论文检索

Shuyao Xiao, Shengling Wang, Ke Chao

Large Language Models (LLMs) are widely used in high-stakes applications, making their interpretability increasingly important. Existing interpretability methods are typically categorized into internal and external perspectives, which are often studied in isolation and tend to overlook two key aspects: causality and temporal dynamics. Explanations are often limited to surface correlations or static dependencies, failing to capture how influences evolve during autoregressive generation. To address these limitations, we propose a causal and dynamic interpretability framework for LLM generation. We first characterize the backdoor-adjusted causal effects of both the generated prefix and the prompt on the current token using the Structural Causal Model. Next, we introduce two metrics to quantify contextual causal influence and question–answer causal influence. Overall, our work provides a unified causal view of internal consistency and external alignment in LLM generation dynamics.

Mingze Xu, Zijing Zhao, Qiming Peng, Houwen Peng, Han Hu, Zhanhui Kang, Yuxing Han

While Multimodal Large Language Models (MLLMs) are advancing rapidly, accurately evaluating their capabilities remains challenging. Current paradigms primarily rely on holistic scoring and static leaderboards, which fail to disentangle fine-grained competencies. Specifically, they suffer from “Outcome Bias” by validating only final answers and ignoring intermediate reasoning. To address these limitations, we introduce ATOM (AnaTomy Of MLLM), a novel MLLM-as-a-judge framework designed to shift the focus from ranking to fine-grained diagnosis. ATOM decomposes complex reasoning into atomic criteria anchored in visual elements, enforcing verification against explicit visual facts. Validated on a newly constructed benchmark with rigorous human rankings, ATOM achieves state-of-the-art accuracy, surpassing the strongest baseline by up to 7.92%. Moving beyond ranking, ATOM bridges the gap between assessment and alignment: by pinpointing atomic-level failures, it establishes a closed-loop mechanism for targeted self-correction. This approach enables models to identify and rectify errors autonomously, successfully resolving up to 39.95% of previously failed queries without human intervention.

Zhizhang Fu, Yuancheng Gu, Chenkai Hu, Hanmeng Liu, Yue Zhang

Test-time scaling has enabled Large Language Models (LLMs) to tackle complex reasoning, yet the limitations of current Chain-of-Thought (CoT) evaluation obscure whether performance gains stem from genuine reasoning or mere verbosity. To address this, (1) we propose a novel neuro-symbolic framework for the non-intrusive, comprehensive process-centric evaluation of reasoning grounded in First-Order Logic. (2) Through this lens, we identify four distinct behavioral prototypes and diagnose the failure modes. (3) We examine the impact of inference mode, training strategy, and model scale. Our analysis reveals that extended token generation is not a prerequisite for deep reasoning. Furthermore, we reveal critical constraints: mixing long and short CoT data in training risks premature saturation and collapse, while distillation into smaller models captures behavioral length but fails to replicate logical efficacy due to intrinsic capacity limits.

Sourajit Saha, Tejas Gokhale

In multimodal information retrieval (MMIR), candidates relevant to an input query need to be retrieved from a database, where the query and database items span different modalities. As real-world databases evolve, repeatedly annotating and indexing data and re-optimizing domain-specific models across modalities is impractical. We present MULTI-SCORE, a fine-tuning-free, two-stage MMIR approach that couples efficient candidate filtering with fine-grained multimodal re-ranking. Stage-1 adopts Matryoshka representations to efficiently filter out low-relevance candidates without expensive similarity computations on full-scale representations for the entire database. Stage-2 re-ranks the filtered candidates by computing their fine-grained multimodal contextual representations with two scoring functions for semantic alignment using chain-of-thought prompting and question-answering. Experiments demonstrate state-of-the-art zero-shot retrieval on 12 MMIR tasks across 32 datasets while outperforming supervised methods on 23 datasets.

Zikang Liu, Peilan Xu

Generative answer engines expose content through selective citation rather than ranked retrieval, fundamentally altering how visibility is determined. This shift calls for new optimization methods beyond traditional search engine optimization. Existing generative engine optimization (GEO) approaches primarily rely on token-level text rewriting, offering limited interpretability and weak control over the trade-off between citation visibility and content quality. We propose FeatGEO, a feature-level, multi-objective optimization framework that abstracts webpages into interpretable structural, content, and linguistic properties. Instead of directly editing text, FeatGEO optimizes over this feature space and uses a language model to realize feature configurations into natural language, decoupling high-level optimization from surface-level generation. Experiments on GEO-Bench across three generative engines demonstrate that FeatGEO consistently improves citation visibility while maintaining or improving content quality, substantially outperforming token-level baselines. Further analyses show that citation behavior is more strongly influenced by document-level content properties than by isolated lexical edits, and that the learned feature configurations generalize across language models of different scales.

