Multi-turn instruction following capability constitutes a core competency of large language models (LLMs) in real-world applications. Existing evaluation benchmarks predominantly focus on fine-grained constraint satisfaction and domain-specific capability assessment, yet overlook the crucial structural dependencies between dialogue turns that distinguish multi-turn from single-turn interactions. These structural dependencies not only reflect user intent but also establish an essential second dimension for the instruction following evaluation beyond constraint satisfaction. To address this gap, we propose StructFlowBench, a multi-turn instruction following benchmark with structural flow modeling. The benchmark defines an innovative structural flow framework with six fundamental inter-turn relationships. These relationships introduce novel structural constraints for model evaluation and also serve as generation parameters for creating customized dialogue flows tailored to specific scenarios. Adopting established LLM-based automatic evaluation methodologies, we conduct systematic evaluations of 13 leading open-source and closed-source LLMs. Experimental results reveal significant deficiencies in current models’ comprehension of multi-turn dialogue structures. The code is available at https://github.com/MLGroupJLU/StructFlowBench.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
HTML: Hierarchical Topology Multi-task Learning for Semantic Parsing in Knowledge Base Question Answering
PDF ↗Knowledge base question answering (KBQA) aims to answer natural language questions by reasoning over structured knowledge bases. Existing approaches often struggle with the complexity of mapping questions to precise logical forms, particularly when dealing with diverse entities and relations. In this paper, we propose Hierarchical Topology Multi-task Learning (HTML), a novel framework that leverages a hierarchical multi-task learning paradigm to enhance the performance of logical form generation. Our framework consists of a main task: generating logical forms from questions, and three auxiliary tasks: entity prediction from the input question, relation prediction for the given entities, and logical form generation based on the given entities and relations. Through joint instruction-tuning, HTML allows mutual guidance and knowledge transfer among the hierarchical tasks, capturing the subtle dependencies between entities, relations, and logical forms. Extensive experiments on public benchmarks show that HTML markedly outperforms both supervised fine-tuning methods and training-free ones based on powerful large language models (e.g., GPT-4), demonstrating its superiority in question understanding and structural knowledge reasoning.
ReKG-MCTS: Reinforcing LLM Reasoning on Knowledge Graphs via Training-Free Monte Carlo Tree Search
PDF ↗Recent advancements in combining knowledge graphs (KGs) with large language models (LLMs) have demonstrated promising potential in complex KG reasoning tasks, yet existing approaches face limitations in path exploration strategies or excessive computational overhead. We propose ReKG-MCTS, a novel training-free framework that synergizes Monte Carlo Tree Search (MCTS) with LLM capabilities to enable dynamic reasoning over KGs. The framework conceptualizes KG reasoning as a decision-making process, where MCTS strategically explores paths over KG while LLMs provide semantic guidance for reasoning paths. The framework consists of four phases: (1) UCB-based node selection that balances exploration-exploitation on KG, (2) path expansion with KG structural constraints, (3) LLM-guided MC rollouts for simulation, and (4) value backpropagation. Experimental results on WebQSP and CWQ demonstrate that ReKG-MCTS outperforms existing training-free methods and achieves competitive performance compared to fine-tuned baselines. These findings suggest a new paradigm for leveraging language models in KG reasoning tasks. The code is available at https://github.com/ShawnKS/rekgmcts.
Multimodal machine translation (MMT) integrates visual information to address ambiguity and contextual limitations in neural machine translation (NMT). Some empirical studies have revealed that many MMT models underutilize visual data during translation. They attempt to enhance cross-modal interactions to enable better exploitation of visual data. However, they only focus on simple interactions between nouns in text and corresponding entities in image, overlooking global semantic alignment, particularly for prepositional phrases and verbs in text which are more likely to be translated incorrectly. To address this, we design a Text-Image In-depth Questioning method to deepen interactions and optimize translations. Furthermore, to mitigate errors arising from contextually irrelevant image noise, we propose a Consistency Constraint strategy to improve our approach’s robustness. Our approach achieves state-of-the-art results on five translation directions of Multi30K and AmbigCaps, with +2.35 BLEU on the challenging MSCOCO benchmark, validating our method’s effectiveness in utilizing visual data and capturing comprehensive textual semantics.
Current backdoor attack defenders in Natural Language Processing (NLP) typically involve data reduction or model pruning, risking losing crucial information. To address this challenge, we introduce a novel backdoor defender, i.e., BadWindtunnel, in which we build a high-noise simulated training environment, similar to the wind tunnel, which allows precise control over training conditions to model the backdoor learning behavior without affecting the final model. We also use the confidence variance as a learning behavior quantification metric in the simulated training, which is based on the characteristics of backdoor-poisoned data (shorted in poisoned data): higher learnability and robustness. In addition, we propose a two-step strategy to further model poisoned data, including target label identification and poisoned data revealing. Extensive experiments demonstrate BadWindtunnel’s superiority, with a 21% higher average reduction in attack success rate than the second-best defender.
