Test-Time Scaling (TTS) methods for enhancing Large Language Model (LLM) reasoning often incur substantial inference costs, due to reliance on long chain-of-thought (CoT) generation, self-consistency sampling methods, or searching under Process Reward Models (PRMs). This paper introduces Guided by Gut (GG), an efficient self-guided TTS framework that enables LLMs to perform step-by-step reasoning at a low cost, without any reward models or verifiers. GG performs a lightweight tree search guided solely by intrinsic confidence signals of the LLM at each reasoning step and improves the reliability of such internal confidence signals by reinforcement learning. Empirical evaluations on challenging mathematical reasoning benchmarks demonstrate that GG enables smaller models (e.g., 1.5B-7B parameters) to achieve accuracy matching or surpassing significantly larger models (e.g., 32B–70B parameters), while reducing GPU memory usage by up to 10×. Compared to TTS with PRMs, GG achieves comparable accuracy with 8× faster inference speeds and 4–5× lower memory usage. Additionally, GG reduces KV cache memory usage by approximately 50% compared to Best-of-N sampling, facilitating more efficient and practical deployment of TTS techniques.
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
Uncertainty quantification (UQ) for large language models (LLMs) is a key building block for safety guardrails of daily LLM applications. Yet, even as LLM agents are increasingly deployed in highly complex tasks, most UQ research still centers on single-turn question-answering. We argue that UQ research must shift to realistic settings with interactive agents, and that a new principled framework for agent UQ is needed. This paper presents three pillars to build a solid ground for future agent UQ research: (1. Foundations) We present the first general formulation of agent UQ that subsumes broad classes of existing UQ setups; (2. Challenges) We identify four technical challenges specifically tied to agentic setups—selection of uncertainty estimator, uncertainty of heterogeneous entities, modeling uncertainty dynamics in interactive systems, and lack of fine-grained benchmarks—with numerical analysis on a real-world agent benchmark, \tau^2-bench; (3. Future Directions) We conclude with noting on the practical implications of agent UQ and remaining open problems as forward-looking discussion for future explorations.
Every Response Counts: Quantifying Uncertainty of LLM-based Multi-Agent Systems through Tensor Decomposition
PDF ↗While Large Language Model-based Multi-Agent Systems (MAS) consistently outperform single-agent systems on complex tasks, their intricate interactions introduce critical reliability challenges arising from communication dynamics and role dependencies. Existing Uncertainty Quantification methods, typically designed for single-turn outputs, fail to address the unique complexities of the MAS. Specifically, these methods struggle with three distinct challenges: the cascading uncertainty in multi-step reasoning, the variability of inter-agent communication paths, and the diversity of communication topologies. To bridge this gap, we introduce MATU, a novel framework that quantifies uncertainty through tensor decomposition. MATU moves beyond analyzing final text outputs by representing entire reasoning trajectories as embedding matrices and organizing multiple execution runs into a higher-order tensor. By applying tensor decomposition, we disentangle and quantify distinct sources of uncertainty, offering a comprehensive reliability measure that is generalizable across different agent structures. We provide comprehensive experiments to show that MATU effectively estimates holistic and robust uncertainty across diverse tasks and communication topologies.
Accommodation and Epistemic Vigilance: A Pragmatic Account of Why LLMs Fail to Challenge Harmful Beliefs
PDF ↗Recent evaluations show that large language models (LLMs) frequently fail to challenge users’ harmful beliefs in domains ranging from medical advice to social reasoning. We present a unifying analysis through the lens of pragmatics: these safety failures can be understood and addressed as LLMs exhibiting excessive accommodation and insufficient epistemic vigilance. We show that the pragmatic factors affecting accommodation and epistemic vigilance in humans (at-issueness, linguistic encoding, and source reliability) influence LLM behaviors in similar ways. We demonstrate how these factors explain performance differences across three safety benchmarks that test models’ ability to challenge harmful beliefs, spanning misinformation (Cancer-Myth, SAGE-Eval) and sycophancy (ELEPHANT). This pragmatic lens further motivates prompting interventions, such as adding the phrase "wait a minute", that drastically improve performance on these difficult benchmarks by shifting pragmatic cues. Our results have practical implications for benchmark design and underscore the importance of pragmatics for understanding model behavior and improving performance.
