As large language models are deployed in high-stakes enterprise applications, from healthcare to finance, ensuring adherence to organization-specific policies has become essential. Yet existing safety evaluations focus exclusively on universal harms. We present COMPASS (Company/Organization Policy Alignment Assessment), the first systematic framework for evaluating whether LLMs comply with organizational allowlist and denylist policies. We apply COMPASS to eight diverse industry scenarios, generating and validating 5,920 queries that test both routine compliance and adversarial robustness through strategically designed edge cases. Evaluating seven state-of-the-art models, we uncover a fundamental asymmetry: models reliably handle legitimate requests (>95% accuracy) but catastrophically fail at enforcing prohibitions, refusing only 13–40% of adversarial denylist violations. These results demonstrate that current LLMs lack the robustness required for policy-critical deployments, establishing COMPASS as an essential evaluation framework for organizational AI safety.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
BioTool: A Comprehensive Tool-Calling Dataset for Enhancing Biomedical Capabilities of Large Language Models
PDF ↗Despite the success of large language models (LLMs) on general-purpose tasks, their performance in highly specialized domains such as biomedicine remains unsatisfactory. A key limitation is the inability of LLMs to effectively leverage biomedical tools, which clinical experts and biomedical researchers rely on extensively in daily workflows. While recent general-domain tool-calling datasets have substantially improved the capabilities of LLM agents, existing efforts in the biomedical domain largely rely on in-context learning and restrict models to a small set of tools. To address this gap, we introduce BioTool, a comprehensive biomedical tool-calling dataset designed for fine-tuning LLMs. BioTool comprises 34 frequently used tools collected from the NCBI, Ensembl, and UniProt databases, along with 7,040 high-quality, human-verified query–API call pairs spanning variation, genomics, proteomics, evolution, and general biology. Fine-tuning a 4-billion-parameter LLM on BioTool yields substantial improvements in biomedical tool-calling performance, outperforming cutting-edge commercial LLMs such as GPT-5.1. Furthermore, human expert evaluations demonstrate that integrating a BioTool-fine-tuned tool caller significantly improves downstream answer quality compared to the same LLM without tool usage, highlighting the effectiveness of BioTool in enhancing the biomedical capabilities of LLMs. The full dataset and evaluation code are available at https://github.com/gxx27/BioTool.
Mind the Pause: Disfluency-Aware Objective Tuning for Multilingual Speech Correction with LLMs
PDF ↗Automatic Speech Recognition (ASR) transcripts often contain disfluencies, such as fillers, repetitions, and false starts, which reduce readability and hinder downstream applications like chatbots and voice assistants. If left unaddressed, such disfluencies can significantly degrade the reliability of downstream systems. Most existing approaches rely on classical models that focus on identifying disfluent tokens for removal. While this strategy is effective to some extent, it often disrupts grammatical structure and semantic coherence, leading to incomplete or unnatural sentences. Recent literature explored the use of large language models (LLMs); however, these efforts have primarily focused on disfluency detection or data augmentation, rather than performing comprehensive correction. We propose a multilingual correction pipeline where a sequence tagger first marks disfluent tokens, and these signals guide instruction fine-tuning of an LLM to rewrite transcripts into fluent text. To further improve reliability, we add a contrastive learning objective that penalizes the reproduction of disfluent tokens, encouraging the model to preserve grammar and meaning while removing disfluent artifacts. Our experiments across three Indian languages, namely Hindi, Bengali, and Marathi show consistent improvements over strong baselines, including multilingual sequence-to-sequence models. These results highlight that detection-only strategies are insufficient. Combining token-level cues with instruction tuning and contrastive learning provides a practical and scalable solution for multilingual disfluency correction in speech-driven NLP systems. We make the codes publicly available at https://github.com/deepak-kumar-98/Mind-the-Pause.
LogicEval: A Systematic Framework for Evaluating Automated Repair Techniques for Logical Vulnerabilities in Real-World Software
PDF ↗Logical vulnerabilities in software stem from flaws in program logic rather than memory safety, which can lead to critical security failures. Although existing automated program-repair techniques primarily focus on repairing memory-corruption vulnerabilities, they struggle with logical vulnerabilities because of their limited semantic understanding of the vulnerable code and its expected behavior. On the other hand, recent successes of large language models (LLMs) in understanding and repairing code are promising. However, no framework currently exists to analyze the capabilities and limitations of such techniques for logical vulnerabilities. We aim to systematically evaluate both traditional and LLM-based repair approaches for addressing real-world logical vulnerabilities. To facilitate our assessment, we created the first-ever dataset, LogicDS, comprising 122 logical vulnerabilities that reflect tangible security impact. We also developed a systematic framework, LogicEval, to evaluate patches for logical vulnerabilities. Evaluations suggest that compilation and testing failures are primarily driven by prompt sensitivity, loss of code context, and difficulty in patch localization.
