ACL 2025 论文列表

Implementing new features in repository-level codebases is a crucial application of code generation models. However, current benchmarks lack a dedicated evaluation framework for this capability. To fill this gap, we introduce FEA-Bench, a benchmark designed to assess the ability of large language models (LLMs) to perform incremental development within code repositories. We collect pull requests from 83 GitHub repositories and use rule-based and intent-based filtering to construct task instances focused on new feature development. Each task instance containing code changes is paired with relevant unit test files to ensure that the solution can be verified. The feature implementation requires LLMs to simultaneously possess code completion capabilities for new components and code editing abilities for other relevant parts in the code repository, providing a more comprehensive evaluation method of LLMs’ automated software engineering capabilities.Experimental results show that LLMs perform significantly worse in the FEA-Bench, highlighting considerable challenges in such repository-level incremental code development.

Real-world decision-making often requires integrating and reasoning over information from multiple modalities. While recent multimodal large language models (MLLMs) have shown promise in such tasks, their ability to perform multi-hop reasoning across diverse sources remains insufficiently evaluated. Existing benchmarks, such as MMQA, face challenges due to (1) data contamination and (2) a lack of complex queries that necessitate operations across more than two modalities, hindering accurate performance assessment. To address this, we present Financial Cross-Modal Multi-Hop Reasoning (FCMR), a benchmark created to analyze the reasoning capabilities of MLLMs by urging them to combine information from textual reports, tables, and charts within the financial domain. FCMR is categorized into three difficulty levels—Easy, Medium, and Hard—facilitating a step-by-step evaluation. In particular, problems at the Hard level require precise cross-modal three-hop reasoning and are designed to prevent the disregard of any modality. Experiments on this new benchmark reveal that even state-of-the-art MLLMs struggle, with the best-performing model (Claude 3.5 Sonnet) achieving only 30.4% accuracy on the most challenging tier. We also conduct analysis to provide insights into the inner workings of the models, including the discovery of a critical bottleneck in the information retrieval phase.

Instruction tuning stands as a crucial advancement in leveraging large language models (LLMs) for enhanced task performance. However, the annotation of instruction datasets has traditionally been expensive and laborious, often relying on manual annotations or costly proprietary LLMs. Recent works explore approaches to synthesize data with open-sourced LLMs but require high-quality human-crafted seed data. In this work, we introduce , an end-to-end framework to synthesize high-quality instruction data with open-sourced LLMs and sampled unlabeled documents, eliminating the necessity for seed data. Starting from diverse pre-screened documents, the framework synthesizes complex and diverse high-quality instruction and response pairs in different stages. We propose a tagging-based prompt method to generate diverse and complex seed data and a UCB-based approach to augment more instruction data with the seed data. A novel Think Different prompt is proposed to address the distributional limitations of the seeds, further boosting the data diversity. Experiments prove that the can generate diverse and complex high-quality data even with a opensource small teacher model. The synthesized instruction data demonstrates performance that is comparable to, or even surpasses, baseline annotation methods with proprietary LLMs or open-sourced LLMs while requiring fewer instruction data samples.

Recent advances in mechanistic interpretability have highlighted the potential of automating interpretability pipelines in analyzing the latent representations within LLMs. While this may enhance our understanding of internal mechanisms, the field lacks standardized evaluation methods for assessing the validity of discovered features. We attempt to bridge this gap by introducing **FADE**: Feature Alignment to Description Evaluation, a scalable model-agnostic framework for automatically evaluating feature-to-description alignment. **FADE** evaluates alignment across four key metrics – *Clarity, Responsiveness, Purity, and Faithfulness* – and systematically quantifies the causes of the misalignment between features and their descriptions. We apply **FADE** to analyze existing open-source feature descriptions and assess key components of automated interpretability pipelines, aiming to enhance the quality of descriptions. Our findings highlight fundamental challenges in generating feature descriptions, particularly for SAEs compared to MLP neurons, providing insights into the limitations and future directions of automated interpretability. We release **FADE** as an open-source package at: [github.com/brunibrun/FADE](https://github.com/brunibrun/FADE).

