论文检索

Dongwon Noh, Donghyeok Koh, Junghun Yuk, Gyuwan Kim, Jae Yong Lee, KyungTae Lim, Cheoneum Park

Prior benchmarks for evaluating the domain-specific knowledge of large language models (LLMs) lack the scalability to handle complex academic tasks. To address this, we introduce ScholarBench, a benchmark centered on deep expert knowledge and complex academic problem-solving, which evaluates the academic reasoning ability of LLMs and is constructed through a three-step process. ScholarBench targets more specialized and logically complex contexts derived from academic literature, encompassing five distinct problem types. Unlike prior benchmarks, ScholarBench evaluates the abstraction, comprehension, and reasoning capabilities of LLMs across eight distinct research domains. To ensure high-quality evaluation data, we define category-specific example attributes and design questions that are aligned with the characteristic research methodologies and discourse structures of each domain. Additionally, this benchmark operates as an English-Korean bilingual dataset, facilitating simultaneous evaluation for linguistic capabilities of LLMs in both languages. The benchmark comprises 5,031 examples in Korean and 5,309 in English, with even state-of-the-art models like o3-mini achieving an average evaluation score of only 0.543, demonstrating the challenging nature of this benchmark.

Chang Yang, Peng Zhang, Jing Zhang, Hui Gao, Changhao Song

The rise of large language models (LLMs) has fundamentally reshaped the technological paradigm of rumor detection, offering transformative opportunities to construct adaptive detection systems while simultaneously ushering in new threats, such as “logically perfect rumors”. This paper aims to unify existing methods in the field of rumor detection and reveal the logical mechanisms behind them. From the perspective of complex systems, we innovatively propose a Cognition-Interaction-Behavior (CIB) tri-level framework for rumor detection based on collective intelligence and explore the synergistic relationship between LLMs and collective intelligence in rumor governance. We identify promising future research directions, including advancing agent-based modeling to capture complex rumor dynamics, addressing emerging challenges unique to the LLM era, and interdisciplinary perspectives. We hope this work lays a theoretical foundation for next-generation rumor detection paradigms and offers valuable insights for advancing the field.

Yue Qiu, Yujan Ting, Pei Dong, Terrence Chen, Weijing Huang

Reinforcement learning (RL) for large language models (LLMs) typically requires clear reward signals, which are often unavailable for open-ended (OE) questions where answer evaluation is ambiguous without scalable expert labeling. We investigate whether LLMs benefit from training on mixed data with varying reward clarity. Our approach combines Multiple-choice questions (MCQs), which offer clear binary rewards, with OE questions, for which we use simpler, potentially noisy rewards such as Jaccard similarity or LLM-based evaluators. We hypothesize that MCQs can stabilize training when mixed with OE questions. Our experiments show this mixed-data approach consistently improves medical question-answering performance across model scales.

Zhirui Chen, Wei Shen, Jiashui Huang, Ling Shao

Despite significant progress, large language models (LLMs) still struggle with long contexts due to memory limitations and their inability to tackle complex and long-context tasks. Additionally, LLMs often suffer from a lack of transparency and are prone to producing hallucinations. To address these challenges, we propose JERR, a novel framework designed to enhance long-context comprehension via graph-based reasoning in LLMs. JERR integrates three key components: synopsis extraction, graph construction, and relational reasoning. First, synopsis is extracted by chunking text strategically, allowing the model to summarize and understand information more efficiently. Second, we build a directed acyclic graph (DAG) to resolve redundancy, ensuring logical consistency and clarity. Finally, we incorporate Monte Carlo Tree Search (MCTS) to help the model navigate complex reasoning paths, ensuring more accurate and interpretable outputs. This framework provides a novel solution that enables LLMs to handle extended contexts and complex reasoning tasks with improved reliability and transparency. Experimental results show that JERR consistently outperforms all baselines on the ROUGE and F1 metrics, achieving the highest scores on the LLM-Rater evaluation.