From Recognition to Reasoning: Benchmarking and Enhancing MLLMs on Real-World Receipt Document Understanding
PDF ↗Extracting structured information from visual documents (Visual Information Extraction, VIE) is a cornerstone of business automation. While recent Multimodal Large Language Models (MLLMs) have shown promising capabilities, existing benchmarks suffer from critical limitations in scale and realism, lack semantic granularity, and fail to cover diverse document types. To bridge this gap, we introduce ReceiptBench, a large-scale, human-annotated benchmark consisting of 10k diverse receipts, organizing information extraction into four hierarchical sub-tasks: (1) Basic Perception for raw text spotting, (2) Format Normalization for strictly following standardization instructions, (3) Semantic Reasoning for inferring implicit attributes from context, and (4) Structure Parsing for handling nested line items. Furthermore, we propose a two-stage training framework incorporating Metric-Aware Group Relative Policy Optimization (GRPO), which translates rigorous evaluation constraints into reinforcement learning signals to enhance structural consistency. Extensive experiments demonstrate that our method yields state-of-the-art performance, surpassing leading proprietary models on complex reasoning tasks. We release our datasets and code at https://github.com/wwwT0ri/ReceiptBench.
Persona-driven simulations are increasingly used in computational social science, yet their validity critically depends on the fidelity of the underlying personas. Constructing virtual populations that are both authentic and scalable remains a central challenge. We introduce Synthia, a persona-generation framework that grounds LLM-generated personas in real social-media posts while delegating narrative construction to language models, using publicly available data from the Bluesky platform. Across multiple social-survey benchmarks, Synthia improves alignment with human opinion distributions over prior state-of-the-art approaches while relying on substantially smaller models. A multi-dimensional fairness and bias analysis shows that Synthia outperforms previous methods for most demographics across different dimensions. Uniquely, Synthia preserves interaction-graph structure among personas grounded in real social network users, enabling network-aware analysis, which we demonstrate through two homophily-focused case studies. Together, these results position Synthia as a practical and reliable framework for constructing scalable, high-fidelity, and equitable virtual populations.
Language models are increasingly deployed in interactive settings where users reason about facts over time rather than in isolation.In such scenarios, correct behavior requires models to maintain and update implicit temporal assumptions established earlier in a conversation. We study this challenge through the lens of temporal scope stability: the ability to preserve, override, or transfer time-scoped factual context across dialogue turns. We introduce ChronoScope, a large-scale diagnostic benchmark designed to isolate temporal scope behavior in controlled multi-turn interactions, comprising over one million deterministically generated question chains grounded in Wikidata. ChronoScope evaluates whether models can correctly retain inferred temporal scope when follow-up questions omit explicit time references, spanning implicit carryover, explicit scope switching, cross-entity transfer, and longer temporal trajectories.Through extensive evaluation of state-of-the-art language models, we find that temporal scope stability is frequently violated in controlled multi-turn settings, with models often drifting toward present-day assumptions despite correct underlying knowledge.These failures intensify with interaction length and persist even under oracle context conditions, revealing a gap between single-turn factual accuracy and coherent temporal reasoning under sequential interaction.We make our dataset and evaluation suite publicly available at https://github.com/yashkumaratri/ChronoScope.
CxMP: A Linguistic Minimal-Pair Benchmark for Evaluating Constructional Understanding in Language Models
PDF ↗Understanding language acquisition in language models remains an open question, yet many benchmarks focus on grammatical acceptability, with far less attention to interpreting meanings conveyed by grammatical forms.We introduce the Linguistic Minimal-Pair Benchmark for Evaluating Constructional Understanding in Language Models (CxMP), grounded in Construction Grammar, which treats form–meaning pairings (constructions) as fundamental linguistic units.It evaluates whether models interpret the semantic information implied by constructions, using a controlled minimal-pairs across nine types.Our results show that constructional understanding develops more gradually and remains limited for some constructions even in large language models (LLMs), whereas performance on grammatical acceptability emerges earlier, with shallow heuristics in CxMP exhibiting a U-shaped pattern.These findings highlight the need to broaden existing linguistic evaluations to capture meanings encoded in linguistic form.