Large language models (LLMs) have shown remarkable capabilities across diverse coding tasks. However, their adoption requires a true understanding of program execution rather than relying on surface-level patterns. Existing benchmarks primarily focus on predicting program properties tied to specific inputs (e.g., code coverage, program outputs). As a result, they provide a narrow view of dynamic code reasoning and are prone to data contamination. We argue that understanding program execution requires evaluating its inherent duality through two complementary reasoning tasks: (i) predicting a program’s observed behavior for a given input, and (ii) inferring how the input must be mutated toward a specific behavioral objective. Both tasks jointly probe a model’s causal understanding of execution flow. We instantiate this duality in DexBench, a benchmark comprising 445 paired instances, and evaluate 13 LLMs. Our results demonstrate that dual-path reasoning provides a robust and discriminative proxy for dynamic code understanding.
MediEval: A Unified Medical Benchmark for Patient-Contextual and Knowledge-Grounded Reasoning in LLMs
PDF ↗Large Language Models (LLMs) are increasingly applied to medicine, yet their adoption is limited by concerns over reliability and safety. Existing evaluations either test factual medical knowledge in isolation or assess patient-level reasoning without verifying correctness, leaving a critical gap. We introduce MediEval, a benchmark that links MIMIC-IV electronic health records (EHRs) to a unified knowledge base built from UMLS and other biomedical vocabularies. MediEval generates diverse factual and counterfactual medical statements within real patient contexts, enabling systematic evaluation across a 4-quadrant framework that jointly considers knowledge grounding and contextual consistency. Using this framework, we identify critical failure modes, including hallucinated support and truth inversion, that current proprietary, open-source, and domain-specific LLMs frequently exhibit. To address these risks, we propose Counterfactual Risk-Aware Fine-tuning (CoRFu), a DPO-based method with an asymmetric penalty targeting unsafe confusions. CoRFu improves by +16.4 macro-F1 points over the base model and eliminates truth inversion errors, demonstrating both higher accuracy and substantially greater safety.
Training models on a carefully chosen portion of data rather than the full dataset is now a standard preprocess for modern ML. From vision coreset selection to large-scale filtering in language models, it enables scalability with minimal utility loss. A common intuition is that training on fewer samples should also reduce privacy risks. In this paper, we challenge this assumption. We show that subset training is not privacy free: the very choices of which data are included or excluded can introduce new privacy surface and leak more sensitive information. Such information can be captured by adversaries either through side-channel metadata from the subset selection process or via the outputs of the target model. To systematically study this phenomenon, we propose CoLA (Choice Leakage Attack), a unified framework for analyzing privacy leakage in subset selection. In CoLA, depending on the adversary’s knowledge of the side-channel information, we define two practical attack scenarios: Subset-aware Side-channel Attacks and Black-box Attacks. Under both scenarios, we investigate two privacy surfaces unique to subset training: (1) Training-membership MIA (TM-MIA), which concerns only the privacy of training data membership, and (2) Selection-participation MIA (SP-MIA), which concerns the privacy of all samples that participated in the subset selection process. Notably, SP-MIA enlarges the notion of membership from model training to the entire data-model supply chain. Experiments on vision and language models show that existing threat models underestimate subset-training privacy risks: the expanded privacy surface leaks both training and selection membership, extending risks from individual models to the broader ML ecosystem.