Yifei Xu, Yingjie Zong, Wang Zhonghua, Sirui Wu, Yuan Rao, Dan Zhang, Shuiguang Deng

News writing empowered by large language models (LLMs) has emerged as a prevalent trend due to their efficiency and scalability. This paradigm necessitates dynamic information acquisition, knowledge structuring, and precise viewpoint articulation. However, current approaches often rely on superficially retrieved information and oversimplified knowledge enumeration, resulting in shallow, repetitive, and unordered outputs. Additionally, the lack of controllability over narrative viewpoints fails to align with user-defined preferences. To address these limitations, we propose an LLM-based multi-agent controllable news writing framework termed CtrlNews. The framework simulates expert questioning through automated role assignment and question generation followed by a three-layer hierarchical gravitational graph iteratively refined via expansion-reflection cycles. Besides, we elaborate a fine-grained viewpoint control mechanism to precisely regulate bias, emotion, and exaggeration attributes. When composing long-form news articles, the controlled viewpoints are extended via emotion-preserving composition and self-reflection refinement to ensure the consistency of viewpoint control and prevent the dilution of the control effect. Experiments on quality and control effect evaluation, news dissemination effect assessment, and human evaluation demonstrate significant improvements across multiple metrics compared to existing methods.

Alina Fastowski, Bardh Prenkaj, Gjergji Kasneci

Ensuring the robustness of factual knowledge in LLMs is critical for reliable applications in tasks such as question answering and reasoning. However, existing evaluation methods predominantly focus on performance-based metrics, often investigating from the perspective of prompt perturbations, which captures only the externally triggered side of knowledge robustness. To bridge this gap, we introduce a principled approach to measure factual robustness from the perspective of the generation process by analyzing token distribution entropy in combination with temperature scaling sensitivity. These two factors build the Factual Robustness Score (FRS), a novel metric which quantifies the stability of a fact against perturbations in decoding conditions, given its initial uncertainty. To validate our approach, we conduct extensive experiments on 5 LLMs across 3 closed-book QA datasets (SQuAD, TriviaQA, and HotpotQA). We show that factual robustness varies significantly – smaller models report an FRS of 0.76, larger ones 0.93 – with accuracy degrading by ~60% under increased uncertainty. These insights demonstrate how entropy and temperature scaling impact factual accuracy, and lay a foundation for developing more robust knowledge retention and retrieval in future models. We release our code at https://github.com/afastowski/frs.

Mingrui Xie, Tianxiang Xu, Qianhai Tang, Shanming Yao, Xiaofeng Zhang, Junliang Du

Large Vision–Language Models (LVLMs) have garnered substantial interest owing to their impressive ability to interpret visual inputs and converse with users.Nevertheless, LVLMs still suffer from object hallucination – generating descriptions for objects that are absent from the image, which undermines reliability and hinders real-world deployment. We propose DAPE-BR, a positional-alignment scheme that (i) preserves the pretrained weight order while globally—- visual–text distances, (ii) embeds an isotropic fused patch-distance metric, and (iii) applies a patch-distance causal mask to enforce spatial causality. Extensive experiments on POPE, MMStar and SQA show that DAPE-BR consistently reduces hallucinations and boosts.

Zihao Guo, Hongtao Lv, Chaoli Zhang, Yibowen Zhao, Yixin Zhang, Lizhen Cui

Prior research indicates that although large language models (LLMs) can precisely articulate the theoretical probability distributions associated with optimal strategic choices, their actual decision-making systematically diverges from these prescriptions—a phenomenon we define as the cognition–behaviour gap in LLMs. For example, in a Rock–Paper–Scissors (RPS) game, LLMs correctly identify the strategy of Nash equilibrium as selecting each action (Rock, Paper, Scissors) with equal probability \frac{1}{3}, but their observed choices systematically deviate from this uniform distribution. Through a comprehensive evaluation of 20 state-of-the-art LLMs, we identify two critical insights: (1) we demonstrate that intrinsic biases inherited from pre-training corpora alone are insufficient to explain the observed deviations; (2) we introduce a semantic-free paradigm that strips away intrinsic biases to isolate pure positional bias-LLMs exhibit distinct position preferences—for example, o1 favours the first option, DeepSeek-V3 peaks the middle and DeepSeek-R1 shows a bimodal bias toward first and last positions. Our findings advocate innovation to bridge the gap between strategic reasoning and decision-making in LLMs.

