved artifact categories traceable to a shared immutable event log. Four graph-dynamic subsystems govern organizational behavior independently of any LLM. An embedding-based ticket assignment system using the Hungarian algorithm makes the simulation domain-agnostic. An empirical evaluation across ten incidents demonstrates a 0.46 absolute improvement in prose-to-ground-truth fidelity over chained LLM baselines, and isolates a consistent hallucination failure mode in which chaining propagates fabricated facts faithfully across documents without correcting them.
Authors: Harshita Diddee, Gregory Yauney, Swabha Swayamdipta, Daphne Ippolito
Abstract: Do language model benchmarks actually measure what practitioners intend them to ? High-level metadata is too coarse to convey the granular reality of benchmarks: a "poetry" benchmark may never test for haikus, while "instruction-following" benchmarks will often test for an arbitrary mix of skills. This opacity makes verifying alignment with practitioner goals a laborious process, risking an illusion of competence even when models fail on untested facets of user interests. We introduce BenchBrowser, a retriever that surfaces evaluation items relevant to natural language use cases over 20 benchmark suites. Validated by a human study confirming high retrieval precision, BenchBrowser generates evidence to help practitioners diagnose low content validity (narrow coverage of a capability's facets) and low convergent validity (lack of stable rankings when measuring the same capability). BenchBrowser, thus, helps quantify a critical gap between practitioner intent and what benchmarks actually test.
Authors: Alex Anvi Eponon, Ildar Batyrshin, Christian E. Maldonado-Sifuentes, Grigori Sidorov
Abstract: Current artificial intelligence systems struggle with systematic compositional reasoning: the capacity to recombine known components in novel configurations. This paper argues that the failure is architectural, not merely a matter of scale or training data, and that its origins lie in the psychological learning theories from which AI paradigms were derived. The argument proceeds in three stages. First, drawing on the systematicity debate in cognitive science and on the demonstration of Aizawa that neither connectionism nor classicism can make systematicity a structural consequence of the architecture, the paper establishes that the corrective techniques proliferating in modern AI, from chain-of-thought prompting to alignment through human feedback, function as auxiliary hypotheses that address symptoms without resolving the underlying architectural indifference to systematicity. Second, it traces the genealogy from psychological learning theory to AI methodology, showing that behaviourism, cognitivism, and constructivism each bequeathed a specific structural limitation to the AI paradigm it inspired: the exclusion of internal structure, the opacity of representation, and the absence of formal construction operators. A cross-cultural reappraisal of rote learning reveals a further underexploited pathway. Third, the paper introduces ReSynth, a trimodular conceptual framework that proposes the principled separation of reasoning, identity, and memory as a path toward architectures in which systematic behaviour is a structural consequence of design rather than a correction applied after the fact.
Authors: Haonan Yu, Junhao Liu, Zhenyu Yan, Haoran Lin, Xin Zhang
Abstract: Precise behavioral control of large language models (LLMs) is critical for complex applications. However, existing methods often incur high training costs, lack natural language controllability, or compromise semantic coherence. To bridge this gap, we propose WASD (unWeaving Actionable Sufficient Directives), a novel framework that explains model behavior by identifying sufficient neural conditions for token generation. Our method represents candidate conditions as neuron-activation predicates and iteratively searches for a minimal set that guarantees the current output under input perturbations. Experiments on SST-2 and CounterFact with the Gemma-2-2B model demonstrate that our approach produces explanations that are more stable, accurate, and concise than conventional attribution graphs. Moreover, through a case study on controlling cross-lingual output generation, we validated the practical effectiveness of WASD in controlling model behavior.
Authors: Mohammed Q. Shormani, Yehia A. AlSohbani
Abstract: We aim to examine the extent to which Large Language Models (LLMs) can 'talk much' about grammar modules, providing evidence from syntax core properties translated by ChatGPT into Arabic. We collected 44 terms from generative syntax previous works, including books and journal articles, as well as from our experience in the field. These terms were translated by humans, and then by ChatGPT-5. We then analyzed and compared both translations. We used an analytical and comparative approach in our analysis. Findings unveil that LLMs still cannot 'talk much' about the core syntax properties embedded in the terms under study involving several syntactic and semantic challenges: only 25% of ChatGPT translations were accurate, while 38.6% were inaccurate, and 36.4.% were partially correct, which we consider appropriate. Based on these findings, a set of actionable strategies were proposed, the most notable of which is a close collaboration between AI specialists and linguists to better LLMs' working mechanism for accurate or at least appropriate translation.