Shichen Li, Zhouyang Wang, Zhongqing Wang, Peifeng Li

Large Language Models (LLMs) can perform sentiment analysis via natural language instructions, yet their predictions are highly sensitive to prompt phrasing. Prior work has shown that sentiment is encoded linearly in LLM representations, but the model’s ability to utilize this information remains surprisingly fragile to prompt variations. We leverage Sparse Autoencoders (SAEs) and circuit-level analysis to uncover causal mechanisms underlying sentiment prediction. We identify a sentiment analysis circuit and find that prompt sensitivity may stem from task activation failure. The model encodes the sentiment feature consistently, but different prompts trigger varying degrees of circuit activation. Based on this insight, we propose a simple inference-time intervention method that amplifies circuit features to compensate for insufficient activation. Experiments across diverse datasets, templates, and languages show consistent improvements, offering an interpretable and training-free alternative to manual prompt engineering.

Ai Jian, Jingqing Ruan, Xing Ma, Dailin Li, Weipeng Zhang, Ke Zeng, Xunliang Cai

Reward models (RMs) are central to reinforcement learning from human feedback (RLHF), providing the critical supervision signals that align large language models (LLMs) with human preferences.Generative reward models (GRMs) provide greater interpretability than traditional scalar RMs, but they come with a critical trade-off: pairwise methods are hindered by a training-inference mismatch, while pointwise methods require expensive absolute annotations.To bridge this gap, we propose the Preference-aware Task-adaptive Reward Model (PaTaRM).Unlike prior approaches, PaTaRM enables robust pointwise training using readily available pairwise data via a novel Preference-Aware Reward (PAR) mechanism, eliminating the need for explicit rating labels. Furthermore, it incorporates a task-adaptive rubric system that dynamically generates instance-specific criteria for precise evaluation.Extensive experiments demonstrate that PaTaRM achieves an average relative improvement of 8.7% over the corresponding base models on RewardBench and RMBench across the Qwen3-8B and Qwen3-14B backbones.Crucially, when used as a reward model for downstream RLHF, it yields an average relative improvement of 13.6% over the corresponding base policies on IFEval and InfoBench, validating its effectiveness for policy alignment.Our code, data, and checkpoints are available at https://huggingface.co/AIJian/PaTaRM

Mohammadtaher Safarzadeh, Hitesh Laxmichand Patel, Afshin Oroojlooy, Graham Horwood, Dan Roth

Large language models (LLMs) have achieved strong performance on natural language to SQL (NL2SQL) benchmarks, yet their reported accuracy may be inflated by contamination from benchmark queries or structurally similar patterns seen during training. We introduce SPENCE (Syntactic Probing and Evaluation of NL2SQL Contamination Effects), a controlled syntactic probing framework for detecting and quantifying such contamination. SPENCE systematically generates syntactic variants of test queries for four widely used NL2SQL datasets—Spider, SParC, CoSQL, and the newer BIRD benchmark. We use SPENCE to evaluate multiple high-capacity LLMs under execution-based scoring. For each model, we measure changes in execution accuracy (Delta Accuracy) across increasing levels of syntactic divergence and quantify rank sensitivity using Kendall’s tau with bootstrap confidence intervals. By aligning these robustness trends with benchmark release dates, we observe a clear temporal gradient: older benchmarks such as Spider exhibit the strongest negative values and thus the highest likelihood of training leakage, whereas the more recent BIRD dataset shows minimal sensitivity and appears largely uncontaminated. Together, these findings highlight the importance of temporally contextualized, syntactic-probing evaluation for trustworthy NL2SQL benchmarking.