Zheni Zeng, Jiayi Chen, Huimin Chen, Yukun Yan, Yuxuan Chen, Zhenghao Liu, Zhiyuan Liu, Maosong Sun

Large language models (LLMs) exhibit human-like intelligence, enabling them to simulate human behavior and support various applications that require both humanized communication and extensive knowledge reserves. Efforts are made to personify LLMs with special training data or hand-crafted prompts, while correspondingly faced with challenges such as insufficient data usage or rigid behavior patterns. Consequently, personified LLMs fail to capture personified knowledge or express persistent opinion. To fully unlock the potential of LLM personification, we propose PersLLM, a framework for better data construction and model tuning. For insufficient data usage, we incorporate strategies such as Chain-of-Thought prompting and anti-induction, improving the quality of data construction and capturing the personality experiences, knowledge, and thoughts more comprehensively. For rigid behavior patterns, we design the tuning process and introduce automated DPO to enhance the specificity and dynamism of the models’ personalities, which leads to a more natural opinion communication. Both automated metrics and expert human evaluations demonstrate the effectiveness of our approach. Case studies in human-machine interactions and multi-agent systems further suggest potential application scenarios and future directions for LLM personification.

Idan Kashani, Avi Mendelson, Yaniv Nemcovsky

Large language models (LLMs) achieve impressive results over various tasks, and ever-expanding public repositories contain an abundance of pre-trained models. Therefore, identifying the best-performing LLM for a given task is a significant challenge. Previous works have suggested learning LLM representations to address this. However, these approaches present limited scalability and require costly retraining to encompass additional models and datasets. Moreover, the produced representation utilizes distinct spaces that cannot be easily interpreted. This work presents an efficient, training-free approach to representing LLMs as linear operators within the prompts’ semantic task space, thus providing a highly interpretable representation of the models’ application. Our method utilizes closed-form computation of geometrical properties and ensures exceptional scalability and real-time adaptability to dynamically expanding repositories. We demonstrate our approach on success prediction and model selection tasks, achieving competitive or state-of-the-art results with notable performance in out-of-sample scenarios.

Yue Xin, Chen Shen, Shaotian Yan, Xiaosong Yuan, Yaoming Wang, Xiaofeng Zhang, Chenxi Huang, Jieping Ye

Chain-of-Thought (CoT) prompting enhances the math reasoning capability of large language models (LLMs) to a large margin. However, the mechanism underlying such improvements remains unexplored. In this paper, we present SalaMAnder (Shapley-based Mathematical Expression Attribution and Metric), a theoretically grounded methodology as well as a mathematically rigorous evaluation metric for quantifying component-level contributions in few-shot CoT reasoning. Concretely, we leverage the Shapley value for mathematical expression attribution and develop an efficient stratified sampling algorithm that significantly reduces the computational complexity. Besides, we develop the CoSP (Cardinality of Shapley Positives) metric through covariance analysis. Comprehensive validation across popular LLM models and diverse mathematical benchmarks demonstrates that the CoSP metric within our SalaMAnder framework exhibits a robust monotonic correlation with model performance, not only providing theoretical explanations for the empirical success of existing few-shot CoT but also establishing mathematically rigorous principles for prompt construction optimization. Furthermore, we verify the reliability of the explanation, based on which we unify the insights of previous work.

Tao Fan, Weijing Chen, Yan Kang, Guoqiang Ma, Hanlin Gu, Yuanfeng Song, Lixin Fan, Qiang Yang

Large Language Models (LLMs) have emerged as a transformative force in artificial intelligence, demonstrating exceptional proficiency across various tasks. However, their deployment in resource-constrained environments and concerns over user data privacy pose significant challenges. In contrast, Small Language Models (SLMs) offer computational efficiency but often lag in performance. To address these issues, we propose FedCoT, a federated framework designed for the Chain-of-Thought (CoT) distillation of knowledge from LLMs to SLMs, while ensuring the preservation of clients’ data privacy. FedCoT ensures secure and efficient knowledge transfer from an LLM on a high-powered server to an SLM on a resource-constrained client, while adhering to privacy requirements. Leveraging perturbed prompts and rationales generated through the CoT approach, the framework enhances the performance of the client’s SLM without compromising user data privacy within a multi-task learning framework. We propose two privacy protection strategies: the Exponential Mechanism Strategy and the Adaptive Exponential Mechanism Strategy, which balance user prompt privacy and the usability of rationales. Empirical evaluation on various text generation tasks demonstrates the effectiveness of FedCoT in training task-specific SLMs with enhanced performance while prioritizing data privacy protection. Our code has been contributed to the FATE open-source project and is now publicly accessible at https://github.com/FederatedAI/FATE-LLM/tree/main/python/fate_llm/algo/fedcot