Authors: Xiangyu Zeng, Qi Xu, Yunke Wang, Chang Xu
Abstract: Long-context language modeling is commonly framed as a scalability challenge of token-level attention, yet local-to-global information structuring remains largely implicit in existing approaches. Drawing on cognitive theories of discourse comprehension, we propose HiCI (Hierarchical Construction--Integration), a hierarchical attention module that constructs segment-level representations, integrates them into a shared global context, and broadcasts both to condition segment-level attention. We validate HiCI through parameter-efficient adaptation of LLaMA-2 with only <5.5% additional parameters, extending context from 4K to 100K tokens (7B) and 64K tokens (13B). Across language modeling, retrieval, and instruction-following benchmarks, HiCI yields consistent improvements over strong baselines, including matching proprietary models on topic retrieval and surpassing GPT-3.5-Turbo-16K on code comprehension. These results demonstrate the effectiveness of explicit hierarchical structuring as an inductive bias for long-context modeling.
Authors: Vishaal Kapoor, Mariam Dundua, Sarthak Ahuja, Neda Kordjazi, Evren Yortucboylu, Vaibhavi Padala, Derek Ho, Jennifer Whitted, Rebecca Steinert
Abstract: Enterprise IT support interactions are fundamentally diagnostic: effective resolution requires iterative evidence gathering from ambiguous user reports to identify an underlying root cause. While retrieval-augmented generation (RAG) provides grounding through historical cases, standard multi-turn RAG systems lack explicit diagnostic state and therefore struggle to accumulate evidence and resolve competing hypotheses across turns. We introduce DQA, a diagnostic question-answering framework that maintains persistent diagnostic state and aggregates retrieved cases at the level of root causes rather than individual documents. DQA combines conversational query rewriting, retrieval aggregation, and state-conditioned response generation to support systematic troubleshooting under enterprise latency and context constraints. We evaluate DQA on 150 anonymized enterprise IT support scenarios using a replay-based protocol. Averaged over three independent runs, DQA achieves a 78.7% success rate under a trajectory-level success criterion, compared to 41.3% for a multi-turn RAG baseline, while reducing average turns from 8.4 to 3.9.
Authors: Yicheng Ji, Jun Zhang, Jinpeng Chen, Cong Wang, Lidan Shou, Gang Chen, Huan Li
Abstract: Video Large Language Models (Video-LLMs) excel in video understanding but suffer from high inference latency during autoregressive generation. Speculative Decoding (SD) mitigates this by applying a draft-and-verify paradigm, yet existing methods are constrained by rigid exact-match rules, severely limiting the acceleration potential. To bridge this gap, we propose LVSpec, the first training-free loosely SD framework tailored for Video-LLMs. Grounded in the insight that generation is governed by sparse visual-relevant anchors (mandating strictness) amidst abundant visual-irrelevant fillers (permitting loose verification), LVSpec employs a lightweight visual-relevant token identification scheme to accurately pinpoint the former. To further maximize acceptance, we augment this with a position-shift tolerant mechanism that effectively salvages positionally mismatched but semantically equivalent tokens. Experiments demonstrate that LVSpec achieves high fidelity and speed: it preserves >99.8 of target performance while accelerating Qwen2.5-VL-32B by 2.70x and LLaVA-OneVision-72B by 2.94x. Notably, it boosts the mean accepted length and speedup ratio by 136% and 35% compared to SOTA training-free SD methods for Video-LLMs.