Diffusion Large Language Models (dLLMs) have emerged as a promising non-autoregressive paradigm for text generation, offering parallel decoding and bidirectional context modeling. However, aligning dLLMs with reinforcement learning (RL) remains a significant challenge, as the marginal likelihood of sequences in masked diffusion is typically intractable, rendering standard policy gradient methods unstable or computationally prohibitive. In this work, we propose **Diffusion-Gibbs Alignment (DGA)**, a novel variational framework that reformulates RL for dLLMs as a distribution matching problem. DGA bypasses the explicit computation of log-probabilities by leveraging a learned energy function to model the relative quality of samples. The optimization is decoupled into two stable steps: (1) contrastive energy ranking to capture global reward structures, and (2) weighted diffusion alignment to update the policy via importance sampling. Empirically, DGA establishes a new state-of-the-art across logical reasoning (Sudoku, Countdown), mathematical reasoning (GSM8K, Math500), and code generation (HumanEval, MBPP) benchmarks. DGA offers a novel variational perspective for dLLM alignment, achieving better performance while simultaneously enhancing training speed and memory efficiency.
When users submit queries to Large Language Models (LLMs), their prompts can often contain sensitive data, forcing a difficult choice: Send the query to a powerful proprietary LLM providers to achieving state-of-the-art performance and risk data exposure, or relying on smaller, local models guarantees data privacy but often results in a degradation of task performance. Prior approaches have relied on static pipelines that use LLM rewriting, which shatters linguistic coherence and indiscriminately removes privacy-sensitive information, including task-critical content. We reformulate this challenge (Privacy-Conscious Delegation) as a sequential decision-making problem and introduce a novel reinforcement learning (RL) framework called Privacy-R1 to solve it. Our framework trains an agent to dynamically route text chunks, learning a policy that optimally balances the trade-off between privacy leakage and task performance. It implicitly distinguishes between replaceable Personally Identifiable Information (PII) (which it shields locally) and task-critical PII (which it strategically sends to the remote model for maximal utility). To validate our approach in complex scenarios, we also introduce a new medical dataset with high PII density. Our framework achieves a new state-of-the-art on the privacy-utility frontier, demonstrating the necessity of learned, adaptive policies for deploying LLMs in sensitive environments. Dataset can be found at: https://github.com/zackhuiiiii/Privacy-R1.
Multi-page Document Visual Question Answering requires reasoning over semantics, layouts, and visual elements in long, visually dense documents. Existing OCR-free methods face a trade-off between capacity and precision: end-to-end models scale poorly with document length, while visual retrieval-based pipelines are brittle and passive. We propose Doc-V^*, an OCR-free agentic framework that casts multi-page DocVQA as sequential evidence aggregation. Doc-V^* begins with a thumbnail overview, then actively navigates via semantic retrieval and targeted page fetching, and aggregates evidence in a structured working memory for grounded reasoning. Trained by imitation learning from expert trajectories and further optimized with Group Relative Policy Optimization, Doc-V^* balances answer accuracy with evidence-seeking efficiency. Across five benchmarks, Doc-V^* outperforms open-source baselines and approaches proprietary models, improving out-of-domain performance by up to 47.9% over RAG baseline. Other results reveal effective evidence aggregation with selective attention, not increased input pages.
When High Accuracy Hides Poor Calibration: Rethinking Confidence Evaluation in Transformer-Based Text Classification with Balanced Brier Score
PDF ↗Transformer-based Small (SLMs) and Large Language Models (LLMs) achieve strong effectiveness in text classification (TC), yet deployment requires reliable confidence estimates. Although miscalibration in Transformers has been reported, evidence for TC under fine-tuning remains limited. We evaluate the calibration of fine-tuned SLMs and LLMs against Logistic Regression, a classical, well-calibrated baseline, and find that, despite superior effectiveness, Transformers remain markedly overconfident. Crucially, we show that widely used calibration metrics, such as Expected Calibration Error and Brier Score, become biased in high-effectiveness regimes, where the dominance of correct predictions masks severe miscalibration on errors, sometimes even suggesting better calibration than Logistic Regression, a well-known calibrated method. To address this limitation, we propose the Balanced Brier Score (BBS), which balances the contribution of correct and incorrect predictions within confidence bins. BBS reveals substantially poorer calibration in both SLMs and LLMs, consistent with qualitative evidence from calibration curves. These findings challenge current calibration assessment practices and provide a more reliable alternative for evaluating confidence quality in Transformer-based TC.