Tianlong Wang, Junzhe Chen, Weibin Liao, Xueting Han, Jing Bai

Reinforcement learning (RL) on self-generated data has emerged as a promising paradigm for improving reasoning in large language models (LLMs). However, RL relies on accurate reward signals, which are scarce in many domains, making it critical to train models that can generalize to unseen problems. Existing methods often focus on task-specific or domain-specific reasoning, lacking consideration for generalization and may degrade performance on other tasks. To address this, we distinguish between abstract plans, representing high-level problem-solving strategies, and concrete solutions, proposing that learning plans develops transferable general reasoning capabilities and promotes better generalization. Building on this insight, we propose PlanLearn, a framework that combines plan-based search with Step-level Advantage Preference Optimization (Step-APO) to optimize plan learning. Experimental results show that PlanLearn, trained exclusively on GSM8K and MATH, not only significantly improves in-domain performance but also enhances out-of-domain benchmarks, such as HumanEval (+12.2%), GPQA (+8.6%), ARC-C (+4.0%), MMLU-STEM (+2.2%), and BBH (+1.8%). The code is available at https://github.com/tianlwang/PlanLearn.

Patrick Sutanto, Joan Santoso, Esther Irawati Setiawan, Aji Prasetya Wibawa

Encoder models offer efficiency for specific tasks, but their performance depend on data availability. While Large Language Models (LLMs) excel at few-shot learning, their direct application in real-world scenarios is often hindered by their high computational cost. To address this challenge, we propose a simple yet effective approach that uses LLMs for data generation and scoring to improve encoder only model performance. We evaluate this framework on few-shot Multiple Choice Question Answering (MCQA), an important task where acquiring labeled data is costly. Our approach utilizes LLMs to create MCQA questions and choices (exploring both direct JSON and decomposed generation methods) and assigns probability scores to these choices. This generated data and the LLM scores are then used to fine-tune smaller and more efficient DeBERTa-v3-base using distillation loss. Extensive experiments on the MMLU benchmark demonstrate that our method improves accuracy from 28.9% to 39.3%, representing a gain of over 10% compared to a baseline finetuned directly on 5-shot examples. This shows the effectiveness of LLM-driven data generation and knowledge distillation for few-shot MCQA.

Mengxiang Zhang, Lingyuan Liu

Knowledge distillation (KD) enables the compression of large language models (LLMs) by transferring knowledge from a high-capacity teacher model to a resource-efficient student model, maintaining competitive performance for tasks such as instruction following. However, conventional white-box KD methods often suffer from training-inference mismatches and suboptimal performance due to the asymmetric nature of Kullback-Leibler divergence (KLD) and reliance on computationally expensive student-generated outputs. To address these challenges, we propose Least-to-Most Prompting Knowledge Distillation (L2M-KD), a novel white-box KD method grounded in curriculum learning (CL) and adaptive loss design. L2M-KD employs a two-pronged approach: (1) a CL strategy that ranks training samples by difficulty using Rouge-L scores, partitioning them into easy-to-hard subsets across multiple stages, and (2) an adaptive KD loss that transitions from KLD to skew KLD, dynamically adjusting teacher guidance to mitigate mode-averaging and over-smoothing. Extensive experiments on instruction-following tasks demonstrate that L2M-KD outperforms existing white-box KD methods, achieving superior student model performance with reduced computational overhead by leveraging ground-truth outputs exclusively. Our findings underscore the efficacy of difficulty-aware training and adaptive teacher guidance, offering a computationally efficient and robust approach to LLM compression.

Jiaqi Wu, Chen Chen, Chunyan Hou, Xiaojie Yuan

With the widespread real-world deployment of large language models (LLMs), ensuring their behavior complies with safety standards has become crucial. Jailbreak attacks exploit vulnerabilities in LLMs to induce undesirable behavior, posing a significant threat to LLM safety. Previous defenses often fail to achieve both effectiveness and efficiency simultaneously. Defenses from a representation perspective offer new insights, but existing interventions cannot dynamically adjust representations based on the harmfulness of the queries. To address this limitation, we propose SafeIntervention (SafeInt), a novel defense method that shields LLMs from jailbreak attacks through safety-aware representation intervention. Built on our analysis of the representations of jailbreak samples, the core idea of SafeInt is to relocate jailbreak-related representations into the rejection region. This is achieved by intervening in the representation distributions of jailbreak samples to align them with those of unsafe samples. We conduct comprehensive experiments covering six jailbreak attacks, two jailbreak datasets, and two utility benchmarks. Experimental results demonstrate that SafeInt outperforms all baselines in defending LLMs against jailbreak attacks while largely maintaining utility. Additionally, we evaluate SafeInt against adaptive attacks and verify its effectiveness in mitigating real-time attacks.