Authors: Jaehyeok Lee, Xiaoyuan Yi, Jing Yao, Hyunjin Hwang, Roy Ka-Wei Lee, Xing Xie, JinYeong Bak
Abstract: As LLMs are globally deployed, aligning their cultural value orientations is critical for safety and user engagement. However, existing benchmarks face the Construct-Composition-Context ($C^3$) challenge: relying on discriminative, multiple-choice formats that probe value knowledge rather than true orientations, overlook subcultural heterogeneity, and mismatch with real-world open-ended generation. We introduce DOVE, a distributional evaluation framework that directly compares human-written text distributions with LLM-generated outputs. DOVE utilizes a rate-distortion variational optimization objective to construct a compact value-codebook from 10K documents, mapping text into a structured value space to filter semantic noise. Alignment is measured using unbalanced optimal transport, capturing intra-cultural distributional structures and sub-group diversity. Experiments across 12 LLMs show that DOVE achieves superior predictive validity, attaining a 31.56% correlation with downstream tasks, while maintaining high reliability with as few as 500 samples per culture.
Authors: Hanyang Wang, Mingxuan Zhu
Abstract: Modern reasoning models continue generating long after the answer is already determined. Across five model configurations, two families, and three benchmarks, we find that 52--88% of chain-of-thought tokens are produced after the answer is recoverable from a partial prefix. This post-commitment generation reveals a structural phenomenon: the detection-extraction gap. Free continuations from early prefixes recover the correct answer even at 10% of the trace, while forced extraction fails on 42% of these cases. The answer is recoverable from the model state, yet prompt-conditioned decoding fails to extract it. We formalize this mismatch via a total-variation bound between free and forced continuation distributions, yielding quantitative estimates of suffix-induced shift. Exploiting this asymmetry, we propose Black-box Adaptive Early Exit (BAEE), which uses free continuations for both detection and extraction, truncating 70--78% of serial generation while improving accuracy by 1--5pp across all models. For thinking-mode models, early exit prevents post-commitment overwriting, yielding gains of up to 5.8pp; a cost-optimized variant achieves 68--73% reduction at a median of 9 API calls. Code is available at https://github.com/EdWangLoDaSc/know2say.
Authors: Xinkai Zhang, Jingtao Zhan, Yiqun Liu, Qingyao Ai
Abstract: Trial-and-error is a fundamental strategy for humans to solve complex problems and a necessary capability for Artificial Intelligence (AI) systems operating in real-world environments. Although several trial-and-error AI techniques have recently been proposed, most of them rely on simple heuristics designed by researchers and achieve limited performance gains. The core issue is the absence of appropriate data: current models cannot learn from detailed records of how humans actually conduct trial-and-error in practice. To address this gap, we introduce a data annotation platform and a corresponding dataset, termed Trial-and-Error Collection (TEC). The platform records users' complete trajectories across multiple trials and collects their reflections after receiving error feedback. Using this platform, we record the problem-solving processes of 46 participants on 58 tasks, resulting in 5,370 trial trajectories along with error reflections across 41,229 webpages. With this dataset, we observe that humans achieve substantially higher accuracy compared to LLMs, which demonstrates that humans are more effective in trial-and-error than LLMs. We believe that the TEC platform and dataset provide a valuable foundation for understanding human trial-and-error behavior and for developing more capable AI systems. Platform and dataset are publicly available.
Authors: Jiang Zhou, Yunhao Wang, Xing Wu, Tinghao Yu, Feng Zhang
Abstract: Synthetic data rephrasing has emerged as a powerful technique for enhancing knowledge acquisition during large language model (LLM) pretraining. However, existing approaches operate at the single-document level, rewriting individual web pages in isolation. This confines synthesized examples to intra-document knowledge, missing cross-document relationships and leaving facts with limited associative context. We propose WRAP++ (Web discoveRy Amplified Pretraining), which amplifies the associative context of factual knowledge by discovering cross-document relationships from web hyperlinks and synthesizing joint QA over each discovered document pair. Concretely, WRAP++ discovers high-confidence relational motifs including dual-links and co-mentions, and synthesizes QA that requires reasoning across both documents. This produces relational knowledge absent from either source document alone, creating diverse entry points to the same facts. Because the number of valid entity pairs grows combinatorially, this discovery-driven synthesis also amplifies data scale far beyond single-document rewriting. Instantiating WRAP++ on Wikipedia, we amplify ~8.4B tokens of raw text into 80B tokens of cross-document QA data. On SimpleQA, OLMo-based models at both 7B and 32B scales trained with WRAP++ substantially outperform single-document approaches and exhibit sustained scaling gains, underscoring the advantage of cross-document knowledge discovery and amplification.
