Large language model (LLM) agents have demonstrated remarkable potential in advancing scientific discovery. However, their capability in the fundamental yet crucial task of reproducing code from research papers, especially in the NLP domain, remains underexplored. This task includes unique complex reasoning challenges in the intellectual synthesis of abstract concepts and the comprehension of code repositories with interdependent files. Motivated by this gap, we present LMR-BENCH, a benchmark designed to systematically evaluate the capability of LLM agents on code reproduction from Language Modeling Research. It consists of 28 code reproduction tasks derived from 23 research papers published in top-tier NLP venues over the past five years, spanning nine fundamental categories. Models are provided with a research paper, a code repository containing one or more masked functions, and instructions for implementing these functions. We conduct extensive experiments in standard prompting and LLM agent settings with state-of-the-art LLMs, evaluating the accuracy of unit tests and performing LLM-based evaluation of code correctness. Experimental results reveal that even the most advanced models still exhibit persistent limitations in scientific reasoning and code synthesis, highlighting critical gaps in LLM agents’ ability to autonomously reproduce scientific research.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
This paper investigates how LLMs encode inputs with typos. We hypothesize that specific neurons and attention heads recognize typos and fix them internally using local and global contexts. We introduce a method to identify typo neurons and typo heads that work actively when inputs contain typos. Our experimental results suggest the following: 1) LLMs can fix typos with local contexts when the typo neurons in either the early or late layers are activated, even if those in the other are not. 2) Typo neurons in the middle layers are the core of typo-fixing with global contexts. 3) Typo heads fix typos by widely considering the context not focusing on specific tokens. 4) Typo neurons and typo heads work not only for typo-fixing but also for understanding general contexts.
The emergence of Multimodal Large Reasoning Models (MLRMs) has enabled sophisticated visual reasoning capabilities by integrating reinforcement learning and Chain-of-Thought (CoT) supervision. However, while these enhanced reasoning capabilities improve performance, they also introduce new and underexplored safety risks. In this work, we systematically investigate the security implications of advanced visual reasoning in MLRMs. Our analysis reveals a fundamental trade-off: as visual reasoning improves, models become more vulnerable to jailbreak attacks. Motivated by this critical finding, we introduce VisCRA (Visual Chain Reasoning Attack), a novel jailbreak framework that exploits the visual reasoning chains to bypass safety mechanisms. VisCRA combines targeted visual attention masking with a two-stage reasoning induction strategy to precisely control harmful outputs. Extensive experiments demonstrate VisCRA’s significant effectiveness, achieving high attack success rates on leading closed-source MLRMs: 76.48% on Gemini 2.0 Flash Thinking, 68.56% on QvQ-Max, and 56.60% on GPT-4o. Our findings highlight a critical insight: the very capability that empowers MLRMs — their visual reasoning — can also serve as an attack vector, posing significant security risks. Warning: This paper contains unsafe examples.
MAKAR: a Multi-Agent framework based Knowledge-Augmented Reasoning for Grounded Multimodal Named Entity Recognition
PDF ↗Grounded Multimodal Named Entity Recognition (GMNER), which aims to extract textual entities, their types, and corresponding visual regions from image-text data, has become a critical task in multimodal information extraction. However, existing methods face two major challenges. First, they fail to address the semantic ambiguity caused by polysemy and the long-tail distribution of datasets. Second, unlike visual grounding which provides descriptive phrases, entity grounding only offers brief entity names which carry less semantic information. Current methods lack sufficient semantic interaction between text and image, hindering accurate entity-visual region matching. To tackle these issues, we propose MAKAR, a Multi-Agent framework based Knowledge-Augmented Reasoning, comprising three agents: Knowledge Enhancement, Entity Correction, and Entity Reasoning Grounding. Specifically, in the named entity recognition phase, the Knowledge Enhancement Agent leverages a Multimodal Large Language Model (MLLM) as an implicit knowledge base to enhance ambiguous image-text content with its internal knowledge. For samples with low-confidence entity boundaries and types, the Entity Correction Agent uses web search tools to retrieve and summarize relevant web content, thereby correcting entities using both internal and external knowledge. In the entity grounding phase, the Entity Reasoning Grounding Agent utilizes multi-step Chain-of-Thought reasoning to perform grounding for each entity. Extensive experiments show that MAKAR achieves state-of-the-art performance on two benchmark datasets. Code is available at: https://github.com/Nikol-coder/MAKAR.