ParaCodex: A Profiling-Guided Autonomous Coding Agent for Reliable Parallel Code Generation and Translation
PDF ↗Parallel programming is central to HPC and AI, but producing code that is correct and fast remains challenging, especially for OpenMP GPU offload, where data movement and tuning dominate. Autonomous coding agents can compile, test, and profile on target hardware, but outputs are brittle without domain scaffolding.We present ParaCodex, an HPC-engineer workflow that turns a Codex-based agent into an autonomous OpenMP GPU offload system using staged hotspot analysis, explicit data planning, correctness gating, and profiling-guided refinement. We evaluate translation from serial CPU kernels to OpenMP GPU offload kernels on HeCBench, Rodinia, and NAS. After excluding five kernels, ParaCodex succeeded on all 31 valid kernels. In 27/31 (87%) of these valid cases, the generated kernels improved GPU time over reference implementations, a result that holds independently on both the A100 and RTX 4060. The resulting OpenMP kernels achieve geometric-mean speedups of 3.1 (A100) and 3.6 (RTX 4060) on HeCBench and 1.5 and 1.1 on Rodinia, and outperform a zero-shot Codex baseline on all suites. We also evaluate CUDA -> OpenMP offload translation on ParEval, where ParaCodex maintains high compilation and validation rates in code-only and end-to-end settings.ParaCodex is available at https://github.com/Scientific-Computing-Lab/ParaCodex
The increasing use of large language models in mental health applications calls for principled evaluation frameworks that assess alignment with psychotherapeutic best practices beyond surface-level fluency. While recent systems exhibit conversational competence, they lack structured mechanisms to evaluate adherence to core therapeutic principles. In this paper, we study the problem of evaluating AI-generated therapist-like responses for clinically grounded appropriateness and effectiveness. We assess each therapists utterance along six therapeutic principles: non-judgmental acceptance, warmth, respect for autonomy, active listening, reflective understanding, and situational appropriateness using a fine-grained ordinal scale. We introduce FAITH-M, a benchmark annotated with expert-assigned ordinal ratings, and propose CARE, a multi-stage evaluation framework that integrates intra-dialogue context, contrastive exemplar retrieval, and knowledge-distilled chain-of-thought reasoning. Experiments show that CARE achieves an F-1 score of 63.34 versus the strong baseline Qwen3 F-1 score of 38.56 which is a 64.26% improvement, which also serves as its backbone, indicating that gains arise from structured reasoning and contextual modeling rather than backbone capacity alone. Expert assessment and external dataset evaluations further demonstrate robustness under domain shift, while highlighting challenges in modeling implicit clinical nuance. Overall, CARE provides a clinically grounded framework for evaluating therapeutic fidelity in AI mental health systems.
Severe acoustic degradation is often caused by overlapping noise, disfluencies, and environmental distortions. This phenomenon results in the dissolution of linguistic structures and the generation of unreliable ASR outputs. Inspired by human speech comprehension, we propose Speech-MLM, a novel multimodal framework that reframes ASR as semantics-guided speech reconstruction. This perspective introduces three core challenges: (C1) collapse of linguistic structure under acoustic degradation, (C2) semantic ambiguity under noise, and (C3) misalignment across modalities. To address these issues, we propose Speech-MLM, a multimodal ASR framework that integrates speech, spectrogram-derived visual cues, and textual variants to enhance robustness. It consists of: (i) Cognitive Structure Extractor that recovers prosodic structure from visualized acoustic features, (ii) Semantic Weaver that learns semantic equivalence across varied textual forms, and (iii) Retrieval-Guided Fusion Learner that unifies modalities within a shared semantic space. Experiments on multiple real-world noisy datasets demonstrate that Speech-MLM achieves an average 38.85% reduction in WER, while also attaining 98.71% BERTScore and 96.7% USE, over advanced baselines, demonstrating substantial gains in semantic robustness and generalization across domains.
Large language models (LLMs) are increasingly deployed in multilingual, real-world applications with user inputs – naturally introducing typographical errors (typos). Yet most benchmarks assume clean input, leaving the robustness of LLMs to typos across languages largely underexplored. To address this gap, we introduce MulTypo, a multilingual typo generation algorithm that simulates human-like errors based on language-specific keyboard layouts and typing behavior. We evaluate 18 open-source LLMs across three model families and five downstream tasks spanning language inference, multi-choice question answering, mathematical reasoning, and machine translation tasks. Our results show that typos consistently degrade performance, particularly in generative tasks and those requiring reasoning – while the natural language inference task is comparatively more robust. Instruction tuning improves clean-input performance but may increase brittleness under noise. We also observe language-dependent robustness: high-resource languages are generally more robust than low-resource ones, and translation from English is more robust than translation into English. Our findings underscore the need for noise-aware training and multilingual robustness evaluation. We release a Python package for MulTypo and make the source code publicly available at https://github.com/cisnlp/multypo.