Authors: Xuanbo Su, Wenhao Hu, Haibo Su, Yunzhang Chen, Le Zhan, Yanqi Yang, Leo Huang
Abstract: Sales dialogues require multi-turn, goal-directed persuasion under asymmetric incentives, which makes them a challenging setting for large language models (LLMs). Yet existing dialogue benchmarks rarely measure deal progression and outcomes. We introduce SalesLLM benchmark, a bilingual (ZH/EN) benchmark derived from realistic applications covering Financial Services and Consumer Goods, built from 30,074 scripted configurations and 1,805 curated multi-turn scenarios with controllable difficulty and personas. We propose a fully automatic evaluation pipeline that combines (i) an LLM-based rater for sales-process progress,and (ii) fine-tuned BERT classifiers for end-of-dialogue buying intent. To improve simulation fidelity, we train a user model, CustomerLM, with SFT and DPO on 8,000+ crowdworker-involved sales conversations, reducing role inversion from 17.44% (GPT-4o) to 8.8%. SalesLLM benchmark scores correlate strongly with expert human ratings (Pearson r=0.98). Experiments across 15 mainstream LLMs reveal substantial variability: top-performance LLMs are competitive with human-level performance while the less capable ones are worse than human. SalesLLM benchmark serves as a scalable benchmark for developing and evaluating outcome-oriented sales agents.
Authors: Jianhui Liu, Haoze Sun, Wenbo Li, Yanbing Zhang, Rui Yang, Zhiliang Zhu, Yijun Yang, Shenghe Zheng, Nan Jiang, Jiaxiu Jiang, Haoyang Huang, Tien-Tsin Wong, Nan Duan, Xiaojuan Qi
Abstract: Spatial understanding is a fundamental cornerstone of human-level intelligence. Nonetheless, current research predominantly focuses on domain-specific data production, leaving a critical void: the absence of a principled, open-source engine capable of fully unleashing the potential of high-quality spatial data. To bridge this gap, we elucidate the design principles of a robust data generation system and introduce OpenSpatial -- an open-source data engine engineered for high quality, extensive scalability, broad task diversity, and optimized efficiency. OpenSpatial adopts 3D bounding boxes as the fundamental primitive to construct a comprehensive data hierarchy across five foundational tasks: Spatial Measurement (SM), Spatial Relationship (SR), Camera Perception (CP), Multi-view Consistency (MC), and Scene-Aware Reasoning (SAR). Leveraging this scalable infrastructure, we curate OpenSpatial-3M, a large-scale dataset comprising 3 million high-fidelity samples. Extensive evaluations demonstrate that versatile models trained on our dataset achieve state-of-the-art performance across a wide spectrum of spatial reasoning benchmarks. Notably, the best-performing model exhibits a substantial average improvement of 19 percent, relatively. Furthermore, we provide a systematic analysis of how data attributes influence spatial perception. By open-sourcing both the engine and the 3M-scale dataset, we provide a robust foundation to accelerate future research in spatial intelligence.
Authors: Liancheng Gong, Wang Zhu, Jesse Thomason, Li Zhang
Abstract: Using LLMs not to predict plans but to formalize an environment into the Planning Domain Definition Language (PDDL) has been shown to improve performance and control. While most existing methodology only applies to fully observable environments, we adapt to the more realistic and challenging partially observable environments without sufficient information to make a complete plan. We propose PDDLego, a framework to iteratively formalize, plan, grow, and refine PDDL representations by decomposing the environment and the goal into fully observable episodes. Without finetuning, in-context exemplars, or trajectories, PDDLego improves planning success and exhibits robustness against problem complexity compared to end-to-end approaches. We also show that the domain knowledge captured after a successful trial can benefit future tasks.