Large Language Models (LLMs) have significantly advanced the fact-checking studies. However, existing automated fact-checking evaluation methods rely on static datasets and classification metrics, which fail to automatically evaluate the justification production and uncover the nuanced limitations of LLMs in fact-checking. In this work, we introduce FACT-AUDIT, an agent-driven framework that adaptively and dynamically assesses LLMs’ fact-checking capabilities. Leveraging importance sampling principles and multi-agent collaboration, FACT-AUDIT generates adaptive and scalable datasets, performs iterative model-centric evaluations, and updates assessments based on model-specific responses. By incorporating justification production alongside verdict prediction, this framework provides a comprehensive and evolving audit of LLMs’ factual reasoning capabilities, to investigate their trustworthiness. Extensive experiments demonstrate that FACT-AUDIT effectively differentiates among state-of-the-art LLMs, providing valuable insights into model strengths and limitations in model-centric fact-checking analysis.

This paper introduces F5-TTS, a fully non-autoregressive text-to-speech system based on flow matching with Diffusion Transformer (DiT). Without requiring complex designs such as duration model, text encoder, and phoneme alignment, the text input is simply padded with filler tokens to the same length as input speech, and then the denoising is performed for speech generation, which was originally proved feasible by E2 TTS. However, the original design of E2 TTS makes it hard to follow due to its slow convergence and low robustness. To address these issues, we first model the input with ConvNeXt to refine the text representation, making it easy to align with the speech. We further propose an inference-time Sway Sampling strategy, which significantly improves our model’s performance and efficiency. This sampling strategy for flow step can be easily applied to existing flow matching based models without retraining. Our design allows faster training and achieves an inference RTF of 0.15, which is greatly improved compared to state-of-the-art diffusion-based TTS models. Trained on a public 100K hours multilingual dataset, our F5-TTS exhibits highly natural and expressive zero-shot ability, seamless code-switching capability, and speed control efficiency. We have released all codes and checkpoints to promote community development, at https://SWivid.github.io/F5-TTS/.

Processing long input remains a significant challenge for large language models (LLMs) due to the scarcity of large-scale long-context training data and the high computational cost of training models for extended context windows. In this paper, we propose **Ada**ptive **Gro**uped **P**ositional **E**ncoding (AdaGroPE), a training-free, plug-and-play method to enhance long-context understanding in existing LLMs. AdaGroPE progressively increases the reuse count of relative positions as the distance grows and dynamically adapts the positional encoding mapping to sequence length, thereby fully exploiting the range of pre-trained position embeddings. Its design is consistent with the principles of rotary position embedding (RoPE) and aligns with human perception of relative distance, enabling robust performance in real-world settings with variable-length inputs. Extensive experiments across various benchmarks demonstrate that our AdaGroPE consistently achieves state-of-the-art performance, surpassing baseline methods and even outperforming LLMs inherently designed for long-context processing on certain tasks.

The study of machine learning-based logical query-answering enables reasoning with large-scale and incomplete knowledge graphs. This paper further advances this line of research by considering the uncertainty in the knowledge. The uncertain nature of knowledge is widely observed in the real world, but does not align seamlessly with the first-order logic underpinning existing studies. To bridge this gap, we study the setting of soft queries on uncertain knowledge, which is motivated by the establishment of soft constraint programming. We further propose an ML-based approach with both forward inference and backward calibration to answer soft queries on large-scale, incomplete, and uncertain knowledge graphs. Theoretical discussions reveal that our method ensures there are no catastrophic cascading errors in our forward inference algorithm while maintaining the same complexity as state-of-the-art inference algorithms for first-order queries. Empirical results justify the superior performance of our approach against previous ML-based methods with number embedding extensions.

