论文检索

Weiyuan Li, Xintao Wang, Siyu Yuan, Rui Xu, Jiangjie Chen, Qingqing Dong, Yanghua Xiao, Deqing Yang

As large language models (LLMs) grow more capable, they face increasingly diverse and complex tasks, making reliable evaluation challenging. The paradigm of LLMs as judges has emerged as a scalable solution, yet prior work primarily focuses on simple settings. Their reliability in complex tasks—where multi-faceted rubrics, unstructured reference answers, and nuanced criteria are critical—remains understudied. In this paper, we constructed ComplexEval Bench, a challenge benchmark designed to systematically expose and quantify Auxiliary Information Induced Biases. We systematically investigated and validated 6 previously unexplored biases across 12 basic and 3 advanced scenarios. Key findings reveal: (1) all evaluated models exhibit significant susceptibility to these biases, with bias magnitude scaling with task complexity; (2) notably, Large Reasoning Models (LRMs) show paradoxical vulnerability. Our in-depth analysis offers crucial insights for improving the accuracy and verifiability of evaluation signals, paving the way for more general and robust evaluation models.

Archie Sage, Jeroen Keppens, Helen Yannakoudakis

As interest grows in applying natural language processing (NLP) techniques to mental health, an expanding body of work explores the automatic detection and classification of cognitive distortions (CDs). CDs are habitual patterns of negatively biased or flawed thinking that distort how people perceive events, judge themselves, and react to the world. Identifying and addressing them is a central goal of therapy. Despite this momentum, the field remains fragmented, with inconsistencies in CD taxonomies, task formulations, and evaluation practices limiting comparability across studies. This survey presents the first comprehensive review of 38 studies spanning two decades, mapping how CDs have been implemented in computational research and evaluating the methods applied. We provide a consolidated CD taxonomy reference, summarise common task setups, and highlight persistent challenges to support more coherent and reproducible research. Alongside our review, we introduce practical resources, including curated evaluation metrics from surveyed papers, a standardised datasheet template, and an ethics flowchart, available online.

Dimitra Niaouri, Mohamed Rayane Ghilene, Michele Linardi, Julien Longhi

Analyzing Socially Unacceptable Discourse (SUD) online is a critical challenge for regulators and platforms amidst growing concerns over harmful content. While Pre-trained Masked Language Models (PMLMs) have proven effective for many NLP tasks, their performance often degrades in multi-label SUD classification due to overlapping linguistic cues across categories. In this work, we propose an artifact-guided pre-training strategy that injects statistically salient linguistic features, referred to as artifacts, into the masked language modelling objective. By leveraging context-sensitive tokens, we guide an importance-weighted masking scheme during pre-training to enhance generalization across discourse types. We further use these artifact signals to inform a lightweight dataset curation procedure that highlights noisy or ambiguous instances. This supports targeted relabeling and filtering, enabling more explainable and consistent annotation with minimal changes to the original data. Our approach provides consistent improvements in 10 datasets extensively used in SUD classification benchmarks.*Disclaimer: This article contains some extracts of unacceptable and upsetting language.*

Eojin Jeon, Mingyu Lee, Sangyun Kim, Junho Kim, Wanzee Cho, Tae-Eui Kam, SangKeun Lee

Emotion context sensitivity—the ability to adjust emotional responses based on contexts—is a core component of human emotional intelligence. For example, being told, “You can come with me if you want,” may elicit joy if the destination is a mall, but provoke fear if the destination is a trap house. As large language models (LLMs) are increasingly deployed in socially interactive settings, understanding this human ability becomes crucial for generating context-appropriate, emotion-aware responses. In this work, we introduce Trace, a novel benchmark for evaluating whether LLMs can understand emotion context sensitivity of humans. This benchmark consists of 1,626 social scenarios and comprises two complementary tests: a sensitivity test, which measures whether models can detect emotional shifts caused by context changes, and a robustness test, which evaluates whether models can maintain stable emotion predictions when context changes are emotionally irrelevant. Each scenario pair keeps the core event constant while systematically varying contextual details—time, place, or agent—based on insights from behavioral theory and emotion psychology. Experimental results show that even the best-performing LLMs lag behind human performance by 20% in the sensitivity test and 15% in the robustness test, indicating substantial room for improvement in emotion-aware reasoning.