MultiMSD: A Corpus for Multilingual Medical Text Simplification from Online Medical References
PDF ↗We release a parallel corpus for medical text simplification, which paraphrases medical terms into expressions easily understood by patients. Medical texts written by medical practitioners contain a lot of technical terms, and patients who are non-experts are often unable to use the information effectively. Therefore, there is a strong social demand for medical text simplification that paraphrases input sentences without using medical terms. However, this task has not been sufficiently studied in non-English languages. We therefore developed parallel corpora for medical text simplification in nine languages: German, English, Spanish, French, Italian, Japanese, Portuguese, Russian, and Chinese, each with 10,000 sentence pairs, by automatic sentence alignment to online medical references for professionals and consumers. We also propose a method for training text simplification models to actively paraphrase complex expressions, including medical terms. Experimental results show that the proposed method improves the performance of medical text simplification. In addition, we confirmed that training with a multilingual dataset is more effective than training with a monolingual dataset.
Recent research has focused on investigating the psychological characteristics of Large Language Models (LLMs), emphasizing the importance of comprehending their behavioral traits. Likert scale personality questionnaires have become the primary tool for assessing these characteristics in LLMs. However, such scales can be skewed by factors such as social desirability, distorting the assessment of true personality traits. To address this issue, we firstly incorporate the forced-choice test, a method known for reducing response bias in human personality assessments, into the evaluation of LLM. Specifically, we evaluated six LLMs: Llama-3.1-8B, GLM-4-9B, GPT-3.5-turbo, GPT-4o, Claude-3.5-sonnet, and Deepseek-V3. We compared the Likert scale and forced-choice test results for LLMs’ Big Five personality scores, as well as their reliability. In addition, we looked at how temperature parameter and language affected LLM personality scores. The results show that the forced-choice test better captures differences between LLMs across various personality dimensions and is less influenced by temperature parameters. Furthermore, we found both broad trends and specific variations in personality scores across models and languages.
We introduce semantic topology, a novel framework for discourse analysis that leverages Circuit Topology to quantify the semantic arrangement of sentences in a text. By mapping recurring themes as series, parallel, or cross relationships, we identify statistical differences in communication patterns in long-form true and fake news. Our analysis of large-scale news datasets reveals that true news are more likely to exhibit more complex topological structures, with greater thematic interleaving and long-range coherence, whereas fake news favor simpler, more linear narratives. These findings suggest that topological features capture stylistic distinctions beyond traditional linguistic cues, offering new insights for discourse modeling.
Hallucination remains a critical challenge for multimodal large language models (MLLMs), undermining their reliability in real-world applications. While fine-grained hallucination detection (FHD) holds promise for enhancing high-quality vision-language data construction and model alignment through enriched feedback signals, automated solutions for this task have yet to be systematically explored. Inspired by the concept of “MLLM as a Judge”, we introduce MHALO, the first comprehensive benchmark specifically designed for evaluating MLLMs’ capability in performing token-level FHD. Our benchmark encompasses 12 distinct hallucination types spanning both multimodal perception and reasoning domains. Through extensive evaluations of 9 selected MLLMs, we reveal substantial performance limitations, with the leading model achieving an average F1_{IoU} of only 40.59%. To address this limitation, we develop HaloDet-4B, a specialized model trained on our curated training data, which significantly outperforms existing models. We hope the benchmark can provide valuable insights for future research on hallucination mitigation in MLLMs. The code and dataset will be publicly available.
Narrative frames are a powerful way of conceptualizing and communicating complex, controversial ideas, however automated frame analysis to date has mostly overlooked this framing device. In this paper, we connect elements of narrativity with fundamental aspects of framing, and present a framework which formalizes and operationalizes such aspects. We annotate and release a data set of news articles in the climate change domain, analyze the dominance of narrative frame components across political leanings, and test LLMs in their ability to predict narrative frames and their components. Finally, we apply our framework in an unsupervised way to elicit components of narrative framing in a second domain, the COVID-19 crisis, where our predictions are congruent with prior theoretical work showing the generalizability of our approach.