Zihan Chen, Yiming Zhang, Wenxiang Geng, Zenghui Ding, Yining Sun

Large Language Models (LLMs) aligned via outcome-based Reinforcement Learning (RL) frequently exhibit a critical failure mode: they achieve high performance on in-distribution benchmarks while demonstrating brittle reasoning capabilities on out-of-distribution (OOD) tasks. We term this phenomenon Reward-Induced Manifold Collapse. We establish a theoretical framework bridging Structural Causal Models (SCM) and the Information Bottleneck (IB) principle to explain this paradox. We define reasoning as a high-complexity causal process and shortcut learning as the exploitation of low-complexity spurious correlations. Under the implicit inductive bias of Stochastic Gradient Descent (SGD), models optimized for outcome rewards are biased toward shortcut solutions whenever the training distribution allows for a “Markovian Screening” of the true causal mechanism. We derive a new generalization bound based on Semantic Coverage Measure (\eta) rather than sample size, showing why data scaling on homogeneous distributions may fail to correct reasoning flaws. We also show that Process Reward Models (PRMs) function as Topological Filters, enforcing step-wise mutual information constraints that render the low-complexity shortcut manifold inadmissible. These results provide a mathematical grounding for the role of process supervision beyond simple credit assignment.

Yue Fan, Hu Zhang, Yunxiao Zhao, Guangjun Zhang, Hao Zhan, Ru Li, Hongye Tan, Yuanlong Wang

Logical reasoning with large language models (LLMs) has made significant progress in recent years. However, existing methods still suffer from insufficient rule semantic grounding and weak rule application mechanisms, making it difficult to achieve precise understanding and effective utilization of rules in complex multi-step reasoning. To address this, we propose Leibniz, a theory-of-mind driven neuro-symbolic reasoning framework. Specifically, we construct a bidirectional reasoning model based on multi-agent collaboration, which characterizes the reasoning process from two complementary perspectives, namely the Evolution Agent and the Reduction Agent. The former transforms belief-unstable propositions into stable ones that are beneficial for proving the target conclusion. The latter performs reverse reduction from the target to resolve belief conflicts and distill new inferential insights. Both share a belief state space and achieve efficient collaborative reasoning through continual belief updating. We integrate natural language and symbolic representations throughout the reasoning process. The experimental results demonstrate that Leibniz surpasses existing state-of-the-art models in reasoning accuracy, and further analyses reveal its substantial advantages in reliability and flexibility.

Hanjun Cho, Jay-Yoon Lee

Existing QA benchmarks typically assume distinct documents with minimal overlap, yet real-world retrieval-augmented generation (RAG) systems operate on corpora such as financial reports, legal codes, and patents, where information is highly redundant and documents exhibit strong inter-document similarity. This mismatch undermines evaluation validity: retrievers can be unfairly undervalued even when they retrieve documents that provide sufficient evidence, because redundancy across documents is not accounted for in evaluation. On the other hand, retrievers that perform well on standard benchmarks often generalize poorly to real-world corpora with highly similar and redundant documents. We present RARE (Redundancy-Aware Retrieval Evaluation), a framework for constructing realistic benchmarks by (i) decomposing documents into atomic facts to enable precise redundancy tracking and (ii) enhancing LLM-based data generation with CRRF. RAG benchmark data usually requires multiple quality criteria, but LLMs often yield trivial outputs. CRRF scores criteria separately and fuses decisions by rank, improving the reliability of generated data. Applying RARE to Finance, Legal, and Patent corpora, we introduce RedQA, where a strong retriever baseline drops from 66.4% PerfRecall@10 on 4-hop General-Wiki to 5.0–27.9% PerfRecall@10 at 4-hop depth, revealing robustness gaps that current benchmarks fail to capture. RARE enables practitioners to build domain-specific RAG evaluations that faithfully reflect real-world deployment conditions.

Qianhong Guo, Wei Xie, Xiaofang Cai, Enze Wang, Shuoyoucheng Ma, Xiaobing Sun, Tian Xia, Kai Chen, Xiaofeng Wang, Baosheng Wang

Although large language models (LLMs) have shown exceptional capabilities across a wide range of tasks, reliable evaluation remains a critical challenge due to data contamination, opaque operation, and subjective preferences. To address these issues, we propose League of LLMs (LOL), a novel benchmark-free evaluation paradigm that organizes multiple LLMs into a self-governed league for multi-round mutual evaluation. LOL integrates four core criteria (dynamic, transparent, objective, and professional) to mitigate key limitations of existing paradigms. Experiments on eight mainstream LLMs in mathematics and programming demonstrate that LOL can effectively distinguish LLM capabilities while maintaining high internal ranking stability (Top-k consistency = 70.7\%). Beyond ranking, LOL reveals empirical findings that are difficult for traditional paradigms to capture. For instance, “memorization-based answering” behaviors are observed in some models, and higher in-family scores are found in the OpenAI model family (\Delta = 9, p < 0.05). Finally, we make our framework and code publicly available as a valuable complement to the current LLM evaluation ecosystem.

