论文检索

Yongan Yu, Qingchen Hu, Xianda Du, Jiayin Wang, Fengran Mo, Renée Sieber

Climate change adaptation requires the understanding of disruptive weather impacts on society, where large language models (LLMs) might be applicable. However, their effectiveness is under-explored due to the difficulty of high-quality corpus collection and the lack of available benchmarks. The climate-related events stored in regional newspapers record how communities adapted and recovered from disasters. However, the processing of the original corpus is non-trivial. In this study, we first develop a disruptive weather impact dataset with a four-stage well-crafted construction pipeline. Then, we propose WXImpactBench, the first benchmark for evaluating the capacity of LLMs on disruptive weather impacts. The benchmark involves two evaluation tasks, multi-label classification and ranking-based question answering. Extensive experiments on evaluating a set of LLMs provide first-hand analysis of the challenges in developing disruptive weather impact understanding and climate change adaptation systems. The constructed dataset and the code for the evaluation framework are available to help society protect against vulnerabilities from disasters.

Zihan Liu, Yang Chen, Mohammad Shoeybi, Bryan Catanzaro, Wei Ping

In this paper, we introduce AceMath, a suite of frontier math models that excel in solving complex math problems, along with highly effective reward models capable of evaluating generated solutions and reliably identifying the correct ones. To develop the instruction-tuned math models, we propose a supervised fine-tuning (SFT) process that first achieves competitive performance across general domains, followed by targeted fine-tuning for the math domain using a carefully curated set of prompts and synthetically generated responses. The resulting model, AceMath-72B-Instruct greatly outperforms Qwen2.5-Math-72B-Instruct, GPT-4o and Claude-3.5 Sonnet. To develop math-specialized reward model, we first construct AceMath-RewardBench, a comprehensive and robust benchmark for evaluating math reward models across diverse problems and difficulty levels. After that, we present a systematic approach to build our math reward models. The resulting model, AceMath-72B-RM, consistently outperforms state-of-the-art reward models. Furthermore, when combining AceMath-72B-Instruct with AceMath-72B-RM, we achieve the highest average rm@8 score across the math reasoning benchmarks.

Sohee Yang, Nora Kassner, Elena Gribovskaya, Sebastian Riedel, Mor Geva

We evaluate how well Large Language Models (LLMs) latently recall and compose facts to answer multi-hop queries like “In the year Scarlett Johansson was born, the Summer Olympics were hosted in the country of”. One major challenge in such evaluation is that LLMs may have developed shortcuts by encountering the head entity “Scarlett Johansson” and the answer entity “United States” in the same training sequences or merely guess the answer based on frequency-based priors. To prevent shortcuts, we exclude test queries where the head and answer entities might have co-appeared during training. Through careful selection of relations and facts and systematic removal of cases where models might guess answers or exploit partial matches, we construct an evaluation dataset SOCRATES (ShOrtCut-fRee lATent rEaSoning). We observe that LLMs demonstrate promising latent multi-hop reasoning abilities without exploiting shortcuts, but only for certain types of queries. For queries requiring latent recall of countries as the intermediate answer, the best models achieve 80% latent composability, but this drops to just 5% for the recall of years. Comparisons with Chain-of-Thought highlight a significant gap between the ability of models to reason latently versus explicitly. Analysis reveals that latent representations of the intermediate answer are constructed more often in queries with higher latent composability, and shows the emergence of latent multi-hop reasoning during pretraining.

Hayden Helm, Aranyak Acharyya, Youngser Park, Brandon Duderstadt, Carey Priebe

Generative models are capable of producing human-expert level content across a variety of topics and domains. As the impact of generative models grows, it is necessary to develop statistical methods to understand collections of available models. These methods are particularly important in settings where the user may not have access to information related to a model’s pre-training data, weights, or other relevant model-level covariates. In this paper we extend recent results on representations of black-box generative models to model-level statistical inference tasks. We demonstrate that the model-level representations are effective for multiple inference tasks.

