论文检索

Xinliang Frederick Zhang, Anhad Mohananey, Alexandra Chronopoulou, Pinelopi Papalampidi, Somit Gupta, Tsendsuren Munkhdalai, Lu Wang, Shyam Upadhyay

Models employing long chain-of-thought (CoT) reasoning have shown superior performance on complex reasoning tasks. Yet, this capability introduces a critical and often overlooked inefficiency—overthinking—models often engage in unnecessarily extensive reasoning even for simple queries, incurring significant computations without accuracy improvements. While prior work has explored solutions to mitigate overthinking, a fundamental gap remains in our understanding of its underlying causes. Most existing analyses are limited to superficial, profiling-based observations, failing to delve into LLMs’ inner workings. This study introduces a systematic, fine-grained analyzer of LLMs’ thought process to bridge the gap, TRACE. We first benchmark the overthinking issue, confirming that long-thinking models are five to twenty times slower on simple tasks with no substantial gains. We then use TRACE to first decompose the thought process into minimally complete sub-thoughts. Next, by inferring discourse relationships among sub-thoughts, we construct granular thought progression graphs and subsequently identify common thinking patterns for topically similar queries. Our analysis reveals two major patterns for open-weight thinking models—Explorer and Late Landing. This finding provides evidence that over-verification and over-exploration are the primary drivers of overthinking in LLMs. Grounded in thought structures, we propose a utility-based definition of overthinking, which moves beyond length-based metrics. This revised definition offers a more insightful understanding of LLMs’ thought progression, as well as practical guidelines for principled overthinking management.

Yusuke Yamauchi, Akiko Aizawa

Psychological research has long utilized circumplex models to structure emotions, placing similar emotions adjacently and opposing ones diagonally. Although frequently used to interpret deep learning representations, these models are rarely directly incorporated into the representation learning of language models, leaving their geometric validity unexplored. This paper proposes a method to induce circular emotion representations within language model embeddings via contrastive learning on a hypersphere. We show that while this circular alignment offers superior interpretability and robustness against dimensionality reduction, it underperforms compared to conventional designs in high-dimensional settings and fine-grained classification. Our findings elucidate the trade-offs involved in applying psychological circumplex models to deep learning architectures.

Yuwei Zhang, Wenhao Yu, Shangbin Feng, Yifan Zhu, Letian Peng, Jayanth Srinivasa, Gaowen Liu, Jingbo Shang

Despite significant advances in large language models (LLMs), their knowledge memorization capabilities remain underexplored, due to the lack of standardized and high-quality testing grounds. In this paper, we introduce a novel, real-world and large-scale knowledge injection benchmark that continuously evolves without human intervention. Specifically, we propose WikiDYK, which leverages recently-added and expert-curated facts from Wikipedia’s “Did You Know...” entries. Each entry is converted into multiple question–answer pairs spanning diverse task formats from easy cloze prompts to complex multi-hop questions. WikiDYK currently contains 12,290 facts and 77,180 questions, and its design allows for seamless extension with future updates from Wikipedia editors. Through extensive experiments using continued pre-training, we reveal a surprising insight: despite their prevalence in modern LLMs, Causal Language Models (CLMs) demonstrate significantly weaker knowledge memorization capabilities compared to Bidirectional Language Models (BiLMs), exhibiting a 23% lower accuracy in terms of reliability. To compensate for the smaller scales of current BiLMs, we introduce a modular collaborative framework utilizing ensembles of BiLMs as external knowledge repositories to integrate with LLMs. Experiment shows that this framework further improves the reliability accuracy by up to 29.1%. Code: https://github.com/zhang-yu-wei/WikiDYK.