Takateru Yamakoshi, Thomas L. Griffiths, R. Thomas McCoy, Robert D. Hawkins

Data-efficient training requires strong inductive biases. To the extent that transformer attention matrices encode syntactic relationships, we would predict that knowledge distillation (KD) targeting attention should selectively accelerate syntax acquisition relative to conventional logit-based KD. To test this hypothesis, we train GPT-2 student models on datasets ranging from 10K to 5M sentences using both distillation methods, evaluating them on both syntactic benchmarks and perplexity. Surprisingly, while logit-based KD dramatically improves data-efficiency, attention-based KD provides minimal benefit even for syntactic tasks. This suggests that output distributions provide sufficient supervisory signal for syntax acquisition, indicating that syntactic knowledge may be distributed throughout the network rather than localized in attention patterns.

Guang Yang, Yujie Zhu

Pre-trained language models (PLMs) are increasingly being applied to code-related tasks. Although PLMs have achieved good results, they do not take into account potential high-order data correlations within the code. We propose three types of high-order correlations in code tokens, i.e. abstract syntax tree family correlation, lexical correlation, and line correlation. We design a tokens and hyperedges generator to capture these high-order data correlations. We improve the architecture of hypergraph neural networks and combine it with adapter tuning to propose a novel hypergraph-based adapter (HGAdapter) to fine-tune PLMs. HGAdapter can encode high-order data correlations and is allowed to be inserted into various PLMs to enhance performance. Experiments were conducted on several public datasets, including six languages of code summarization and code clone detection tasks. Our methods improved the performance of PLMs in datasets to varying degrees. Experimental results validate the introduction of high-order data correlations that contribute to improved effectiveness.

Yongju Jia, Jiarui Ma, Xiangxian Li, Baiqiao Zhang, Xianhui Cao, Juan Liu, Yulong Bian

Pre-trained vision-language models (VLMs), such as CLIP, have demonstrated impressive capability in visual tasks, but their fine-tuning often suffers from bias in class-imbalanced scenes. Recent works have introduced large language models (LLMs) to enhance VLM fine-tuning withsupplementaryy semantic information. However, they often overlook inherent class imbalance in VLMs’ pre-training, which may lead to bias accumulation in downstream tasks. To address this problem, this paper proposes a Multi-dimensional Dynamic Prompt Routing (MDPR) framework. MDPR constructs a comprehensive knowledge base for classes, spanning multiple visual-semantic dimensions. During fine-tuning, the dynamic routing mechanism aligns global visual classes, retrieves optimal prompts, and balances fine-grained semantics, yielding stable predictions through logits fusion. Extensive experiments on long-tailed benchmarks, including CIFAR-LT, ImageNet-LT, and Places-LT, demonstrate that MDPR achieves comparable results with current SOTA methods. Ablation studies further confirm the effectiveness of our semantic library for tail classes and show that our dynamic routing operates with a slight increase in computational overhead, making MDPR a flexible and efficient enhancement for VLM fine-tuning under data imbalance. The codes are available in https://github.com/Sha843/MDPR.

Bhiman Kumar Baghel, Emma Jordan, Zheyuan Ryan Shi, Xiang Lorraine Li

Large Language Models (LLMs) are widely deployed in downstream tasks, but keeping their knowledge up-to-date via retraining or fine-tuning is often computationally expensive. Model editing provides a more efficient alternative by updating a targeted subset of parameters, which often follows the locate-and-edit paradigm. Despite this efficiency, existing methods are limited: edits may fail to inject knowledge (UnderEdit) or unintentionally disrupt unrelated neighboring knowledge (OverEdit). To address these challenges, we propose two complementary methods: **iterative model editing**, which applies successive edits to mitigate UnderEdit, and **neighbor-assisted model editing**, which incorporates neighboring knowledge during editing to reduce OverEdit. Our extensive experiments show that these techniques improve editing performance across multiple LLMs, algorithms, and benchmarks, reducing UnderEdit by up to 38 percentage points and OverEdit by up to 6, while remaining broadly applicable to any locate-and-edit method.