Prafulla Kumar Choubey, Xiangyu Peng, Shilpa Bhagavath, Caiming Xiong, Shiva Kumar Pentyala, Chien-Sheng Wu

Automated service agents require well-structured workflows to deliver consistent and accurate responses to customer queries. However, such workflows are often undocumented, and their automatic extraction from conversations remains largely unexplored. In this work, we present a novel framework for extracting and evaluating dialog workflows from historical interactions. Our extraction process involves two key stages: (1) a retrieval step to select relevant conversations based on key procedural elements, and (2) a structured workflow generation step using question-answer-based chain-of-thought (QA-CoT) prompting. To comprehensively evaluate the quality of the extracted workflows, we introduce an automated simulation framework with agent and customer bots that measures their effectiveness in resolving customer issues. Extensive experiments on the ABCD and SynthABCD datasets show that our QA-CoT technique improves workflow extraction by 12.16% in average macro accuracy over the baseline. Moreover, our evaluation method closely aligns with human assessments, offering a reliable and scalable framework for future research.

Emily Corvi, Hannah Washington, Stefanie Reed, Chad Atalla, Alexandra Chouldechova, P. Alex Dow, Jean Garcia-Gathright, Nicholas J Pangakis, Emily Sheng, Dan Vann 等

Representational harms are widely recognized among fairness-related harms caused by generative language systems. However, their definitions are commonly under-specified. We make a theoretical contribution to the specification of representational harms by introducing a framework, grounded in speech act theory (Austin 1962), that conceptualizes representational harms caused by generative language systems as the perlocutionary effects (i.e., real-world impacts) of particular types of illocutionary acts (i.e., system behaviors). Building on this argument and drawing on relevant literature from linguistic anthropology and sociolinguistics, we provide new definitions of stereotyping, demeaning, and erasure. We then use our framework to develop a granular taxonomy of illocutionary acts that cause representational harms, going beyond the high-level taxonomies presented in previous work. We also discuss the ways that our framework and taxonomy can support the development of valid measurement instruments. Finally, we demonstrate the utility of our framework and taxonomy via a case study that engages with recent conceptual debates about what constitutes a representational harm and how such harms should be measured.

Jinghan Zhang, Xiting Wang, Fengran Mo, Yeyang Zhou, Wanfu Gao, Kunpeng Liu

Multi-step processes via large language models (LLMs) have proven effective for solving complex reasoning tasks. However, the depth of exploration of the reasoning procedure can significantly affect the task performance. Existing methods to automatically decide the depth often lead to high cost and a lack of flexibility. To address these issues, we propose Entropy-based Exploration Depth Conduction (Entro-duction), a novel method that dynamically adjusts the exploration depth during multi-step reasoning by monitoring LLM’s output entropy and variance entropy. We employ these two features to capture the model’s uncertainty of the current step and the fluctuation of uncertainty across consecutive reasoning steps. Based on the observed entropy changes, the LLM selects whether to deepen, expand, or stop exploration according to the probability, which facilitates the trade-off between the reasoning accuracy and exploration effectiveness. Experimental results across four benchmark datasets demonstrate the efficacy of Entro-duction.

Yingtai Xiao, Yuqing Zhu, Sirat Samyoun, Wanrong Zhang, Jiachen T. Wang, Jian Du

Large language models (LLMs) demonstrate strong capabilities in in-context learning, but verifying the correctness of their generated responses remains a challenge. Prior work has explored attribution at the sentence level, but these methods fall short when users seek attribution for specific keywords within the response, such as numbers, years, or names. To address this limitation, we propose TokenShapley, a novel token-level attribution method that combines Shapley value-based data attribution with KNN-based retrieval techniques inspired by recent advances in KNN-augmented LLMs. By leveraging a precomputed datastore for contextual retrieval and computing Shapley values to quantify token importance, TokenShapley provides a fine-grained data attribution approach. Extensive evaluations on four benchmarks show that TokenShapley outperforms state-of-the-art baselines in token-level attribution, achieving a 11–23% improvement in accuracy.

Xuanliang Zhang, Dingzirui Wang, Baoxin Wang, Longxu Dou, Xinyuan Lu, Keyan Xu, Dayong Wu, Qingfu Zhu

Scientific question answering (SQA) is an important task aimed at answering questions based on papers. However, current SQA datasets have limited reasoning types and neglect the relevance between tables and text, creating a significant gap with real scenarios. To address these challenges, we propose a QA benchmark for scientific tables and text with diverse reasoning types (SCITAT). To cover more reasoning types, we summarize various reasoning types from real-world questions. To reason on both tables and text, we require the questions to incorporate tables and text as much as possible. Based on SCITAT, we propose a baseline (CAR), which combines various reasoning methods to address different reasoning types and process tables and text at the same time. CAR brings average improvements of 4.1% over other baselines on SCITAT, validating its effectiveness. Error analysis reveals the challenges of SCITAT, such as complex numerical calculations and domain knowledge.