Xueqing Peng, Lingfei Qian, Yan Wang, Ruoyu Xiang, Yueru He, Yang Ren, Mingyang Jiang, Vincent Jim Zhang, Yuqing Guo, Jeff Zhao 等

Real-world financial analysis involves information across multiple languages and modalities, from reports and news to scanned filings and meeting recordings. Yet most existing evaluations of LLMs in finance remain text-only, monolingual, and largely saturated by current models. To bridge these gaps, we present MultiFinBen, the first expert-annotated multilingual (five languages) and multimodal (text, vision, audio) benchmark for evaluating LLMs in realistic financial contexts. MultiFinBen introduces two new task families: multilingual financial reasoning, which tests cross-lingual evidence integration from filings and news, and financial OCR, which extracts structured text from scanned documents containing tables and charts. Rather than aggregating all available datasets, we apply a structured, difficulty-aware selection based on advanced model performance, ensuring balanced challenge and removing redundant tasks. Evaluating 21 leading LLMs shows that even frontier multimodal models like GPT-4o achieve only 46.01% overall, stronger on vision and audio but dropping sharply in multilingual settings. These findings expose persistent limitations in multilingual, multimodal, and expert-level financial reasoning. All datasets, evaluation scripts, and leaderboards are publicly released.

Arash Asgari, Huan Wu, Amirreza Naziri, Mojtaba Kolahdouzi, Laleh Seyyed-Kalantari

Bias evaluation in large language models (LLMs) uses many metrics and benchmarks, but lacks a systematic way to measure agreement across bias metrics and models. As a result, improvements observed under one metric may contradict another, and model rankings may reflect benchmark-specific artifacts rather than stable bias profiles. In this work, we introduce Metric Agreement Score (MeAS) and Model Agreement Score (MoAS), which quantify cross-metric and cross-model agreement in bias rankings, respectively. We apply these measures to eight LLMs, seven bias metrics, and nine corpora. Our results reveal disagreement among both metrics and models: Contrary to expectations, we find that metrics within the same category (generation-based and probabilistic) often behave independently of each other. For instance, HONEST shows independence with toxicity metrics, and the Context Association Test shows no correlation with Language Modeling Bias metric. At the model level, DeepSeek-family models invert bias rankings relative to most others, indicating that the model family strongly shapes specific bias profiles. These findings challenge the assumption that bias mitigation is universally transferable and highlight the need for agreement-aware evaluation.

Itay Yona, Amir Sarid, Michael Karasik, Yossi Gandelsman

We introduce Doublespeak, a simple in-context representation hijacking attack against language models. The attack works by systematically replacing a harmful keyword (e.g., bomb) with a benign token (e.g., carrot) across multiple in-context examples, provided as a prefix to a harmful request. We demonstrate that this substitution leads to the internal representation of the benign token converging toward that of the harmful one, effectively embedding the harmful semantics under a euphemism. As a result, superficially innocuous prompts (e.g., "How to build a carrot?") are internally interpreted as disallowed instructions ("How to build a bomb?"), thereby bypassing the model’s safety alignment. We use interpretability tools to show this semantic shift occurs progressively across layers. Doublespeak is optimization-free, broadly transferable across model families, and achieves strong success rates on closed-source systems, reaching 74% on Llama-3.3-70B-Instruct with a single-sentence context override. Our findings highlight a new attack surface in LM latent space, indicating that current alignment strategies are insufficient and should instead operate at the representation level.

Jiateng Liu, Zhenhailong Wang, Xiaojiang Huang, Yingjie Li, Xiang Li, Chenlei Guo, Xing Fan, Ruhi Sarikaya, Heng Ji