Lvzhou Luo, Yixuan Cao, Ping Luo

Retrieval-augmented generation improves the factual accuracy of Large Language Models (LLMs) by incorporating external context, but often suffers from irrelevant retrieved content that hinders effectiveness. Context compression addresses this issue by filtering out irrelevant information from context before LLM generation. However, existing methods struggle to adaptively adjust compression rates for different context, maintain low latency and integrate information across multiple documents. To overcome these limitations, We introduce AttnComp, an adaptive, efficient and context-aware compression framework. By leveraging the attention mechanism of LLMs to identify relevant information, AttnComp employs a Top-P compression algorithm to retain the minimal set of documents whose cumulative attention weights exceeds a predefined threshold. In addition to compression, AttnComp estimates response confidence by assessing the overall relevance of the retrieved content, enabling users to gauge response reliability. Experiments demonstrate that AttnComp outperforms existing compression methods and uncompressed baselines, achieving higher accuracy with substantial compression rates and lower latency.

Zoey Liu, Masoud Jasbi, Christan Grant, Kenji Sagae, Emily Prud’hommeaux

Building an NLP training set for understudied languages, including Indigenous and endangered languages, often faces challenges due to varying degrees of resource limitations in the speaker communities. What are some reasonable approaches for training set construction in these cases? We address this question with POS tagging as the test case. Although many might consider POS tagging “a solved problem”, it remains a crucial task for descriptive linguistics and language documentation and requires laborious manual annotation. Drawing data from 12 language families, we compare in-context learning, active learning (AL), and random sampling. Our results suggest: (1) for communities whose language data can be ethically shared with an API, using only 1,000 randomly sampled tokens as prompt examples, the proprietary GPT-4.1-mini can deliver desirable performance (F1>0.83) on par with that from a training set of thousands of tokens in AL iterations; (2) in cases where communities prefer not to share data, 4,500-5,500 tokens selected from AL can yield reasonable results at a pace statistically significantly faster than random sampling, evidenced by growth curve modeling.

Aviv Slobodkin, Hagai Taitelbaum, Yonatan Bitton, Brian Gordon, Michal Sokolik, Nitzan Bitton Guetta, Almog Gueta, Royi Rassin, Dani Lischinski, Idan Szpektor

Subject-driven text-to-image (T2I) generation aims to produce images that align with a given textual description, while preserving the visual identity from a referenced subject image. Despite its broad downstream applicability—ranging from enhanced personalization in image generation to consistent character representation in video rendering—progress in this field is limited by the lack of reliable automatic evaluation. Existing methods either assess only one aspect of the task (i.e., textual alignment or subject preservation), misalign with human judgments, or rely on costly API-based evaluation. To address this gap, we introduce RefVNLI, a cost-effective metric that evaluates both textual alignment and subject preservation in a single run. Trained on a large-scale dataset derived from video-reasoning benchmarks and image perturbations, RefVNLI outperforms or statistically matches existing baselines across multiple benchmarks and subject categories (e.g., Animal, Object), achieving up to 6.4-point gains in textual alignment and 5.9-point gains in subject preservation.

Donghee Han, Hwanjun Song, Mun Yong Yi

Recent studies have explored integrating large langucage models (LLMs) into recommendation systems but face several challenges, including training-induced bias and bottlenecks from serialized architecture.To effectively address these issues, we propose a Query-to-Recommendation, a parallel recommendation framework that decouples LLMs from candidate pre-selection and instead enables direct retrieval over the entire item pool. Our framework connects LLMs and recommendation models in a parallel manner, allowing each component to independently utilize its strengths without interfering with the other. In this framework, LLMs are utilized to generate feature-enriched item descriptions and personalized user queries, allowing for capturing diverse preferences and enabling rich semantic matching in a zero-shot manner. To effectively combine the complementary strengths of LLM and collaborative signals, we introduce an adaptive reranking strategy. Extensive experiments demonstrate an improvement in performance up to 57%, while also improving the novelty and diversity of recommendations.