SciRIFF: A Resource to Enhance Language Model Instruction-Following over Scientific Literature
PDF ↗We present ScIRIFF (Scientific Resource for Instruction-Following and Finetuning), a dataset of 137K instruction-following instances for training and evaluation, covering 54 tasks. These tasks span five core scientific literature understanding capabilities: information extraction, summarization, question answering, claim verification, and classification. ScIRIFF is unique in being the only entirely expert-written, high-quality instruction-following dataset designed for extracting and synthesizing information from research literature across diverse scientific fields. It features complex instructions with long input contexts, detailed task descriptions, and structured outputs. To demonstrate its utility, we finetune a series of large language models (LLMs) using a mix of general domain and ScIRIFF instructions. On nine out-of-distribution held-out tasks (referred to as SciRIFF-Eval), LLMs finetuned on SciRIFF achieve 70.6% average improvement over our baselines trained only on general-domain instructions. ScIRIFF facilitates the development and evaluation of LLMs to help researchers navigate the rapidly growing body of scientific literature.
Currently, Large Language Models (LLMs) have achieved remarkable results in machine translation. However, their performance in multi-domain translation (MDT) is less satisfactory, the meanings of words can vary across different domains, highlighting the significant ambiguity inherent in MDT. Therefore, evaluating the disambiguation ability of LLMs in MDT, remains an open problem. To this end, we present an evaluation and analysis of LLMs on disambiguation in multi-domain translation (DMDTEval), our systematic evaluation framework consisting of three aspects: (1) we construct a translation test set with multi-domain ambiguous word annotation, (2) we curate a diverse set of disambiguation prompt strategies, and (3) we design precise disambiguation metrics, and study the efficacy of various prompt strategies on multiple state-of-the-art LLMs. We conduct comprehensive experiments across 4 language pairs and 13 domains, our extensive experiments reveal a number of crucial findings that we believe will pave the way and also facilitate further research in the critical area of improving the disambiguation of LLMs.
Text-to-SQL, which maps natural language to SQL queries, has benefited greatly from recent advances in Large Language Models (LLMs). While LLMs offer various paradigms for this task, including prompting and supervised fine-tuning (SFT), SFT approaches still face challenges such as complex multi-stage pipelines and poor robustness to noisy schema information. To address these limitations, we present JOLT-SQL, a streamlined single-stage SFT framework that jointly optimizes schema linking and SQL generation via a unified loss. JOLT-SQL employs discriminative schema linking, enhanced by local bidirectional attention, alongside a confusion-aware noisy schema sampling strategy with selective attention to improve robustness under noisy schema conditions. Experiments on the Spider and BIRD benchmarks demonstrate that JOLT-SQL achieves state-of-the-art execution accuracy among comparable-size open-source models, while significantly improving both training and inference efficiency.
This paper introduces ThoughtProbe, a novel inference-time framework that leverages the hidden reasoning features of Large Language Models (LLMs) to improve their reasoning performance. Unlike previous works that manipulate the hidden representations to steer LLM generation, we harness them as discriminative signals to guide the tree-structured response space exploration. In each node expansion, a classifier serves as a scoring and ranking mechanism that efficiently allocates computational resources by prioritizing higher score candidates for continuation. After completing the tree expansion, we collect answers from all branches to form a candidate answer pool. We then propose a branch-aggregation method that marginalizes over all supporting branches by aggregating their CoT scores, thereby identifying the optimal answer from the pool. Experimental results show that our framework’s comprehensive exploration not only covers valid reasoning chains but also effectively identifies them, achieving significant improvements across multiple arithmetic reasoning benchmarks.