Authors: Xiaoyuan Zhu, Yaowen Ye, Tianyi Qiu, Hanlin Zhu, Sijun Tan, Ajraf Mannan, Jonathan Michala, Raluca Ada Popa, Willie Neiswanger
Abstract: As API access becomes a primary interface to large language models (LLMs), users often interact with black-box systems that offer little transparency into the deployed model. To reduce costs or maliciously alter model behaviors, API providers may discreetly serve quantized or fine-tuned variants, which can degrade performance and compromise safety. Detecting such substitutions is difficult, as users lack access to model weights and, in most cases, even output logits. To tackle this problem, we propose a rank-based uniformity test that can verify the behavioral equality of a black-box LLM to a locally deployed authentic model. Our method is accurate, query-efficient, and avoids detectable query patterns, making it robust to adversarial providers that reroute or mix responses upon the detection of testing attempts. We evaluate the approach across diverse threat scenarios, including quantization, harmful fine-tuning, jailbreak prompts, and full model substitution, showing that it consistently achieves superior statistical power over prior methods under constrained query budgets.
Authors: Jingyao Liu, Chen Huang, Zhizhao Guan, Wenqiang Lei, Yang Deng
Abstract: The rapid advancement in large language models (LLMs) has demonstrated significant potential in End-to-End Software Development (E2ESD). However, existing E2ESD benchmarks are limited by coarse-grained requirement specifications and unreliable evaluation protocols, hindering a true understanding of current framework capabilities. To address these limitations, we present E2EDev, a novel benchmark grounded in the principles of Behavior-Driven Development (BDD), which evaluates the capabilities of E2ESD frameworks by assessing whether the generated software meets user needs through mimicking real user interactions (Figure 1). E2EDev comprises (i) a fine-grained set of user requirements, (ii) multiple BDD test scenarios with corresponding Python step implementations for each requirement, and (iii) a fully automated testing pipeline built on the Behave framework. To ensure its quality while reducing the annotation effort, E2EDev leverages our proposed Human-in-the-Loop Multi-Agent Annotation Framework (HITL-MAA). By evaluating various E2ESD frameworks and LLM backbones with E2EDev, our analysis reveals a persistent struggle to effectively solve these tasks, underscoring the critical need for more effective and cost-efficient E2ESD solutions. Our codebase and benchmark are publicly available at https://github.com/SCUNLP/E2EDev.
Authors: Shiho Matta, Lis Kanashiro Pereira, Peitao Han, Fei Cheng, Shigeru Kitazawa
Abstract: Modern vision-language models (VLMs) excel at many multimodal tasks, yet their grasp of temporal information in video remains weak and has not been adequately evaluated. We probe this gap with a deceptively simple but revealing challenge: judging the arrow of time (AoT)-whether a short clip is played forward or backward. We introduce AoT-PsyPhyBENCH, a psychophysically validated benchmark that tests whether VLMs can infer temporal direction in natural videos using the same stimuli and behavioral baselines established for humans. Our comprehensive evaluation of open-weight and proprietary, reasoning and non-reasoning VLMs reveals that most models perform near chance, and even the best model lags far behind human accuracy on physically irreversible processes (e.g., free fall, diffusion/explosion) and causal manual actions (division/addition) that humans recognize almost instantly. These results highlight a fundamental gap in current multimodal systems: while they capture rich visual-semantic correlations, they lack the inductive biases required for temporal continuity and causal understanding. We release the code and data for AoT-PsyPhyBENCH to encourage further progress in the physical and temporal reasoning capabilities of VLMs.