Zhen Yang, Mingyang Zhang, Feng Chen, Ganggui Ding, Liang Hou, Xin Tao, Ying-Cong Chen

Recent progress in large language models (LLMs) has focused on test-time scaling to improve reasoning via increased inference computation, but often at the cost of efficiency. We revisit test-time behavior and uncover a simple yet underexplored phenomenon: reasoning uncertainty is highly localized—only a small subset of high-entropy tokens dominantly affects output correctness. Motivated by this, we propose Minimal Test-Time Intervention (MTI), a training-free framework that enhances reasoning accuracy and stability with minimal overhead. MTI includes: (i) Selective CFG intervention, applying classifier-free guidance only at uncertain positions; and (ii) Lightweight negative-prompt guidance, reusing the main model’s KV cache to approximate unconditional decoding efficiently. MTI yields consistent gains across general, coding, and STEM tasks—e.g., +9.28% average improvement on six benchmarks for DeepSeek-R1-7B and +11.25% on AIME2024 using Ling-mini-2.0—while remaining highly efficient.

Liang Lin, Miao Yu, Moayad Aloqaily, Zhenhong Zhou, Kun Wang, Linsey Pang, Prakhar Mehrotra, Qingsong Wen

Backdoor attacks are a significant threat to large language models (LLMs), often embedded via public checkpoints, yet existing defenses rely on impractical assumptions about trigger settings. To address this challenge, we propose Locphylax, a defense framework that requires no prior knowledge of trigger settings. Locphylax is based on the key observation that when deliberately injecting known backdoors into an already-compromised model, both existing unknown and newly injected backdoors aggregate in the representation space. Locphylax leverages this through a two-stage process: first, aggregating backdoor representations by injecting known triggers, and then, performing recovery fine-tuning to restore benign outputs. Extensive experiments across multiple LLM architectures demonstrate that: (I) Locphylax reduces the average Attack Success Rate to 4.41% across multiple benchmarks, outperforming existing baselines by 28.1%–69.3%. (II) Clean accuracy and utility are preserved within 0.5% of the original model, ensuring negligible impact on legitimate tasks. (III) The defense generalizes across different types of backdoors, confirming its robustness in practical deployment scenarios. Our code is available at https://anonymous.4open.science/r/Locphylax.

Hanwen Shen, Ting Ying, Jiajie Lu, Shanshan Wang

Although debiased large language models (LLMs) excel at handling known or low-bias prompts, they often fail on unfamiliar and high-bias prompts. We demonstrate via out-of-distribution (OOD) detection that these high-bias prompts cause a distribution shift, degrading static model performance. To enable real-time correction, we propose CAP-TTA, a test-time adaptation framework. CAP-TTA triggers context-aware LoRA updates only when a bias-risk score exceeds a set threshold. By utilizing an offline precomputed diagonal preconditioner, it ensures fast and stable optimization. Across multiple benchmarks and human evaluations, CAP-TTA effectively reduces toxicity/bias score with significantly lower latency than standard optimization methods (e.g., AdamW or SGD). Furthermore, it prevents catastrophic forgetting, and substantially improves narrative fluency over state-of-the-art baselines without compromising debiasing performance.

Shuxin Liu, Jianhao Zhang

LLMs have static pre-trained knowledge, leading to obsolescence and hallucinations. Knowledge Editing (KE) addresses these issues and typically requires multi-layer modifications. However, existing state-of-the-art methods, largely following the Locate–Select–Assign–Edit (LSAE) paradigm, rely on fixed-layer selection and uniform residual assignment, ignoring the heterogeneous causal efficacy of different layers. To bridge this, we propose \textbf{CAKE} (\textbf{C}ausal-Guided \textbf{A}daptive \textbf{K}nowledge \textbf{E}diting), a collaborative editing method within the more general Locate–Weight–Assign–Edit (LWAE) paradigm that: (1) selectively identifies critical layers via causal tracing scores; and (2) adaptively allocates editing burdens based on causal weights rather than uniform assumptions. We formulate residual assignment as a constrained quadratic optimization problem and derive a solution for optimal residual allocation, showing that aligning edits with causal efficacy mitigates recursive error accumulation. Furthermore, we establish a generalized weight shift error bound, under which existing paradigms emerge as special, restricted cases. Experimental results demonstrate that CAKE achieves SOTA performance with comparable overhead, validating the superiority of causal-guided adaptation.