State-of-the-art large language models (LLMs) have demonstrated impressive code generation capabilities but struggle with real-world software engineering tasks, such as revising source code to address code reviews, hindering their practical use. Code review comments are often implicit, ambiguous, and colloquial, requiring models to grasp both code and human intent. This challenge calls for evaluating large language models’ ability to bridge both technical and conversational contexts. While existing work has employed the automated code refinement (ACR) task to resolve these comments, current evaluation methods fall short, relying on text matching metrics that provide limited insight into model failures and remain susceptible to training data contamination.To address these limitations, we introduce a novel evaluation benchmark, \textbf{CodeReviewQA} that enables us to conduct fine-grained assessment of model capabilities and mitigate data contamination risks.In CodeReviewQA, we decompose the generation task of code refinement into \textbf{three essential reasoning steps}: \textit{change type recognition} (CTR), \textit{change localisation} (CL), and \textit{solution identification} (SI). Each step is reformulated as multiple-choice questions with varied difficulty levels, enabling precise assessment of model capabilities, while mitigating data contamination risks. Our comprehensive evaluation spans 72 recently released large language models on \textbf{900 manually curated, high-quality examples} across nine programming languages. Our results show that CodeReviewQA is able to expose specific model weaknesses in code review comprehension, disentangled from their generative automated code refinement results.
Debate, Reflect, and Distill: Multi-Agent Feedback with Tree-Structured Preference Optimization for Efficient Language Model Enhancement
PDF ↗Large Language Models (LLMs) continue to set new standards in knowledge-intensive and complex reasoning tasks, yet their high computational demands limit widespread adoption. While distilling large models into smaller ones offers a sustainable solution, current techniques—such as static knowledge distillation, resource-intensive reinforcement learning from human feedback, or limited self-reflection—struggle to yield substantial and lasting performance gains. In this paper, we present a novel Debate and Reflect (D&R) framework that orchestrates multi-turn debates between smaller models and stronger teacher models, eliciting actionable feedback (e.g., error analysis, corrective strategies) to guide student models. Further, we introduce Tree-structured Direct Preference Optimization (T-DPO) to efficiently leverage these debate logs, organizing interactions into a hierarchical format for effective training. Empirical evaluations across diverse NLP benchmarks demonstrate that our approach significantly improves smaller-model accuracy, robustness, and generalization, outperforming conventional baselines by a large margin.
Multimodal Large Language Models (MLLMs) are measured on numerous benchmarks like image captioning, visual question answer, and reasoning. However, these benchmarks often include overly simple or uninformative samples, making it difficult to effectively distinguish the performance of different MLLMs. Additionally, evaluating models across many benchmarks creates a significant computational burden. To address these issues, we propose LIME (Less Is More for MLLM Evaluation), a refined and efficient benchmark curated using a semi-automated pipeline. This pipeline filters out uninformative samples and eliminates answer leakage by focusing on tasks that require image-based understanding. Our experiments show that LIME reduces the number of samples by 76% and evaluation time by 77%, while it can more effectively distinguish different models’ abilities. Notably, we find that traditional automatic metrics like CIDEr are insufficient for evaluating MLLMs’ captioning performance, and excluding the caption task score yields a more accurate reflection of overall model performance. All code and data are available at https://anonymous.4open.science/r/LIME-49CD
The key to effective alignment lies in high-quality preference data. Recent research has focused on automated alignment, which involves developing alignment systems with minimal human intervention. However, prior research has predominantly focused on developing data generation methods, while insufficient attention has been paid to quality control mechanisms and often produces inaccurate and unhelpful data, leading to unpredictable benefits during iterative optimization. In this paper, we present Self-Steering Optimization (SSO), an algorithm that autonomously generates high-quality preference data, eliminating manual annotation requirements. SSO employs a specialized optimization objective to build a data generator from the policy model itself, which is used to produce accurate and on-policy data. We demonstrate SSO‘s effectiveness through comprehensive experiments on two series of models: Llama 3 and Qwen 2. Our evaluation across diverse benchmarks shows that SSO consistently outperforms baselines in human preference alignment and reward optimization. Further analysis validates SSO as a scalable framework for preference optimization, benefiting the advancement in automated alignment techniques.
HellaSwag-Pro: A Large-Scale Bilingual Benchmark for Evaluating the Robustness of LLMs in Commonsense Reasoning
PDF ↗Large language models (LLMs) have shown remarkable capabilities in commonsense reasoning; however, some variations in questions can trigger incorrect responses. Do these models truly understand commonsense knowledge, or just memorize expression patterns? To investigate this question, we present the first extensive robustness evaluation of LLMs in commonsense reasoning. We introduce HellaSwag-Pro, a large-scale bilingual benchmark consisting of 11,200 cases, by designing and compiling seven types of question variants. To construct this benchmark, we propose a two-stage method to develop Chinese HellaSwag, a finely annotated dataset comprising 12,000 instances across 56 categories. We conduct extensive experiments on 41 representative LLMs, revealing that these LLMs are far from robust in commonsense reasoning. Furthermore, this robustness varies depending on the language in which the LLM is tested. This work establishes a high-quality evaluation benchmark, with extensive experiments offering valuable insights to the community in commonsense reasoning for LLMs.