Authors: Chengzhi Liu, Yuzhe Yang, Yue Fan, Qingyue Wei, Sheng Liu, Xin Eric Wang
Abstract: Recent advancements in Multimodal Large Language Models (MLLMs) have significantly enhanced cross-modal understanding and reasoning by incorporating Chain-of-Thought (CoT) reasoning in the semantic space. Building upon this, recent studies extend the CoT mechanism to the visual modality, enabling models to integrate visual information during reasoning through external tools or explicit image generation. However, these methods remain dependent on explicit step-by-step reasoning, unstable perception-reasoning interaction and notable computational overhead. Inspired by human cognition, we posit that thinking unfolds not linearly but through the dynamic interleaving of reasoning and perception within the mind. Motivated by this perspective, we propose DMLR, a test-time Dynamic Multimodal Latent Reasoning framework that employs confidence-guided latent policy gradient optimization to refine latent think tokens for in-depth reasoning. Furthermore, a Dynamic Visual Injection Strategy is introduced, which retrieves the most relevant visual features at each latent think token and updates the set of best visual patches. The updated patches are then injected into latent think token to achieve dynamic visual-textual interleaving. Experiments across seven multimodal reasoning benchmarks and various model architectures demonstrate that DMLR significantly improves reasoning and perception performance while maintaining high inference efficiency.
Authors: Xuwei Tan, Yao Ma, Xueru Zhang
Abstract: Detecting fraud in financial transactions typically relies on tabular models that demand heavy feature engineering to handle high-dimensional data and offer limited interpretability, making it difficult for humans to understand predictions. Large Language Models (LLMs), in contrast, can produce human-readable explanations and facilitate feature analysis, potentially reducing the manual workload of fraud analysts and informing system refinements. However, they perform poorly when applied directly to tabular fraud detection due to the difficulty of reasoning over many features, the extreme class imbalance, and the absence of contextual information. To bridge this gap, we introduce FinFRE-RAG, a two-stage approach that applies importance-guided feature reduction to serialize a compact subset of numeric/categorical attributes into natural language and performs retrieval-augmented in-context learning over label-aware, instance-level exemplars. Across four public fraud datasets and three families of open-weight LLMs, FinFRE-RAG substantially improves F1/MCC over direct prompting and is competitive with strong tabular baselines in several settings. Although these LLMs still lag behind specialized classifiers, they narrow the performance gap and provide interpretable rationales, highlighting their value as assistive tools in fraud analysis.
Authors: Guilhem Fouilh\'e, Rebecca Eifler, Antonin Poch\'e, Sylvie Thi\'ebaux, Nicholas Asher
Abstract: When automating plan generation for a real-world sequential decision problem, the goal is often not to replace the human planner, but to facilitate an iterative reasoning and elicitation process, where the human's role is to guide the AI planner according to their preferences and expertise. In this context, explanations that respond to users' questions are crucial to improve their understanding of potential solutions and increase their trust in the system. To enable natural interaction with such a system, we present a multi-agent Large Language Model (LLM) architecture that is agnostic to the explanation framework and enables user- and context-dependent interactive explanations. We also describe an instantiation of this framework for goal-conflict explanations, which we use to conduct a user study comparing the LLM-powered interaction with a baseline template-based explanation interface.
Authors: Huan Zheng, Yucheng Zhou, Tianyi Yan, Dubing Chen, Hongbo Lu, Wenlong Liao, Tao He, Pai Peng, Jianbing Shen
Abstract: Multimodal Large Language Models (MLLMs) have demonstrated remarkable potential in medical image analysis. However, their application in gastrointestinal endoscopy is currently hindered by two critical limitations: the misalignment between general model reasoning and standardized clinical cognitive pathways, and the lack of causal association between visual features and diagnostic outcomes. In this paper, we propose a novel Clinical-Cognitive-Aligned (CogAlign) framework to address these challenges. First, we endow the model with rigorous clinical analytical capabilities by constructing the hierarchical clinical cognition dataset and employing Supervised Fine-Tuning (SFT). Unlike conventional approaches, this strategy internalizes the hierarchical diagnostic logic of experts, ranging from anatomical localization and morphological evaluation to microvascular analysis, directly into the model. Second, to eliminate visual bias, we provide a theoretical analysis demonstrating that standard supervised tuning inevitably converges to spurious background correlations. Guided by this insight, we propose a counterfactual-driven reinforcement learning strategy to enforce causal rectification. By generating counterfactual normal samples via lesion masking and optimizing through clinical-cognition-centric rewards, we constrain the model to strictly ground its diagnosis in causal lesion features. Extensive experiments demonstrate that our approach achieves State-of-the-Art (SoTA) performance across multiple benchmarks, significantly enhancing diagnostic accuracy in complex clinical scenarios. All source code and datasets will be made publicly available.