The empathy dialogue system requires understanding emotions and their underlying causes. However, existing datasets mainly focus on emotion labels, while cause annotations are added post hoc through costly and subjective manual processes. This leads to three limitations: subjective bias in cause labels, weak rationality due to ambiguous cause-emotion relationships, and high annotation costs that hinder scalability. To address these challenges, we propose ECC (Emotion-Cause Conversation Dataset), a scalable dataset with 2.4K dialogues, which is also the first dialogue dataset where conversations and their emotion-cause labels are automatically generated synergistically during creation. We create an automatic extension framework EC-DD for ECC that utilizes knowledge and large language models (LLMs) to automatically generate conversations, and train a causality-aware empathetic response model CAER on this dataset. Experimental results show that ECC can achieve comparable or even superior performance to artificially constructed empathy dialogue datasets. Our code will be publicly released on https://github.com/Yuan-23/ECC
Visual Question Generation (VQG) research focuses predominantly on natural images while neglecting the diagram, which is a critical component in educational materials. To meet the needs of pedagogical assessment, we propose the Diagram-Driven Course Questions Generation (DDCQG) task and construct DiagramQG, a comprehensive dataset with 15,720 diagrams and 25,798 questions across 37 subjects and 371 courses. Our approach employs course and input text constraints to generate course-relevant questions about specific diagram elements. We reveal three challenges of DDCQG: domain-specific knowledge requirements across courses, long-tail distribution in course coverage, and high information density in diagrams. To address these, we propose the Hierarchical Knowledge Integration framework (HKI-DDCQG), which utilizes trainable CLIP for identifying relevant diagram patches, leverages frozen vision-language models for knowledge extraction, and generates questions with trainable T5. Experiments demonstrate that HKI-DDCQG outperforms existing models on DiagramQG while maintaining strong generalizability across natural image datasets, establishing a strong baseline for DDCQG.
The rapid growth of text-to-video (T2V) diffusion models has raised concerns about privacy, copyright, and safety due to their potential misuse in generating harmful or misleading content. These models are often trained on numerous datasets, including unauthorized personal identities, artistic creations, and harmful materials, which can lead to uncontrolled production and distribution of such content. To address this, we propose VideoEraser, a training-free framework that prevents T2V diffusion models from generating videos with undesirable concepts, even when explicitly prompted with those concepts. Designed as a plug-and-play module, VideoEraser can seamlessly integrate with representative T2V diffusion models via a two-stage process: Selective Prompt Embedding Adjustment (SPEA) and Adversarial-Resilient Noise Guidance (ARNG). We conduct extensive evaluations across four tasks, including object erasure, artistic style erasure, celebrity erasure, and explicit content erasure. Experimental results show that VideoEraser consistently outperforms prior methods regarding efficacy, integrity, fidelity, robustness, and generalizability. Notably, VideoEraser achieves state-of-the-art performance in suppressing undesirable content during T2V generation, reducing it by 46% on average across four tasks compared to baselines.
QualBench: Benchmarking Chinese LLMs with Localized Professional Qualifications for Vertical Domain Evaluation
PDF ↗The rapid advancement of Chinese LLMs underscores the need for vertical-domain evaluations to ensure reliable applications. However, existing benchmarks often lack domain coverage and provide limited insights into the Chinese working context. Leveraging qualification exams as a unified framework for expertise evaluation, we introduce \textbf{QualBench}, the first multi-domain Chinese QA benchmark dedicated to localized assessment of Chinese LLMs. The dataset includes over 17,000 questions across six vertical domains, drawn from 24 Chinese qualifications to align with national policies and professional standards. Results reveal an interesting pattern of Chinese LLMs consistently surpassing non-Chinese models, with the Qwen2.5 model outperforming the more advanced GPT-4o, emphasizing the value of localized domain knowledge in meeting qualification requirements. The average accuracy of 53.98% reveals the current gaps in domain coverage within model capabilities. Furthermore, we identify performance degradation caused by LLM crowdsourcing, assess data contamination, and illustrate the effectiveness of prompt engineering and model fine-tuning, suggesting opportunities for future improvements through multi-domain RAG and Federated Learning. Data and code are publicly available at https://github.com/mengze-hong/QualBench.