Large language model agents rely on in-context policy documents encoding diverse business rules. As businesses scale, these documents grow, creating substantial computational overhead and motivating internalization methods that embed policy into model priors. Prior work focuses on generic prompts, but we find agentic policies span multiple complexity levels and demand heavier reasoning, posing greater challenges. We introduce an agentic benchmark generator with Controllable Complexity in agent policy across four levels, enabling systematic evaluation of agents under increasing complexity and providing a testbed for policy internalization. Our analysis shows that workflow-governing policy specifications are the hardest to reason over, and that SFT on gold trajectories with chain-of-thought is data-hungry and struggles at high complexity. We propose Category-Aware Policy Continued Pretraining, an automated pipeline that analyzes policies, extracts key specifications, categorizes them into factual, behavioral, and conditional types, and isolates those driving workflow complexity. This enables targeted “therapy” by synthesizing specialized training data for each type and improving internalization via an autoregressive pretraining loss. Extensive experiments show our synthetic data and objective consistently improve performance. Combined with SFT, our method outperforms the baseline across different settings, especially in data-sparse and high-complexity regimes, with gains up to 41% and 22% on Qwen-3-32B. Overall, we achieve 97.3% prompt reduction on our benchmark, and on \tau-Bench we further improve performance while reducing prompt requirements with very limited SFT data.

Binchi Zhang, Xujiang Zhao, Jundong Li, Haifeng Chen, Zhengzhang Chen

Large language models (LLMs) are increasingly deployed in culturally sensitive real-world tasks. However, existing cultural alignment approaches fail to align LLMs’ broad cultural values with the specific goals of downstream tasks and suffer from cross-culture interference. We propose CultureManager, a novel pipeline for task-specific cultural alignment. CultureManager synthesizes task-aware cultural data in line with target task formats, grounded in culturally relevant web search results. To prevent conflicts between cultural norms, it manages multi-culture knowledge learned in separate adapters with a culture router that selects the appropriate one to apply. Experiments across five national cultures and ten culture-sensitive tasks show consistent improvements over prompt-based and fine-tuning baselines. Our results demonstrate the necessity of task adaptation and modular culture management for effective cultural alignment.

Qizhuo Xie, Yunhui Liu, Yu Xing, Qianzi Hou, Xudong Jin, Tao Zheng, Tieke He

Large Language Models (LLMs) have shown immense potential in Knowledge Graph Completion (KGC), yet bridging the modality gap between continuous graph embeddings and discrete LLM tokens remains a critical challenge. While recent quantization-based approaches attempt to align these modalities, they typically treat quantization as flat numerical compression, resulting in semantically entangled codes that fail to mirror the hierarchical nature of human reasoning. In this paper, we propose GS-Quant, a novel framework that generates semantically coherent and structurally stratified discrete codes for KG entities. Unlike prior methods, GS-Quant is grounded in the insight that entity representations should follow a linguistic coarse-to-fine logic. We introduce a Granular Semantic Enhancement module that injects hierarchical knowledge into the codebook, ensuring that earlier codes capture global semantic categories while later codes refine specific attributes. Furthermore, a Generative Structural Reconstruction module imposes causal dependencies on the code sequence, transforming independent discrete units into structured semantic descriptors. By expanding the LLM vocabulary with these learned codes, we enable the model to reason over graph structures isomorphically to natural language generation. Experimental results demonstrate that GS-Quant significantly outperforms existing text-based and embedding-based baselines.

Haikang Deng, Po-Nien Kung, Nanyun Peng

Large language models (LLMs) are increasingly adept at solving complex problems, such as mathematical reasoning and automatic evaluation. However, performance often degrades when prompts intertwine task instructions with rigid formatting requirements. This entanglement creates competing goals for the model, hindering its reasoning capabilities. To address this, we introduce Deco-G, a decoding framework that explicitly decouples format adherence from problem solving. Deco-G delegates format adherence to a separate Format Estimation Module (FEM), which performs probabilistic lookahead to estimate future format compliance rate and reweighs token probabilities, allowing the LLM to focus solely on task resolution. To make this approach both practical and efficient, we introduce three key innovations: instruction-aware distillation, a flexible trie-building algorithm, and HMM state pruning. Experiments across mathematical reasoning, event argument extraction, and LLM-as-a-judge demonstrate that Deco-G constantly gains over prompting or structured generation baselines, with guaranteed format compliance.