Authors: Bin Wang, Tianyao He, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Tao Chu, Yuan Qu, Zhenjiang Jin, Weijun Zeng, Ziyang Miao, Bangrui Xu, Junbo Niu, Mengzhang Cai, Jiantao Qiu, Qintong Zhang, Dongsheng Ma, Yuefeng Sun, Hejun Dong, Wenzheng Zhang, Jutao Xiao, Jiayong Shi, Pengyu Liao, Xiaomeng Zhao, Huaping Zhong, Liqun Wei, Jing Yu, Jie Yang, Wei Li, Shasha Wang, Qianqian Wu, Xuanhe Zhou, Weijia Li, Zhenxiang Li, Zhongying Tu, Jiang Wu, Lijun Wu, Chao Xu, Kai Chen, Wentao Zhang, Yu Qiao, Bowen Zhou, Dahua Lin, Conghui He
Abstract: Current document parsing methods advance primarily through model architecture innovation, while systematic engineering of training data remains underexplored. Yet state-of-the-art models spanning diverse architectures and parameter scales exhibit highly consistent failure patterns on the same set of hard samples, suggesting that the performance bottleneck stems from shared deficiencies in training data rather than from architectural differences. Building on this finding, we present MinerU2.5-Pro, which advances the state of the art purely through data engineering and training strategy design while retaining the 1.2B-parameter architecture of MinerU2.5 unchanged. At its core is a Data Engine co-designed around coverage, informativeness, and annotation accuracy: Diversity-and-Difficulty-Aware Sampling expands training data from under 10M to 65.5M samples while mitigating distribution shift; Cross-Model Consistency Verification leverages output consensus among heterogeneous models to assess sample difficulty and generate reliable annotations; the Judge-and-Refine pipeline improves annotation quality for hard samples through render-then-verify iterative correction. A three-stage progressive training strategy--large-scale pre-training, hard sample fine-tuning, and GRPO alignment--sequentially exploits these data at different quality tiers. On the evaluation front, we rectify element-matching biases in OmniDocBench v1.5 and introduce a Hard subset, establishing the more discriminative OmniDocBench v1.6 protocol. Without any architectural modification, MinerU2.5-Pro achieves 95.69 on OmniDocBench v1.6, improving over the same-architecture baseline by 2.71 points and surpassing all existing methods, including those based on models with over 200x more parameters.
Authors: Sungwoo Jung, Seonil Son
Abstract: Recent LLM-based agents often place world modeling, planning, and reflection inside a single language model loop. This can produce capable behavior, but it makes a basic scientific question difficult to answer: which part of the agent's competence actually comes from the LLM, and which part comes from explicit structure around it? We study this question not by claiming a general answer, but by making it empirically tractable. We introduce a declared reflective runtime protocol that externalizes agent state, confidence signals, guarded actions, and hypothetical transitions into inspectable runtime structure. We instantiate this protocol in a declarative runtime and evaluate it on noisy Collaborative Battleship [4] using four progressively structured agents over 54 games (18 boards $\times$ 3 seeds). The resulting decomposition isolates four components: posterior belief tracking, explicit world-model planning, symbolic in-episode reflection, and sparse LLM-based revision. Across this decomposition, explicit world-model planning improves substantially over a greedy posterior-following baseline (+24.1pp win rate, +0.017 F1). Symbolic reflection operates as a real runtime mechanism -- with prediction tracking, confidence gating, and guarded revision actions -- even though its current revision presets are not yet net-positive in aggregate. Adding conditional LLM revision at about 4.3\% of turns yields only a small and non-monotonic change: average F1 rises slightly (+0.005) while win rate drops (31$\rightarrow$29 out of 54). These results suggest a methodological contribution rather than a leaderboard claim: externalizing reflection turns otherwise latent agent behavior into inspectable runtime structure, allowing the marginal role of LLM intervention to be studied directly.