Sharath Naganna, Saprativa Bhattacharjee, Biplab Banerjee, Pushpak Bhattacharyya

Humblebragging is a phenomenon in which individuals present self-promotional statements under the guise of modesty or complaints. For example, a statement like, “Ugh, I can’t believe I got promoted to lead the entire team. So stressful!”, subtly highlights an achievement while pretending to be complaining. Detecting humblebragging is important for machines to better understand the nuances of human language, especially in tasks like sentiment analysis and intent recognition. However, this topic has not yet been studied in computational linguistics. For the first time, we introduce the task of automatically detecting humblebragging in text. We formalize the task by proposing a 4-tuple definition of humblebragging and evaluate machine learning, deep learning, and large language models (LLMs) on this task, comparing their performance with humans. We also create and release a dataset called HB-24, containing 3,340 humblebrags generated using GPT-4o. Our experiments show that detecting humblebragging is non-trivial, even for humans. Our best model achieves an F1-score of 0.88. This work lays the foundation for further exploration of this nuanced linguistic phenomenon and its integration into broader natural language understanding systems.

Yihang Cheng, Lan Zhang, Junyang Wang, Mu Yuan, Yunhao Yao

Retrieval-augmented generation (RAG) improves the service quality of large language models by retrieving relevant documents from credible literature and integrating them into the context of the user query.Recently, the rise of the cloud RAG service has made it possible for users to query relevant documents conveniently.However, directly sending queries to the cloud brings potential privacy leakage.In this paper, we are the first to formally define the privacy-preserving cloud RAG service to protect the user query and propose RemoteRAG as a solution regarding privacy, efficiency, and accuracy.For privacy, we introduce (n,\epsilon)-DistanceDP to characterize privacy leakage of the user query and the leakage inferred from relevant documents.For efficiency, we limit the search range from the total documents to a small number of selected documents related to a perturbed embedding generated from (n,\epsilon)-DistanceDP, so that computation and communication costs required for privacy protection significantly decrease.For accuracy, we ensure that the small range includes target documents related to the user query with detailed theoretical analysis.Experimental results also demonstrate that RemoteRAG can resist existing embedding inversion attack methods while achieving no loss in retrieval under various settings.Moreover, RemoteRAG is efficient, incurring only 0.67 seconds and 46.66KB of data transmission (2.72 hours and 1.43 GB with the non-optimized privacy-preserving scheme) when retrieving from a total of 10^5 documents.

Mufan Qiu, Xinyu Hu, Fengwei Zhan, Sukwon Yun, Jie Peng, Ruichen Zhang, Bhavya Kailkhura, Jiekun Yang, Tianlong Chen

Foundation models for single-cell RNA sequencing (scRNA-seq) have shown promising capabilities in capturing gene expression patterns. However, current approaches face critical limitations: they ignore biological prior knowledge encoded in gene regulatory relationships and fail to leverage multi-omics signals that could provide complementary regulatory insights. In this paper, we propose GRNFormer, a new framework that systematically integrates multi-scale Gene Regulatory Networks (GRNs) inferred from multi-omics data into RNA foundation model training. Our framework introduces two key innovations. First, we introduce a pipeline for constructing hierarchical GRNs that capture regulatory relationships at both cell-type-specific and cell-specific resolutions. Second, we design a structure-aware integration framework that addresses the information asymmetry in GRNs through two technical advances: (1) A graph topological adapter using multi-head cross-attention to weight regulatory relationships dynamically, and (2) a novel edge perturbation strategy that perturb GRNs with biologically-informed co-expression links to augment graph neural network training. Comprehensive experiments have been conducted on three representative downstream tasks across multiple model architectures to demonstrate the effectiveness of GRNFormer. It achieves consistent improvements over state-of-the-art (SoTA) baselines: \mathbf{3.6\\\%} increase in drug response prediction correlation, \mathbf{9.6\\\%} improvement in single-cell drug classification AUC, and \mathbf{1.1\\\%} average gain in gene perturbation prediction accuracy.