Yiqing Zhang, Xiaozhong Liu, Fabricio Murai

Trustworthy biomedical question answering (QA) systems must not only provide accurate answers but also justify them with current, verifiable evidence. Retrieval-augmented approaches partially address this gap but lack mechanisms to iteratively refine poor queries, whereas self-reflection methods kick in only after full retrieval is completed. In this context, we introduce PubMed Reasoner, a biomedical QA agent composed of three stages: **self-critic query refinement** evaluates MeSH terms for coverage, alignment, and redundancy to enhance PubMed queries based on partial (metadata) retrieval; **reflective retrieval** processes articles in batches until sufficient evidence is gathered; and **evidence-grounded response generation** produces answers with explicit citations. PubMed Reasoner with a GPT-4o backbone achieves **78.32%** accuracy on PubMedQA, slightly surpassing human experts, and showing consistent gains on MMLU Clinical Knowledge. Moreover, LLM-as-judge evaluations prefer our responses across: reasoning soundness, evidence grounding, clinical relevance, and trustworthiness. By orchestrating retrieval-first reasoning over authoritative sources, our approach provides practical assistance to clinicians and biomedical researchers while controlling compute and token costs.

Yuqing Yang, Qi Zhu, Zhen Han, Boran Han, Zhengyuan Shen, Shuai Wang, Vassilis N. Ioannidis, Huzefa Rangwala

While large language models (LLMs) perform well on table tasks, they still make data referencing errors (DREs), i.e., incorrectly citing or omitting table values, despite understanding the table structure. Beyond final-answer accuracy, DREs directly compromise the correctness and reliability of intermediate reasoning steps. Yet prior studies have only offered limited, small-scale analyses. In this work, we present the first systematic evaluation of tabular data referencing errors across different models and tasks. Our results show that DREs occur across all tested models (1.7B to 20B parameters). Furthermore, we demonstrate that incorporating data referencing as a critic significantly improves answer accuracy up to 12.0%, through critic-based filtering and rejection sampling. Finally, we trained a lightweight 4B-parameter critic model that achieves an average F1 score of 78.2% in detecting both in-distribution and out-of-distribution DREs, and effectively assists inference for larger models.

Dayu Wang, Jiaye Yang, Weikang Li, Jiahui Liang, Yang Li, Deguo Xia, Jizhou Huang

Large Language Models (LLMs) often suffer from "Reasoning Collapse" on challenging mathematical reasoning tasks, where stochastic sampling produces lexical variations of the same erroneous logic rather than genuine semantic exploration. We observe that failed reasoning traces are often associated with a low-rank bias manifold in the model’s hidden-state geometry, which reduces exploration toward corrective solution directions. To address this, we propose Spectral Orthogonal Exploration (SOE), a geometric inference framework under a "Student Guides Teacher" paradigm. Instead of using a weak auxiliary agent for imitation, SOE uses it as an orthogonal probe to introduce semantically heterogeneous reasoning signals into the teacher’s orthogonal complement of its dominant subspace. This intervention steers the teacher toward more diverse reasoning trajectories and improves exploration beyond standard sampling. Experiments on mathematical benchmarks show that SOE improves average accuracy by 62.4% and average sampling efficiency by 113.7% over baseline methods, suggesting that geometric interventions can be effective for mitigating reasoning collapse in mathematical reasoning. We further provide preliminary evidence that SOE is also effective on logic and code generation benchmarks. Code is available at https://github.com/dayuwang401/spectral-orthogonal-exploration.

Binchi Zhang, Zhengzhang Chen, Zaiyi Zheng, Jundong Li, Haifeng Chen

Large Language Models (LLMs) have achieved remarkable success in natural language processing by encoding extensive knowledge, but their utility relies on timely updates as human knowledge keeps evolving. In this paper, we investigate the problem of LLM knowledge updates, which requires simultaneously unlearning unwanted information and learning new knowledge. Existing approaches that tackle unlearning and learning separately encounter *task conflicts* and *knowledge management issues* when applied to comprehensive knowledge updates.In this paper, we validate our findings with theoretical analysis and empirical evidence, and propose LOKA, a conflict-aware framework for Large language mOdel Knowledge updAtes. During training, LOKA introduces an adaptive knowledge memory approach in which updated knowledge is allocated across multiple memory units. During inference, LOKA retrieves the most relevant memory unit from the knowledge memory and integrates it with the original LLM to apply updated knowledge, while a learning-based router controls the activation of the knowledge memory to improve knowledge utilization. Extensive experiments demonstrate the efficacy of LOKA in achieving accurate, flexible, and conflict-aware knowledge updates.