Taichi Aida, Danushka Bollegala

In Semantic Change Detection (SCD), it is a common problem to obtain embeddings that are both interpretable and high-performing. However, improving interpretability often leads to a loss in the SCD performance, and vice versa. To address this problem, we propose SCDTour, a method that orders and merges interpretable axes to alleviate the performance degradation of SCD. SCDTour considers both (a) semantic similarity between axes in the embedding space, as well as (b) the degree to which each axis contributes to semantic change. Experimental results show that SCDTour preserves performance in semantic change detection while maintaining high interpretability. Moreover, agglomerating the sorted axes produces a more refined set of word senses, which achieves comparable or improved performance against the original full-dimensional embeddings in the SCD task. These findings demonstrate that SCDTour effectively balances interpretability and SCD performance, enabling meaningful interpretation of semantic shifts through a small number of refined axes.

Armin Tourajmehr, Mohammad Reza Modarres, Yadollah Yaghoobzadeh

Large language models (LLMs) have demonstrated notable creative abilities in generating literary texts, including poetry and short stories. However, prior research has primarily centered on English, with limited exploration of non-English literary traditions and without standardized methods for assessing creativity. In this paper, we evaluate the capacity of LLMs to generate Persian literary text enriched with culturally relevant expressions. We build a dataset of user-generated Persian literary spanning 20 diverse topics and assess model outputs along four creativity dimensions—originality, fluency, flexibility, and elaboration—by adapting the Torrance Tests of Creative Thinking. To reduce evaluation costs, we adopt an LLM as a judge for automated scoring and validate its reliability against human judgments using intraclass correlation coefficients, observing strong agreement. In addition, we analyze the models’ ability to understand and employ four core literary devices: simile, metaphor, hyperbole, and antithesis. Our results highlight both the strengths and limitations of LLMs in Persian literary text generation, underscoring the need for further refinement.

ChenZhuo Zhao, Ziqian Liu, Xinda Wang, Junting Lu, Chaoyi Ruan

Prompt optimization is a practical and widely applicable alternative to fine tuning for improving large language model performance. Yet many existing methods evaluate candidate prompts by sampling full outputs, often coupled with self critique or human annotated preferences, which limits scalability, especially for smaller models or models that are not instruction tuned. We present PMPO (Probabilistic Metric Prompt Optimization), a unified framework that uses token level cross entropy as a direct, lightweight evaluation signal. PMPO locates low quality prompt segments via a masking based analysis and iteratively rewrites them to propose improved variants. Crucially, during evaluation, PMPO selects among variants by minimizing loss in a single forward pass, eliminating output sampling and human or judge based scoring for selection while still using standard generation only to propose rewrites. This unified, loss based strategy supports both supervised and preference based tasks. Across model sizes and datasets, PMPO outperforms prior prompt optimizers: it achieves the highest average accuracy on BBH, performs strongly on GSM8K and AQuA RAT, and raises AlpacaEval 2.0 win rates by over 19 points. These results demonstrate PMPO’s effectiveness, efficiency, and broad applicability.

Yinhong Liu, Jianfeng He, Hang Su, Ruixue Lian, Yi Nian, Jake W. Vincent, Srikanth Vishnubhotla, Robinson Piramuthu, Saab Mansour

Multimodal Dialogue Summarization (MDS) is a critical task with wide-ranging applications. To support the development of effective MDS models, robust automatic evaluation methods are essential for reducing both cost and human effort. However, such methods require a strong meta-evaluation benchmark grounded in human annotations. In this work, we introduce MDSEval, the first meta-evaluation benchmark for MDS, consisting image-sharing dialogues, corresponding summaries, and human judgments across eight well-defined quality aspects. To ensure data quality and richfulness, we propose a novel filtering framework leveraging Mutually Exclusive Key Information (MEKI) across modalities. Our work is the first to identify and formalize key evaluation dimensions specific to MDS. Finally, we benchmark state-of-the-art modal evaluation methods, revealing their limitations in distinguishing summaries from advanced MLLMs and their susceptibility to various bias.