What statistical properties might support learning abstract grammatical knowledge from linear input? We address this question by examining the statistical distribution of function words. Function words have been argued to aid acquisition through three distributional properties: high frequency, reliable syntactic association, and phrase-boundary alignment. We conduct a cross-linguistic corpus analysis of 186 languages, which confirms that all three properties are universal. Using counterfactual language modeling and ablation experiments on English, we show that preserving these properties facilitates acquisition in neural learners, with a Goldilocks effect: function words must be frequent enough to be reliable, yet diverse enough to remain informative to structural dependency. Probing analyses further reveal that different learning conditions produce systematically different reliance on function words.
Graph-based retrieval-augmented generation (GraphRAG) systems construct knowledge graphs over document collections to support multi-hop reasoning. While prior work shows that GraphRAG responses may leak retrieved subgraphs, the feasibility of *query-efficient* reconstruction of the hidden graph structure remains unexplored under realistic query budgets. We study a budget-constrained black-box setting where an adversary adaptively queries the system to steal its latent entity–relation graph. We propose AGEA (Agentic Graph Extraction Attack), a framework that leverages a novelty-guided exploration–exploitation strategy, external graph memory modules, and a two-stage graph extraction pipeline combining lightweight discovery with LLM-based filtering. We evaluate AGEA on medical, agriculture, and literary datasets across Microsoft-GraphRAG and LightRAG systems. Under identical query budgets, AGEA significantly outperforms prior attack baselines, recovering up to 90% of entities and relationships while maintaining high precision. These results demonstrate that modern GraphRAG systems are highly vulnerable to structured, agentic extraction attacks, even under strict query limits. The code is available at https://github.com/shuashua0608/AGEA.
Not All Tokens Matter: Towards Efficient LLM Reasoning via Token Significance in Reinforcement Learning
PDF ↗Large language models (LLMs) show strong reasoning abilities but often produce unnecessarily long explanations that reduce efficiency. Although reinforcement learning (RL) has been used to improve reasoning, most methods focus on accuracy and rely on uniform length-based rewards that overlook the differing contributions of individual tokens, often harming correctness. We revisit length optimization in RL through the perspective of token significance. Observing that many chain-of-thought (CoT) tokens contribute little to the final answer, we introduce a significance-aware length reward that selectively penalizes insignificance tokens, reducing redundancy while preserving essential reasoning. We also propose a dynamic length reward that encourages more detailed reasoning early in training and gradually shifts toward conciseness as learning progresses. Integrating these components into standard policy optimization yields a framework that improves both reasoning efficiency and accuracy. Experiments across multiple benchmarks demonstrate substantial reductions in response length while preserving or improving correctness, highlighting the importance of modeling token significance for efficient LLM reasoning.
Language models (LMs) are increasingly used in collaboration: multiple LMs trained by different parties collaborate through routing systems, multi-agent debate, model merging, and more. Critical safety risks remain in this decentralized paradigm: what if some of the models in multi-LLM systems are compromised or malicious? We first quantify the impact of malicious models by engineering four categories of malicious LMs, plug them into four types of popular model collaboration systems, and evaluate the compromised system across 10 datasets. We find that malicious models have a severe impact on the multi-LLM systems, especially for reasoning and safety domains where performance is lowered by 7.12% and 7.94% on average. We then propose mitigation strategies to alleviate the impact of malicious components, by employing external supervisors that oversee model collaboration to disable/mask them out to reduce their influence. On average, these strategies recover 95.31% of the initial performance, while making model collaboration systems fully resistant to malicious models remains an open research question.
Question answering over visually rich documents (VRDs) requires reasoning not only over isolated content but also over documents’ structural organization and cross-page dependencies. However, conventional retrieval-augmented generation (RAG) methods encode content in isolated chunks during ingestion, losing structural and cross-page dependencies, and retrieve a fixed number of pages at inference, regardless of the specific demands of the question or context. This often results in incomplete evidence retrieval and degraded answer quality for multi-page reasoning tasks. To address these limitations, we propose LAD-RAG, a novel Layout-Aware Dynamic RAG framework. During ingestion, LAD-RAG constructs a symbolic document graph that captures layout structure and cross-page dependencies, adding it alongside standard neural embeddings to yield a more holistic representation of the document. During inference, an LLM agent dynamically interacts with the neural and symbolic indices to adaptively retrieve the necessary evidence based on the query. Experiments on MMLongBench-Doc, LongDocURL, DUDE, and MP-DocVQA demonstrate that LAD-RAG improves retrieval, achieving over 90% perfect recall on average without any top-k tuning, and outperforming baseline retrievers by up to 20% in recall at comparable noise levels, yielding higher QA accuracy with minimal latency.