Jonathan Cook, Tim Rocktäschel, Jakob Nicolaus Foerster, Dennis Aumiller, Alex Wang

Much recent progress in Large Language Model (LLM) performance has been driven by verifiable feedback in deterministic domains like mathematics and code. However, scaling reinforcement learning (RL) and test-time compute in domains for which strict verification is infeasible remains a challenge. A common approach is to use an LLM-as-judge, which often relies on opaque, monolithic scores. In this work, we propose that AI feedback is most effective when decomposed into granular, prompt-specific checklists. To transform these checklists into a scalar reward, we introduce DIVA: DIscriminative VAriance weighting, a dynamic aggregation scheme that prioritises checklist items based on their ability to distinguish quality across a candidate pool. This ensures the reward signal focuses on the most salient criteria for a given prompt and response group, rather than being diluted by trivial or redundant constraints. Our approach yields an 11.8% win-rate improvement on AlpacaEval 2.0 using Qwen3-8B, outperforming holistic reward models and existing checklist baselines. Beyond training, we show that these checklists serve as a structured policy improvement operator at inference time - by using the model’s own checklist evaluation as localised contextual feedback, the model can iteratively refine its output. This self-correction mechanism outperforms free-form sequential self-correction, offering a unified and interpretable framework for scaling both training-time and test-time performance in domains lacking strict verifiers.

Yilun Zhu, Yuan Zhuang, Nikhita Vedula, Dushyanta Dhyani, Shaoyuan Xu, Mohsen Bayati, Bryan Wang, Shervin Malmasi

Many applications of LLM-based text regression require predicting a full conditional distribution rather than a single point value. We study distributional regression under empirical-quantile supervision, where each input is paired with multiple observed quantile outcomes, and the target distribution is represented by a dense grid of quantiles. We address two key limitations of current approaches: the lack of local grounding for distribution estimates, and the reliance on shared representations that create an indirect bottleneck between inputs and quantile outputs. In this paper, we introduce Quantile Token Regression, which, to our knowledge, is the first work to insert dedicated quantile tokens into the input sequence, enabling direct input-output pathways for each quantile through self-attention. We further augment these quantile tokens with retrieval, incorporating semantically similar neighbor instances and their empirical distributions to ground predictions with local evidence from similar instances. We also provide the first theoretical analysis of loss functions for quantile regression, clarifying which distributional objectives each optimizes. Experiments on the Inside Airbnb and StackSample benchmark datasets with LLMs ranging from 1.7B to 14B parameters show that quantile tokens with neighbors consistently outperform baselines (\sim4 points lower MAPE and 2\times narrower prediction intervals), with especially large gains on smaller and more challenging datasets where quantile tokens produce substantially sharper and more accurate distributions.

Ehsan Hoseinzade, Ke Wang, Anandharaju Durai Raju

Table annotation is crucial for making web and enterprise tables usable in downstream NLP applications. Unlike textual data where learning semantically rich token or sentence embeddings often suffice, tables are structured combinations of columns wherein useful representations must jointly capture column’s semantics and the inter-column relationships. Existing models learn by linearizing the 2D table into a 1D token sequence and encoding it with pretrained language models (PLMs) such as BERT. However, this leads to limited semantic quality and weaker generalization to unseen or rare values compared to modern LLMs, and degraded structural modeling due to 2D-to-1D flattening and context-length constraints. We propose TabEmb, which directly targets these limitations by decoupling semantic encoding from structural modeling. An LLM first produces semantically rich embeddings for each column, and a graph-based module over columns then injects relationships into the embeddings, yielding joint semantic–structural representations for table annotation. Experiments show that TabEmb consistently outperforms strong baselines on different table annotation tasks. Source code and datasets are available at https://github.com/hoseinzadeehsan/TabEmb