Mohamed Imed Eddine Ghebriout, Gaël Guibon, Ivan Lerner, Emmanuel Vincent

Dialogue summarization aims to distill the core meaning of a conversation into a concise text. This is crucial for reducing the complexity and noise inherent in dialogue-heavy applications. While recent approaches typically train language models to mimic human-written summaries, such supervision is costly and often results in outputs that lack task-specific focus limiting their effectiveness in downstream applications, such as medical tasks. In this paper, we propose QUARTZ, a framework for task-oriented utility-based dialogue summarization. QUARTZ starts by generating multiple summaries and task-oriented question-answer pairs from a dialogue in a zero-shot manner using a pool of large language models (LLMs). The quality of the generated summaries is evaluated by having LLMs answer task-related questions before (i) selecting the best candidate answers and (ii) identifying the most informative summary based on these answers. Finally, we fine-tune the best LLM on the selected summaries. When validated on multiple datasets, QUARTZ demonstrates its effectiveness by achieving competitive results in various zero-shot settings, rivaling fully-supervised State-of-the-Art (SotA) methods. Code will be released publicly.

Adi Simhi, Itay Itzhak, Fazl Barez, Gabriel Stanovsky, Yonatan Belinkov

Prior work on large language model (LLM) hallucinations has associated them with model uncertainty or inaccurate knowledge. In this work, we define and investigate a distinct type of hallucination, where a model can consistently answer a question correctly, but a seemingly trivial perturbation, which can happen in real-world settings, causes it to produce a hallucinated response with high certainty. This phenomenon, which we dub CHOKE (Certain Hallucinations Overriding Known Evidence), is particularly concerning in high-stakes domains such as medicine or law, where model certainty is often used as a proxy for reliability. We show that CHOKE examples are consistent across prompts, occur in different models and datasets, and are fundamentally distinct from other hallucinations. This difference leads existing mitigation methods to perform worse on CHOKE examples than on general hallucinations. Finally, we introduce a probing-based mitigation that outperforms existing methods on CHOKE hallucinations. These findings reveal an overlooked aspect of hallucinations, emphasizing the need to understand their origins and improve mitigation strategies to enhance LLM safety.

Yusuf Sali, Sıtkı Can Toraman

Out-of-Distribution (OOD) detection is a challenging task that requires great generalization capability for the practicality and safety of task-oriented dialogue systems (TODS). With the dawn of large language models (LLMs), their enhanced ability to handle diverse patterns and contexts may aid in addressing this challenging task. In this paper, we investigate the current performance of LLMs in the near-OOD setting, where OOD queries belong to the same domain but different intents. To take advantage of out-of-the-shelf capabilities of LLMs, we do not use fine-tuning. We study the performance of one of the leading frontier models, GPT-4o, in 3 well-known public datasets and 3 in-house datasets, using 10 different methods and prompt variations. We study the performance of different prompts and techniques in Gemini 1.5 Flash and Llama 3.1-70b. We investigate the effect of increasing the number of In-Distribution (ID) intents. We propose a novel hybrid method that is cost-efficient, high-performing, highly robust, and versatile enough to be used with smaller LLMs without sacrificing performance. This is achieved by combining ID success of smaller text classification models and high generalization capabilities of LLMs in OOD detection.

Yerin Hwang, Dongryeol Lee, Taegwan Kang, Yongil Kim, Kyomin Jung

As large language models (LLMs) take on growing roles as automated evaluators in practical settings, a critical question arises: Can individuals persuade an LLM judge to assign unfairly high scores? This study is the first to reveal that strategically embedded persuasive language can bias LLM judges when scoring mathematical reasoning tasks, where correctness should be independent of stylistic variation. Grounded in Aristotle’s rhetorical principles, we formalize seven persuasion techniques (Majority, Consistency, Flattery, Reciprocity, Pity, Authority, Identity) and embed them into otherwise identical responses. Across six math benchmarks, we find that persuasive language leads LLM judges to assign inflated scores to incorrect solutions, by up to 8% on average, with Consistency causing the most severe distortion. Notably, increasing model size does not substantially mitigate this vulnerability. Further analysis demonstrates that combining multiple persuasion techniques amplifies the bias, and pairwise evaluation is likewise susceptible. Moreover, the persuasive effect persists under counter-prompting strategies, highlighting a critical vulnerability in LLM-as-a-Judge pipelines and underscoring the need for robust defenses against persuasion-based attacks.