CARFT: Boosting LLM Reasoning via Contrastive Learning with Annotated Chain-of-Thought-based Reinforced Fine-Tuning
PDF ↗Reasoning capability plays a significantly critical role in the the broad applications of Large Language Models (LLMs). To enhance the reasoning performance of LLMs, diverse Reinforcement Learning (RL)-based fine-tuning approaches have been proposed to address the limited generalization capability of LLMs trained solely via Supervised Fine-Tuning (SFT). Despite their effectiveness, two major limitations hinder the advancement of LLMs. First, vanilla RL-based approaches ignore annotated Chain-of-Thought (CoT) and incorporate unstable reasoning path sampling, which typically results in model collapse, unstable training process, and suboptimal performance. Second, existing SFT approaches generally overemphasize the annotated CoT, potentially leading to performance degradation due to insufficient exploitation of potential CoT. In this paper, we propose a Contrastive learning with annotated CoT-based Reinforced Fine-Tuning approach, i.e., CARFT, to enhance the reasoning performance of LLMs while addressing the aforementioned limitations. Specifically, we propose learning a representation for each CoT. Based on this representation, we design novel contrastive signals to guide the fine-tuning process. Our approach not only fully exploits the available annotated CoT but also stabilizes the fine-tuning procedure by incorporating an additional unsupervised learning signal. We conduct comprehensive experiments and in-depth analysis with three baseline approaches, two foundation models, and two datasets to demonstrate significant advantages of CARFT in terms of robustness, performance (up to 10.15%), and efficiency (up to 30.62%).
Large Language Models (LLMs) rely heavily on large-scale reasoning data, but as such data becomes increasingly scarce, model self-improvement offers a promising alternative. However, this process can lead to model collapse, as the model’s output becomes overly deterministic with reduced diversity. In this work, we identify a new risk beyond model collapse, which we term the Superficial Self-Improved Reasoners phenomenon. This phenomenon indicates that while self-improvement enhances in-domain (ID) reasoning accuracy, it degrades the model’s generalized reasoning capability on out-of-domain (OOD) datasets, as the model tends to memorize the training data. Our analyses of layer importance and parameter changes reveal that reasoning-critical layers receive fewer updates compared to less relevant layers during self-improvement. To address this, we propose Iterative Model Merging (IMM), which balances reasoning improvements and generalization by merging the weights of the original and self-improved models. IMM effectively mitigates model collapse and improves generalized reasoning capability. Code is available at https://github.com/xiangchi-yuan/merge_syn
Discovering customer intentions is crucial for automated service agents, yet existing intent clustering methods often fall short due to their reliance on embedding distance metrics and neglect of underlying semantic structures. To address these limitations, we propose an **LLM-in-the-loop (LLM-ITL)** intent clustering framework, integrating the language understanding capabilities of LLMs into conventional clustering algorithms. Specifically, this paper (1) examines the effectiveness of fine-tuned LLMs in semantic coherence evaluation and intent cluster naming, achieving over 95% accuracy aligned with human judgments; (2) designs an LLM-ITL framework that facilitates the iterative discovery of coherent intent clusters and the optimal number of clusters; and (3) introduces context-aware techniques tailored for customer service dialogue. Since existing English benchmarks lack sufficient semantic diversity and intent coverage, we further present a comprehensive Chinese dialogue intent dataset comprising over 100k real customer service calls with 1,507 human-annotated clusters. The proposed approaches significantly outperform LLM-guided baselines, achieving notable improvements in clustering quality, cost efficiency, and downstream applications. Combined with several best practices, our findings highlight the prominence of LLM-in-the-loop techniques for scalable dialogue data mining.
LLM-Driven Implicit Target Augmentation and Fine-Grained Contextual Modeling for Zero-Shot and Few-Shot Stance Detection
PDF ↗Stance detection aims to identify the attitude expressed in text towards a specific target. Recent studies on zero-shot and few-shot stance detection focus primarily on learning generalized representations from explicit targets. However, these methods often neglect implicit yet semantically important targets and fail to adaptively adjust the relative contributions of text and target in light of contextual dependencies. To overcome these limitations, we propose a novel two-stage framework: First, a data augmentation framework named Hierarchical Collaborative Target Augmentation (HCTA) employs Large Language Models (LLMs) to identify and annotate implicit targets via Chain-of-Thought (CoT) prompting and multi-LLM voting, significantly enriching training data with latent semantic relations. Second, we introduce DyMCA, a Dynamic Multi-level Context-aware Attention Network, integrating a joint text-target encoding and a content-aware mechanism to dynamically adjust text-target contributions based on context. Experiments on the benchmark dataset demonstrate that our approach achieves state-of-the-art results, confirming the effectiveness of implicit target augmentation and fine-grained contextual modeling.