Lei Hsiung, Tianyu Pang, Yung-Chen Tang, Linyue Song, Tsung-Yi Ho, Pin-Yu Chen, Yaoqing Yang

Recent advancements in large language models (LLMs) have underscored their vulnerability to safety alignment jailbreaks, particularly when subjected to downstream fine-tuning. However, existing mitigation strategies primarily focus on reactively addressing jailbreak incidents after safety guardrails have been compromised, removing harmful gradients during fine-tuning, or continuously reinforcing safety alignment throughout fine-tuning. As such, they tend to overlook a critical upstream factor: the role of the original safety-alignment data. This paper therefore investigates the degradation of safety guardrails through the lens of representation similarity between upstream alignment datasets and downstream fine-tuning tasks. Our experiments demonstrate that high similarity between these datasets significantly weakens safety guardrails, making models more susceptible to jailbreaks. Conversely, low similarity between these two types of datasets yields substantially more robust models and thus reduces harmfulness score by up to 10.33%. By highlighting the importance of upstream dataset design in the building of durable safety guardrails and reducing real-world vulnerability to jailbreak attacks, these findings offer actionable insights for fine-tuning service providers to prioritize upstream models with low jailbreak risk.

António Loison, Quentin Macé, Antoine Edy, Victor Xing, Tom Balough, Gabriel de Souza P. Moreira, Bo Liu, Manuel Faysse, Celine Hudelot, Gautier Viaud

Retrieval-Augmented Generation (RAG) pipelines must address challenges beyond simple single-document retrieval, such as interpreting visual elements (tables, charts, images), synthesizing information across documents, and providing accurate source grounding. Existing benchmarks fail to capture this complexity, often focusing on textual data, single-document comprehension, or evaluating retrieval and generation in isolation. We introduce ViDoRe V3, a comprehensive multimodal RAG benchmark featuring multi-type queries over visually rich document corpora. It covers 10 datasets across diverse professional domains, comprising ~26,000 document pages paired with 3,099 human-verified queries, each available in 6 languages. Through 12,000 hours of human annotation effort, we provide high-quality annotations for retrieval relevance, bounding box localization, and verified reference answers. Our evaluation of state-of-the-art RAG pipelines reveals that visual retrievers outperform textual ones, late-interaction models and textual reranking substantially improve performance, and hybrid or purely visual contexts enhance answer generation quality. However, current models still struggle with non-textual elements, open-ended queries, and fine-grained visual grounding. To encourage progress in addressing these challenges, the benchmark is released under a commercially permissive license.

Leonardo Ranaldi

AI agents struggle to operate within open and dynamic environments because they lack a fundamental capacity: the autonomous generation of abstractions. Current models remain static entities, incapable of compressing the infinite complexity of the real world into generalisable concepts once their training phase has concluded.We introduce EVA (Evolving Agents), a novel paradigm for autonomous learning driven by pseudo-symbolic abstraction. EVA introduces a meta-control system that dynamically orchestrates observation and active interaction to distil on-the-fly abstract representations of states, actions, and goals. By disentangling contextual noise from pure logical reasoning, these pseudo-symbolic abstractions allow the agent to construct a highly robust internal curriculum.EVA leverages these self-generated abstractions to form an internal curriculum. This continuous compression of raw sensorimotor experience into reusable concepts allows the agent to independently guide its own exploration, planning, and error correction. Structured upon a bi-level evolutionary-developmental (Evo/Devo) framework, EVA demonstrates how the dynamic refinement of abstractions enables rapid adaptation to unforeseen scenarios. This approach resolves the domain mismatch problem and lays the groundwork for truly autonomous, continuously evolving AI models.