Heming Xia, Cunxiao Du, Rui Li, Chak Tou Leong, Yongqi Li, Wenjie Li

Large reasoning models (LRMs) have demonstrated remarkable proficiency in tackling complex tasks through step-by-step thinking. However, this lengthy reasoning process incurs substantial computational and latency overheads, hindering the practical deployment of LRMs. This work presents a new approach to mitigating overthinking in LRMs via black-box persuasive prompting. By treating LRMs as black-box communicators, we investigate how to persuade them to generate concise responses without compromising accuracy. We introduce Whisper, an iterative refinement framework that generates high-quality persuasive prompts from diverse perspectives. Experiments across multiple benchmarks demonstrate that Whisper consistently reduces token usage while preserving performance. Notably, Whisper achieves a 3\times reduction in average response length on simple GSM8K questions for the Qwen3 series and delivers an average \sim40% token reduction overall. For closed-source APIs, Whisper reduces token usage on MATH-500 by 46% for Claude-3.7 and 50% for Gemini-2.5. Further analysis reveals the broad applicability of Whisper across data domains, model scales, and families, underscoring the potential of black-box persuasive prompting as a practical strategy for enhancing LRM efficiency.

Angqing Jiang, Jianlyu Chen, Zhefang, Yongcan Wang, Xinpeng Li, Keyu Ding, Defu Lian

Effective medical text retrieval requires both high accuracy and low latency. While LLM-based embedding models possess powerful retrieval capabilities, their prohibitive latency and high computational cost limit their application in real-time scenarios. Furthermore, the lack of comprehensive and high-fidelity benchmarks hinders progress in Chinese medical text retrieval. In this work, we introduce the **C**hinese **Med**ical **T**ext **E**mbedding **B**enchmark (**CMedTEB**), a benchmark spanning three kinds of practical embedding tasks: retrieval, reranking, and semantic textual similarity (STS). Distinct from purely automated datasets, CMedTEB is curated via a rigorous multi-LLM voting pipeline validated by clinical experts, ensuring gold-standard label quality while effectively mitigating annotation noise. On this foundation, we propose the **C**hinese Medical **A**symmetric **RE**triever (**CARE**), an asymmetric architecture that pairs a lightweight BERT-style encoder for online query encoding with a powerful LLM-based encoder for offline document encoding. However, optimizing such an asymmetric retriever with two structurally different encoders presents distinctive challenges. To address this, we introduce a novel two-stage training strategy that progressively bridges the query and document representations. Extensive experiments demonstrate that CARE surpasses state-of-the-art symmetric models on CMedTEB, achieving superior retrieval performance without increasing inference latency.

Shuyang Hou, Yi Hu, Muhan Zhang

Recent advancements in large language models (LLMs) have significantly enhanced their reasoning capabilities. However, they continue to struggle with basic character-level tasks, such as counting letters in words—a problem rooted in their tokenization process. While existing benchmarks have highlighted this weakness through basic character operations, such failures are often dismissed due to lacking practical relevance. Yet, many real-world applications, such as navigating text-based maps or interpreting structured tables, rely heavily on precise sub-token understanding. In this regard, we introduce SubTokenTest, a comprehensive benchmark that assesses sub-token understanding through **practical, utility-driven** tasks. Our benchmark includes ten tasks across four domains and isolates tokenization-related failures by decoupling performance from complex reasoning. We provide a comprehensive evaluation of nine advanced LLMs. Additionally, we investigate the impact of test-time scaling on sub-token reasoning and explore how character-level information is encoded within the hidden states.

Shanghao Li, Jinda Han, Yibo Wang, Yuanjie Zhu, Zihe Song, Langzhou He, Kenan Kamel A Alghythee, Philip S. Yu

In many reasoning tasks, large language models (LLMs) rely on structured external knowledge, such as graphs and tables, which is typically linearized into sequential token representations. However, even when sufficient knowledge is available, LLMs can still produce hallucinated outputs, and the underlying mechanisms behind such failures remain poorly understood. We investigate these mechanisms and find that hallucinations arise from systematic internal dynamics rather than random noise. First, attention disproportionately concentrates toward shortcut-like structural cues rather than distributing across the full context. Second, feed-forward representations fail to ground the provided knowledge, causing the model to revert to parametric memory. Moreover, our results indicate that hallucination is consistently associated with failures in semantic grounding within feed-forward layers, while attention allocation exhibits greater task-dependent variability. Finally, we show that these mechanistic patterns generalize beyond single-hop graphs to multi-hop and tabular settings, enabling effective hallucination detection across structured knowledge formats.