Language Models Don’t Know What You Want: Evaluating Personalization in Deep Research Needs Real Users
PDF ↗Deep Research (DR) tools (e.g. OpenAI DR) help researchers cope with ballooning publishing counts. Such tools can synthesize scientific papers to answer researchers’ queries, but lack understanding of their users. We change that in MyScholarQA (MySQA), a personalized DR tool that: 1) infers a profile of a user’s research interests; 2) proposes personalized actions for a user’s input query; and 3) writes a multi-section report for the query that follows user-approved actions. We first test MySQA with NLP’s standard protocol: we design a benchmark of synthetic users and LLM judges, where MySQA beats baselines in citation metrics and personalized action-following. However, we suspect this process does not cover all aspects of personalized DR users value, so we interview users in an online version of MySQA to unmask them. We reveal nine nuanced errors of personalized DR undetectable by our LLM judges, and we study qualitative feedback to form lessons for future DR design. In all, we argue for a pillar of personalization that easy-to-use LLM judges can lead NLP to overlook: real progress in personalization is only possible with real users.
Effective and efficient access to relevant information is essential for disaster management. However, no retrieval model is specialized for disaster management, and existing general-domain models fail to handle the varied search intents inherent to disaster management scenarios, resulting in inconsistent and unreliable performance. To this end, we introduce DMRetriever, the first series of dense retrieval models (33M to 7.6B) tailored for this domain. It is trained through a novel three-stage framework of bidirectional attention adaptation, unsupervised contrastive pre-training, and difficulty-aware progressive instruction fine-tuning, using high-quality data generated through an advanced data refinement pipeline. Comprehensive experiments demonstrate that DMRetriever achieves state-of-the-art performance across all six search intents at every model scale. Moreover, DMRetriever is highly parameter-efficient, with 596M model outperforming baselines over 13.3 larger and 33M model exceeding baselines with only 7.6% of their parameters. All codes, data, and checkpoints are available at https://github.com/KaiYin97/DMRETRIEVER.
Hallucinations in machine translation (MT)—outputs that may be fluent yet unfaithful to the source content—remain a critical obstacle. They hinder the reliable deployment of MT systems in real-world applications. Despite growing attention to this phenomenon, progress has been constrained by the lack of large-scale, high-quality benchmarks dedicated to hallucination detection. We introduce HAT (Hallucination Annotation for Translation), a novel dataset designed to advance research on this problem. HAT comprises 350,959 span-level annotated samples across 38 language pairs, with approximately 8,000–10,000 samples per pair partitioned into training, development, and test sets. Annotations were produced by professional translators under rigorous quality control protocols to ensure reliability. We provide a detailed analysis of hallucination distributions and establish benchmark performance using a diverse set of baselines, including automatic MT evaluation metrics as well as large language models. By providing the first large-scale, systematically annotated resource for hallucination detection in MT, HAT enables the development of more faithful translation models and lays the groundwork for future research on building trustworthy machine translation systems.
Model Internal Sleuthing: Finding Lexical Identity and Inflectional Features in Modern Language Models
PDF ↗Large transformer-based language models dominate modern NLP, yet our understanding of how they encode linguistic information relies primarily on studies of early models like BERT and GPT-2. We systematically probe 25 models from BERT Base to Qwen2.5-7B focusing on two linguistic properties: lexical identity and inflectional features across 6 diverse languages. We find a consistent pattern: inflectional features are linearly decodable throughout the model, while lexical identity is prominent early but increasingly weakens with depth. Further analysis of the representation geometry reveals that models with aggressive mid-layer dimensionality compression show reduced steering effectiveness in those layers, despite probe accuracy remaining high. Pretraining analysis shows that inflectional structure stabilizes early while lexical identity representations continue evolving. Taken together, our findings suggest that transformers maintain inflectional features across layers, while trading off lexical identity for compact, predictive representations.