Dense retrieval has become a core technique in applications like web search and retrieval-augmented generation. Despite their empirical success, it remains unclear whether these models truly understand semantics. To address this gap, this paper conducts a systematic investigation by introducing SURE, a benchmark for Semantic Understanding in dense REtrieval built upon the MSMARCO, NQ, and FiQA datasets. SURE characterizes semantic understanding in dense retrieval along three dimensions: semantic precision, semantic abstraction, and semantic equivalence. We evaluate ten representative models ranging from 110M to 8B parameters, including both general-purpose and domain-specific models. Results show that current dense retrievers struggle to distinguish fine-grained semantic differences across texts with varying information density, and to recognize semantic consistency under lexical paraphrasing. Moreover, larger models do not necessarily exhibit stronger semantic understanding, and diverse training data generally enhances semantic understanding on challenging retrieval tasks.
Coreference resolution is typically evaluated using aggregate statistical metrics such as CoNLL-F1, which measure structural overlap between predicted and gold clusters. While widely used, these metrics offer limited diagnostic insights, penalizing errors without revealing whether a system struggles with specific semantic categories, such as people, locations, or events, and making it difficult to interpret model capabilities or derive actionable improvements. We address this gap by introducing a semantically-enhanced evaluation framework for coreference resolution. Our approach overlays Concept and Named Entity Recognition (CNER) onto coreference outputs, assigning semantic labels to nominal mentions and propagating them to entire coreference clusters. This enables the computation of typed scores aimed at evaluating mention extraction and linking capabilities stratified by semantic class. Across our experiments on OntoNotes, LitBank, and PreCo, we show that our framework uncovers systematic weaknesses that remain obscured by aggregate metrics. Furthermore, we show that these diagnostics can be used to design targeted, low-cost data augmentation strategies, achieving measurable out-of-domain improvements.
Large Language Models (LLMs) are increasingly deployed as long-term interactive agents, yet their limited context windows make it difficult to sustain coherent behavior over extended interactions. Existing memory systems for LLMs often store isolated records and retrieve fragments, limiting their ability to consolidate evolving experience and resolve conflicts. We introduce EverMemOS, a self-organizing memory operating system that implements an engram-inspired lifecycle for computational memory. First, Episodic Trace Formation converts dialogue streams into MemCells that capture episodic traces, atomic facts, and time-bounded foresight. Second, Semantic Consolidation organizes MemCells into thematic MemScenes, distilling stable semantic structures and updating user profiles. Finally, Reconstructive Recollection performs MemScene-guided agentic retrieval to compose the necessary and sufficient context for downstream reasoning. Experiments on LoCoMo, LongMemEval, and PersonaMem-v2 show that EverMemOS significantly outperforms state-of-the-art methods on memory-augmented reasoning tasks.
Mind’s Eye: A Benchmark of Visual Abstraction, Transformation and Composition for Multimodal LLMs
PDF ↗Multimodal large language models (MLLMs) have achieved impressive progress on vision language benchmarks, yet their capacity for cognitive and psychological reasoning remains largely unexplored. We introduce Mind’s Eye, a multiple-choice benchmark of eight visuo-cognitive tasks inspired by classic human intelligence tests and organized under a novel A–R–T taxonomy: Abstraction, Relation, and Transformation. The tasks probe core processes of fluid intelligence such as pattern induction, analogical Relation mapping, and mental Transformation. We evaluate a diverse suite of closed-source and open-source MLLMs and compare their performance with human participants. Humans achieve 80% accuracy, while top performing MLLMs remain below 50%. Error analysis reveals failures in (i) visual attention allocation, (ii) internal perceptual manipulation, (iii) over reliance on domain priors, and (iv) weak abstraction of underlying visual concepts. Our findings suggest that current MLLMs exhibit limited fluid reasoning and visuo-cognitive integration compared with human participants, highlighting the need for cognitively grounded evaluation frameworks like Mind’s Eye.