Yue Wan, Xiaowei Jia, Xiang Lorraine Li

Chain-of-thought (CoT) prompting has been widely adopted to enhance the reasoning capabilities of large language models (LLMs). However, the effectiveness of CoT reasoning is inconsistent across tasks with different reasoning types. This work presents a novel perspective to understand CoT behavior through the lens of confirmation bias in cognitive psychology. Specifically, we examine how model internal beliefs, approximated by direct question-answering probabilities, affect both reasoning generation (Q \to R) and reasoning-guided answer prediction (QR \to A) in CoT. By decomposing CoT into a two-stage process, we conduct a thorough correlation analysis in model beliefs, rationale attributes, and stage-wise performance. Our results provide strong evidence of confirmation bias in LLMs, such that model beliefs not only skew the reasoning process but also influence how rationales are utilized for answer prediction. Furthermore, the interplay between task vulnerability to confirmation bias and the strength of beliefs also provides explanations for CoT effectiveness across reasoning tasks and models. Overall, this study provides a valuable insight for the needs of better prompting strategies that mitigate confirmation bias to enhance reasoning performance. Code is available at https://github.com/yuewan2/biasedcot.

Janek Bevendorff, Matti Wiegmann, Emmelie Richter, Martin Potthast, Benno Stein

The detection of texts generated by LLMs has quickly become an important research problem. Many supervised and zero-shot detectors have already been proposed, yet their effectiveness and precision remain disputed. Current research therefore focuses on making detectors robust against domain shifts and on building corresponding benchmarks. In this paper, we show that the actual limitations hindering progress in LLM detection lie elsewhere: LLM detection is often implicitly modeled as an authorship attribution task, while its true nature is that of authorship verification. We systematically analyze the current research with respect to this misunderstanding, conduct an in-depth comparative analysis of the benchmarks, and validate our claim using state-of-the-art LLM detectors. Our contributions open the realm of authorship analysis technology for understanding and tackling the problem of LLM detection.

Haoran Wang, Zhenyu Hou, Yao Wei, Jie Tang, Yuxiao Dong

Large language models (LLMs) have advanced rapidly from conversational problem solving to addressing real-world tasks involving tool use, such as software engineering (SWE). Recent LLM-powered toolkits, such as OpenAI Codex and Cursor, have offered end-to-end automation of the software development process. However, building effective SWE agents remains challenging due to the lack of high-quality training data and effective test cases. To address this issue, we present SWE-Dev, an SWE agent built upon open-source LLMs. First, we develop a robust pipeline to synthesize test cases for patch evaluation. Second, we scale up agent trajectories to construct the training data for building SWE-Dev. Experiments on the SWE-bench-Verified benchmark show that the SWE-Dev models can achieve top performance among all open SWE agents. Specifically, the success rates of the SWE-Dev 7B and 32B parameter models reach 23.4% and 36.6%, respectively, outperforming state-of-the-art open-source models. All code, models, and datasets are publicly available at https://github.com/THUDM/SWE-Dev.

Debela Gemechu, Ramon Ruiz-Dolz, Henrike Beyer, Chris Reed

While Large Language Models (LLMs) have demonstrated promising results on a range of reasoning benchmarks—particularly in formal logic, mathematical tasks, and Chain-of-Thought prompting—less is known about their capabilities in unconstrained natural language reasoning. Argumentative reasoning, a form of reasoning naturally expressed in language and central to everyday discourse, presents unique challenges for LLMs due to its reliance on context, implicit assumptions, and value judgments. This paper addresses a gap in the study of reasoning in LLMs by presenting the first large-scale evaluation of their unconstrained natural language reasoning capabilities based on natural language argumentation. The paper offers three contributions: (i) the formalisation of a new strategy designed to evaluate argumentative reasoning in LLMs: argument-component selection; (ii) the creation of the Argument Reasoning Tasks (ART) dataset, a new benchmark for argument-component selection based on argument structures for natural language reasoning; and (iii) an extensive experimental analysis involving four different models, demonstrating the limitations of LLMs on natural language reasoning tasks.