Bridging the gap between visual and language remains a pivotal challenge for the multimodal community. Traditional VQA benchmarks encounter a modality gap and over-reliance on language priors, whereas human cognition excels at intuitive semiosis, associating abstract visual symbols to linguistic semantics. Inspired by this neurocognitive mechanism, we focus on emojis, the visual cipher conveying abstract textual semantics. Specifically, we propose a novel task of generating abstract linguistics from emoji sequence images, where such reasoning underpins critical applications in cryptography, thus challenging MLLMs’ reasoning of decoding complex semantics of visual ciphers. We introduce eWe-bench (Express What you SeE), assessing MLLMs’ capability of intuitive semiosis like humans. Our data construction framework ensures high visual sensitivity and data quality, which can be extended to future data enhancement. Evaluation results on advanced MLLMs highlight critical deficiencies in visual intuitive symbolic reasoning. We believe our interesting insights for advancing visual semiosis in MLLMs will pave the way for cryptographic analysis and high-level intuitive cognition intelligence of MLLMs.

Large Language Models (LLMs) have significantly impacted the field of Math Word Problems (MWPs), transforming how these problems are approached and solved, particularly in educational contexts. However, existing evaluations often focus on final accuracy, neglecting the critical aspect of reasoning capabilities. This work addresses that gap by evaluating LLMs’ abilities to detect and correct reasoning mistakes. We present a novel dataset, MWP-MISTAKE, containing MWPs with both correct and incorrect reasoning steps generated through rule-based methods and smaller language models. Our comprehensive benchmarking of state-of-the-art models such as GPT-4o and GPT4 uncovers important insights into their strengths and limitations. While GPT-4o excels in mistake detection and rectification, gaps remain, particularly in handling complex datasets and novel problems. Additionally, we identify concerns with data contamination and memorization, which affect LLM reliability in real-world applications. While OpenAI’ O1 model demonstrates 90% accuracy in reasoning and final answers on complex tasks, it remains weak in mistake detection. Our findings highlight the need for improved reasoning evaluations and suggest ways to enhance LLM generalization and robustness in math problem-solving.

Large Language Models (LLMs) have demonstrated impressive capabilities in natural language processing tasks, such as text generation and semantic understanding. However, their performance on numerical reasoning tasks, such as basic arithmetic, numerical retrieval, and magnitude comparison, remains surprisingly poor. This gap arises from their reliance on surface-level statistical patterns rather than understanding numbers as continuous magnitudes. Existing benchmarks primarily focus on either linguistic competence or structured mathematical problem-solving, neglecting fundamental numerical reasoning required in real-world scenarios. To bridge this gap, we propose NumericBench, a comprehensive benchmark to evaluate six fundamental numerical capabilities: number recognition, arithmetic operations, contextual retrieval, comparison, summary, and multi-step reasoning. NumericBench includes datasets ranging from synthetic number lists to crawled real-world data, addressing challenges like long contexts, noise, and multi-step reasoning. Extensive experiments on state-of-the-art LLMs, including GPT-4 and DeepSeek, reveal persistent weaknesses in numerical reasoning, highlighting the urgent need to improve numerically-aware language modeling. The benchmark is released in: https://github.com/TreeAI-Lab/NumericBench.

The integration of Artificial Intelligence (AI) into mental health education and training (MHET) has become a promising solution to meet the increasing demand for skilled mental health professionals. This systematic review analyses 38 studies on AI-powered conversational agents (CAs) in MHET, selected from a total of 1003 studies published between 2019 and 2024. Following the PRISMA protocol, we reviewed papers from computer science, medicine, and interdisciplinary databases, assessing key aspects such as technological approaches, data characteristics, application areas, and evaluation methodologies. Our findings reveal that AI-based approaches, including Large Language Models (LLMs), dominate the field, with training as the application area being the most prevalent. These technologies show promise in simulating therapeutic interactions but face challenges such as limited public datasets, lack of standardised evaluation frameworks, and difficulty in ensuring authentic emotional responses, along with gaps in ethical considerations and clinical efficacy. This review presents a comprehensive framework for understanding the role of CAs in MHET while providing valuable recommendations to guide future research.