Model merging techniques like task arithmetic, which combines model parameters through weighted averaging, have proven effective. However, the success of task arithmetic relies on the linearity between model weight differences and output feature changes, which is often lacking in conventional fine-tuned models. In this work, we employ neuron description methods to analyze and classify neurons based on their functionalities. We theoretically demonstrate that grouping Multi-Layer Perceptron (MLP) neurons by functionality enhances model linearity. Building on this, we propose a neuron-based task arithmetic merging method that consistently improves performance across various tasks and model scales. Our approach is complementary to existing merging techniques, achieving superior results in merging models fine-tuned on fundamental tasks like Math, Code and Translation.
Does Your Voice Assistant Remember? Analyzing Conversational Context Recall and Utilization in Voice Interaction Models
PDF ↗Recent advancements in multi-turn voice interaction models have improved user-model communication. However, while closed-source models effectively retain and recall past utterances, whether open-source models share this ability remains unexplored. To fill this gap, we systematically evaluate how well open-source interaction models utilize past utterances using ContextDialog, a benchmark we proposed for this purpose. Our findings show that speech-based models have more difficulty than text-based ones, especially when recalling information conveyed in speech, and even with retrieval-augmented generation, models still struggle with questions about past utterances. These insights highlight key limitations in open-source models and suggest ways to improve memory retention and retrieval robustness.
M-RangeDetector: Enhancing Generalization in Machine-Generated Text Detection through Multi-Range Attention Masks
PDF ↗The increasing capability and widespread usage of large language models (LLMs) highlight the desirability of automatic detection of machine-generated text. Existing supervised detectors often overfit within their training domains, as they have primarily learned domain-specific textual features, such as word frequency, syntax, and semantics. In this paper, we introduce a domain-independent feature, namely the difference of writing strategy between LLMs and human, to improve the out-of-domain generalization capability of detectors. LLMs focus on the preceding range tokens when generating a token, while human consider multiple ranges, including bidirectional, global, and local contexts. The attention mask influences the range of tokens to which the model can attend. Therefore, we propose a method called M-RangeDetector, which integrates four distinct attention masking strategies into a Multi-Range Attention module, enabling the model to capture diverse writing strategies. Specifically, with the global mask, band mask, dilated mask, and random mask, our method learns various writing strategies for machine-generated text detection. The experimental results on three datasets demonstrate the superior generalization capability of our method.
Social network simulation is developed to provide a comprehensive understanding of social networks in the real world, which can be leveraged for a wide range of applications such as group behavior emergence, policy optimization, and business strategy development. However, billions of individuals and their evolving interactions involved in social networks pose challenges in accurately reflecting real-world complexities. In this study, we propose a comprehensive Social network Simulation System (GA\text{-}S^3) that leverages newly designed Group Agents to make intelligent decisions regarding various online events. Unlike other intelligent agents that represent an individual entity, our group agents model a collection of individuals exhibiting similar behaviors, facilitating the simulation of large-scale network phenomena with complex interactions at a manageable computational cost. Additionally, we have constructed a social network benchmark from 2024 popular online events that contains fine-grained information on Internet traffic variations. The experiment demonstrates that our approach is capable of achieving accurate and highly realistic prediction results.
Boosting Vulnerability Detection of LLMs via Curriculum Preference Optimization with Synthetic Reasoning Data
PDF ↗Large language models (LLMs) demonstrate considerable proficiency in numerous coding-related tasks; however, their capabilities in detecting software vulnerabilities remain limited. This limitation primarily stems from two factors: (1) the absence of reasoning data related to vulnerabilities, which hinders the models’ ability to capture underlying vulnerability patterns; and (2) their focus on learning semantic representations rather than the reason behind them, thus failing to recognize semantically similar vulnerability samples. Furthermore, the development of LLMs specialized in vulnerability detection is challenging, particularly in environments characterized by the scarcity of high-quality datasets. In this paper, we propose a novel framework ReVD that excels at mining vulnerability patterns through reasoning data synthesizing and vulnerability-specific preference optimization. Specifically, we construct forward and backward reasoning processes for vulnerability and corresponding fixed code, ensuring the synthesis of high-quality reasoning data. Moreover, we design the triplet supervised fine-tuning followed by curriculum online preference optimization for enabling ReVD to better understand vulnerability patterns. The extensive experiments conducted on PrimeVul and SVEN datasets demonstrate that ReVD sets new state-of-the-art for LLM-based software vulnerability detection, e.g., 12.24%-22.77% improvement in the accuracy. The source code and data are available at https://github.com/Xin-Cheng-Wen/PO4Vul.