This survey reviews LLM-based multi-agent systems for clinical and healthcare workflows, including diagnosis, triage, consultation, discharge, mental health, and EHR-linked decision support. We define AI hospitals as workflow-level clinical systems in which agents take explicit roles, hand off shared state, use EHR- or guideline-grounded tools, and operate with safety gates and audit-ready logs. We argue that these systems should be compared at the workflow level, rather than only by model components or end-task accuracy, because clinical action, evidence, and accountability are expressed through state transitions and handoffs. We organize the literature through a workflow-level taxonomy covering roles and handoffs, memory and evidence, tools, and reasoning, control, and escalation. We further synthesize major workflow settings and task families, introduce a four-layer evaluation stack spanning safety, process, outcome, and operations, and connect model capabilities to workflow observables relevant to deployment. Finally, we present Integration Readiness Levels (IRL1-IRL6), task-level instrumentation requirements, and recurring workflow failure modes as a practical framework for comparing, evaluating, and deploying clinical LLM agents and AI hospitals.
Thinking-Based Non-Thinking: Solving the Reward Hacking Problem in Training Hybrid Reasoning Models via Reinforcement Learning
PDF ↗Large reasoning models (LRMs) have attracted much attention due to their exceptional performance. However, their performance mainly stems from thinking, a long Chain of Thought (CoT), which significantly increase computational overhead. To address this overthinking problem, existing work focuses on using reinforcement learning (RL) to train hybrid reasoning models that automatically decide whether to engage in thinking or not based on the complexity of the query. Unfortunately, using RL will suffer the the reward hacking problem, e.g., the model engages in thinking but is judged as not doing so, resulting in incorrect rewards.To mitigate this problem, existing works either employ supervised fine-tuning (SFT), which incurs high computational costs, or enforce uniform token limits on non-thinking responses, which yields limited mitigation of the problem.In this paper, we propose Thinking-Based Non-Thinking (TNT). It does not employ SFT, and sets different maximum token usage for responses not using thinking across various queries by leveraging information from the solution component of the responses using thinking. Experiments on five mathematical benchmarks demonstrate that TNT reduces token usage by around 50\\%$ compared to DeepSeek-R1-Distill-Qwen-1.5B/7B and DeepScaleR-1.5B, while significantly improving accuracy. In fact, TNT achieves the optimal trade-off between accuracy and efficiency among all tested methods. Additionally, the probability of reward hacking problem in TNT’s responses, which are classified as not using thinking, remains below $10\\%$ across all tested datasets.
Large Language Models (LLMs) have recently advanced the field of Automated Theorem Proving (ATP), attaining substantial performance gains through widely adopted test-time scaling strategies, notably reflective Chain-of-Thought (CoT) reasoning and increased sampling passes. However, they both introduce significant computational overhead for inference. Moreover, existing cost analyses typically regulate only the number of sampling passes, while neglecting the substantial disparities in sampling costs introduced by different scaling strategies. In this paper, we systematically compare the efficiency of different test-time scaling strategies for ATP models and demonstrate the inefficiency of the current state-of-the-art (SOTA) open-source approaches. We then investigate approaches to significantly reduce token usage and sample passes while maintaining the original performance. Specifically, we propose two complementary methods that can be integrated into a unified EconRL pipeline for amplified benefits: (1) a dynamic Chain-of-Thought (CoT) switching mechanism designed to mitigate unnecessary token consumption, and (2) Diverse parallel-scaled reinforcement learning (RL) with trainable prefixes to enhance pass rates under constrained sampling passes. Experiments on miniF2F and ProofNet demonstrate that our EconProver-GD achieves comparable performance to baseline methods with only 12% of the computational cost. This work provides actionable insights for deploying lightweight ATP models without sacrificing performance.
Multi-Task Representation Alignment on Language Understanding: A Mutual Information Perspective
PDF ↗Multi-task learning (MTL) enables joint learning over multiple tasks based on shared representations, but suffers from task interference issue during optimization. Existing works mainly focus on task balancing or probabilistic modeling but fail to address the issue since they struggle to learn sufficient representations for all target tasks. To address this, we propose a multi-task representation alignment (MTRA) framework to achieve task-specific alignment and self-alignment on the shared representations from a mutual information perspective. MTRA ensures that the learned representations contain task-relevant features while mitigating the negative effects of task-irrelevant features. First, we design a task-specific alignment objective to align the shared representations and task-specific representations with the expected targets of all tasks via information maximization. Besides, we design a self-alignment objective to eliminate task-irrelevant features via conditional information minimization. Experiments on two multi-task language benchmarks show that MTRA outperforms 13 representative MTL methods under the same settings, particularly under label-noisy and data-constrained conditions. Further analysis shows that the learned shared representations exhibit sufficient task informativeness and superior alignment properties.