Yuhui Zhang, Yuchang Su, Yiming Liu, Serena Yeung-Levy

Negation is a fundamental linguistic phenomenon that can entirely reverse the meaning of a sentence. As vision language models (VLMs) continue to advance and are deployed in high-stakes applications, assessing their ability to comprehend negation becomes essential. To address this, we introduce NegVQA, a visual question answering (VQA) benchmark consisting of 7,379 two-choice questions covering diverse negation scenarios and image-question distributions. We construct NegVQA by leveraging large language models to generate negated versions of questions from existing VQA datasets. Evaluating 20 state-of-the-art VLMs across seven model families, we find that these models struggle significantly with negation, exhibiting a substantial performance drop compared to their responses to the original questions. Furthermore, we uncover a U-shaped scaling trend, where increasing model size initially degrades performance on NegVQA before leading to improvements. Our benchmark reveals critical gaps in VLMs’ negation understanding and offers insights into future VLM development. Project page available at https://yuhui-zh15.github.io/NegVQA/.

Mert Inan, Anthony Sicilia, Malihe Alikhani

Deaf and Hard-of-Hearing (DHH) users increasingly utilize Large Language Models (LLMs), yet face significant challenges due to these models’ limited understanding of sign language grammar, multimodal sign inputs, and Deaf cultural contexts. Further, current approaches that try to address these limitations, frequently reduce sign language processing (SLP) to traditional translation tasks, neglecting the multimodal and linguistic complexity inherent in signed languages. In this paper, we present an empirical investigation informed by learning theory into natively integrating sign language support within LLMs, directly addressing the documented needs of DHH users. We introduce the first text-based and multimodal LLMs capable of sign language processing called SignAlignLM, and propose new prompting and fine-tuning strategies incorporating sign linguistic rules and conventions. We show that LLMs can be generalized interfaces for both spoken and signed languages if trained with a multitasking paradigm. Our code and model checkpoints are open-source.

Yuting Huang, Chengyuan Liu, Yifeng Feng, Yiquan Wu, Chao Wu, Fei Wu, Kun Kuang

As Large Language Models (LLMs) are widely applied in various domains, the safety of LLMs is increasingly attracting attention to avoid their powerful capabilities being misused. Existing jailbreak methods create a forced instruction-following scenario, or search adversarial prompts with prefix or suffix tokens to achieve a specific representation manually or automatically. However, they suffer from low efficiency and explicit jailbreak patterns, far from the real deployment of mass attacks to LLMs. In this paper, we point out that simply rewriting the original instruction can achieve a jailbreak, and we find that this rewriting approach is learnable and transferable. We propose the **R**ewrite to **J**ailbreak (R2J) approach, a transferable black-box jailbreak method to attack LLMs by iteratively exploring the weakness of the LLMs and automatically improving the attacking strategy. The jailbreak is more efficient and hard to identify since no additional features are introduced. Extensive experiments and analysis demonstrate the effectiveness of R2J, and we find that the jailbreak is also transferable to multiple datasets and various types of models with only a few queries. We hope our work motivates further investigation of LLM safety. The code can be found at [https://github.com/ythuang02/R2J/.](https://github.com/ythuang02/R2J/)

Weijie Shi, Hao Chen, Jiaming Li, Yao Zhao, Yazhong Zhang, Qijin Chen, Jipeng Zhang, Ruiyuan Zhang, Jia Zhu, Jiajie Xu 等

Retrieval-augmented language models (RALMs) aim to incorporate external knowledge to address the issues of factual hallucination and knowledge obsolescence faced by large language models (LLMs). Inevitably, the retrieved passages based on similarity search may be irrelevant to the given question, and the aggregation of these passages can confuse the model to give a correct answer. To improve the performance of RALM in such conditions, we propose layer-knowledge guided attention for RALMs, which harnesses the layer-wise knowledge of LLMs to optimize per-layer attention on useful passages, making the model pay attention to the most relevant content and ignore irrelevant ones. Specifically, we first systematically study LLM’s attention patterns and their relationship with the accuracy of RALM responses, where middle-focus attentions play a crucial role in selectively gathering relevant information. Based on this, a layer-wise passage estimator leverages the varied knowledge encoded across LLM layers to assess not only passage relevance scores but also associated confidences. Finally, a relevance-aware passage fusion enables selective attention to relevant passages, mitigating distractibility and positional bias of causal attention. Experiments show that our method outperforms existing methods on RALM benchmarks.