Eliminating toxicity from Large Language Models (LLMs) is crucial for ensuring user safety. However, current methods have limitations in the analysis and utilization of toxic samples, failing to fully harness their potential. Through comparative analysis of toxic and safe samples, we discover that toxic samples exhibit diversity and, within this diversity, there lies specificity. These findings suggest that leveraging these characteristics of toxic samples could enhance the performance of algorithms in detoxifying LLMs. To this end, we propose a novel diverse detoxification framework, DivDetox, which comprises two innovative components: a Multi-Category-Induced Personalized Sample Generation (MPSG) strategy and a Scaled Contrastive DPO (SC-DPO) approach. The former is designed to elicit a variety of personalized toxic responses from the LLM, while the latter is constructed to precisely and fully utilize these toxic responses. Experiments on benchmark datasets across different model scales and different detoxification tasks verify the effectiveness of our architecture.
In-Context derived Vector (ICV) methods extract task-relevant representations from large language models (LLMs) and reinject them during inference, achieving comparable performance to few-shot In-Context Learning (ICL) without repeated demonstration processing. However, existing ICV methods remain sensitive to ICL-specific factors, often use coarse or semantically fragmented representations as the source of the vector, and rely on heuristic-based injection positions, limiting their applicability.To address these issues, we propose Dynamic Vector (DyVec), which incorporates an Exhaustive Query Rotation (EQR) strategy to extract robust semantically aggregated latent representations by mitigating variance introduced by ICL. It then applies Dynamic Latent Segmentation and Injection to adaptively partition representations based on task complexity and leverages REINFORCE-based optimization to learn optimal injection positions for each segment.Experiments results show that DyVec outperforms few-shot ICL, LoRA, and prior ICV baselines. Further analysis highlights the effectiveness of dynamically segmenting and injecting semantically aggregated latent representations. DyVec provides a lightweight and data-efficient solution for inference-time task adaptation.
Pragmatic Inference Chain (PIC) Improving LLMs’ Reasoning of Authentic Implicit Toxic Language
PDF ↗The rapid development of large language models (LLMs) gives rise to ethical concerns about their performance, while opening new avenues for developing toxic language detection techniques. However, LLMs’ unethical output and their capability of detecting toxicity have primarily been tested on language data that do not demand complex meaning inference, such as the biased associations of ‘he’ with programmer and ‘she’ with household. Nowadays toxic language adopts a much more creative range of implicit forms, thanks to advanced censorship. In this study, we collect authentic toxic interactions that evade online censorship and that are verified by human annotators as inference-intensive. To evaluate and improve LLMs’ reasoning of the authentic implicit toxic language, we propose a new prompting method, Pragmatic Inference Chain (PIC), drawn on interdisciplinary findings from cognitive science and linguistics. The PIC prompting significantly improves the success rate of GPT-4o, Llama-3.1-70B-Instruct, DeepSeek-v2.5, and DeepSeek-v3 in identifying implicit toxic language, compared to five baseline prompts, such as CoT and rule-based baselines. In addition, it also facilitates the models to produce more explicit and coherent reasoning processes, hence can potentially be generalized to other inference-intensive tasks, e.g., understanding humour and metaphors.
Model merging has emerged as a promising approach for updating large language models (LLMs) by integrating multiple domain-specific models into a cross-domain merged model. Despite its utility and plug-and-play nature, unmonitored mergers can introduce significant security vulnerabilities, such as backdoor attacks and model merging abuse. In this paper, we identify a novel and more realistic attack surface where a malicious merger can extract targeted personally identifiable information (PII) from an aligned model with model merging. Specifically, we propose Merger-as-a-Stealer, a two-stage framework to achieve this attack: First, the attacker fine-tunes a malicious model to force it to respond to any PII-related queries. The attacker then uploads this malicious model to the model merging conductor and obtains the merged model. Second, the attacker inputs direct PII-related queries to the merged model to extract targeted PII. Extensive experiments demonstrate that Merger-as-a-Stealer successfully executes attacks against various LLMs and model merging methods across diverse settings, highlighting the effectiveness of the proposed framework. Given that this attack enables character-level extraction for targeted PII without requiring any additional knowledge from the attacker, we stress the necessity for improved model alignment and more robust defense mechanisms to mitigate such threats.