Hiba Ahsan, Arnab Sen Sharma, Silvio Amir, David Bau, Byron C Wallace

We know from prior work that LLMs encode social biases, and that this manifests in clinical tasks. In this work we adopt tools from mechanistic interpretability to unveil sociodemographic representations and biases within LLMs in the context of healthcare. Specifically, we ask: Can we identify activations within LLMs that encode sociodemographic information (e.g., gender, race)? We find that, in three open weight LLMs, gender information is highly localized in MLP layers and can be reliably manipulated at inference time via patching. Such interventions can surgically alter generated clinical vignettes for specific conditions, and also influence downstream clinical predictions which correlate with gender, e.g., patient risk of depression. We find that representation of patient race is somewhat more distributed, but can also be intervened upon, to a degree. To our knowledge, this is the first application of mechanistic interpretability methods to LLMs for healthcare.

Wataru Hashimoto, Hidetaka Kamigaito, Taro Watanabe

Decoding strategies manipulate the probability distribution underlying the output of a language model and can therefore affect both generation quality and its uncertainty. In this study, we investigate the impact of decoding strategies on uncertainty estimation in Large Language Models (LLMs). Our experiments show that Contrastive Search, which mitigates repetition, yields better uncertainty estimates on average across a range of preference-aligned LLMs. In contrast, the benefits of these strategies sometimes diverge when the model is only post-trained with supervised fine-tuning, i.e. without explicit alignment.

Ruihan Jin, Pengpeng Shao, Zhengqi Wen, Jinyang Wu, Mingkuan Feng, Shuai Zhang, Jianhua Tao

The rapid advancements in large language models (LLMs) have led to the emergence of routing techniques, which aim to efficiently select the optimal LLM from diverse candidates to tackle specific tasks, optimizing performance while reducing costs. Current LLM routing methods are limited in effectiveness due to insufficient exploration of the intrinsic connection between user queries and the characteristics of LLMs. To address this issue, in this paper, we present **RadialRouter**, a novel framework for LLM routing which employs a lightweight Transformer-based backbone with a radial structure named **RadialFormer** to articulate the query-LLMs relationship. The optimal LLM selection is performed based on the final states of RadialFormer. The pipeline is further refined by an objective function that combines Kullback-Leibler divergence with the query-query contrastive loss to enhance robustness. Experimental results on RouterBench show that RadialRouter significantly outperforms existing routing methods by 9.2% and 5.8% in the *Balance* and *Cost First* scenarios, respectively. Additionally, its adaptability toward different performance-cost trade-offs and the dynamic LLM pool demonstrates practical application potential.

Yuchen Wu, Liang Ding, Li Shen, Dacheng Tao

Large language models (LLMs) encode vast amounts of world knowledge but remain static once trained, making timely integration of emerging facts prohibitively expensive via full retraining. Knowledge-editing techniques have thus emerged to inject or overwrite specific facts into LLMs, yet they either over-rely on superficial cues or incur complex, iterative pipelines that collapse under noisy, multi-hop conditions. We introduce **Reason-KE**, an end-to-end reasoning-chain-based editing framework that steers a pretrained LLM through four structured stages—fact acknowledgment, relevance determination, selective application, and final reasoning—to filter distractors in a single pass. Trained on MQuAKE-CF with up to four irrelevant facts, Reason-KE elevates Qwen2.5-7B’s multi-hop QA accuracy to 90.2% (↑17.6 pp) while suffering merely 6.3% drop under heavy distraction and <1% when answers are leaked. Our quantitative analysis confirms Reason-KE’s resilience and efficiency, establishing a new state of the art for reliable LLM knowledge updates. The code will be released.