Personalized AI assistants, a hallmark of the human-like capabilities of Large Language Models (LLMs), are a challenging application that intertwines multiple problems in LLM research. Despite the growing interest in the development of personalized assistants, the lack of an open-source conversational dataset tailored for personalization remains a significant obstacle for researchers in the field. To address this research gap, we introduce HiCUPID, a new benchmark to probe and unleash the potential of LLMs to deliver personalized responses. Alongside a conversational dataset, HiCUPID provides a Llama-3.2-based automated evaluation model whose assessment closely mirrors human preferences. We release our dataset, evaluation model, and code at https://github.com/12kimih/HiCUPID.

Instruction-tuning enhances the ability of large language models (LLMs) to follow user instructions more accurately, improving usability while reducing harmful outputs. However, this process may increase the model’s dependence on user input, potentially leading to the unfiltered acceptance of misinformation and the generation of hallucinations. Existing studies primarily highlight that LLMs are receptive to external information that contradict their parametric knowledge, but little research has been conducted on the direct impact of instruction-tuning on this phenomenon. In our study, we investigate the impact of instruction-tuning on LLM susceptibility to misinformation. Our analysis reveals that instruction-tuned LLMs are significantly more likely to accept misinformation when it is presented by the user. A comparison with base models shows that instruction-tuning increases reliance on user-provided information, shifting susceptibility from the assistant role to the user role. Furthermore, we explore additional factors influencing misinformation susceptibility, such as the role of the user in prompt structure, misinformation length, and the presence of warnings in the system prompt. Our findings underscore the need for systematic approaches to mitigate unintended consequences of instruction-tuning and enhance the reliability of LLMs in real-world applications.

Large Language Models (LLMs) are increasingly deployed as human assistants across various domains where they help to make choices. However, the mechanisms behind LLMs’ choice behavior remain unclear, posing risks in safety-critical situations. Inspired by the intrinsic and extrinsic motivation framework within the classic human behavioral model of Self-Determination Theory and its established research methodologies, we investigate the factors influencing LLMs’ choice behavior by constructing a virtual QA platform that includes three different experimental conditions, with four models from GPT and Llama series participating in repeated experiments. Our findings indicate that LLMs’ behavior is influenced not only by intrinsic attention bias but also by extrinsic social influence, exhibiting patterns similar to the Matthew effect and Conformity. We distinguish independent pathways of these two factors in LLMs’ behavior by self-report. This work provides new insights into understanding LLMs’ behavioral patterns, exploring their human-like characteristics.

We investigate the prevalence of anthropomorphic language in the reporting of AI technology, focussed on NLP and LLMs. We undertake a corpus annotation focussing on one year of ACL long-paper abstracts and news articles from the same period. We find that 74% of ACL abstracts and 88% of news articles contain some form of anthropomorphic description of AI technology. Further, we train a regression classifier based on BERT, demonstrating that we can automatically label abstracts for their degree of anthropomorphism based on our corpus. We conclude by applying this labelling process to abstracts available in the entire history of the ACL Anthology and reporting on diachronic and inter-venue findings, showing that the degree of anthropomorphism is increasing at all examined venues over time.

Personalized dialogue systems have advanced considerably with the integration of user-specific personas into large language models (LLMs). However, while LLMs can effectively generate personalized responses, the influence of persona sentiment on dialogue quality remains underexplored. In this work, we conduct a large-scale analysis of dialogues generated using a range of polarized user profiles. Our experiments reveal that dialogues involving negatively polarized users tend to overemphasize persona attributes. In contrast, positively polarized profiles yield dialogues that selectively incorporate persona information, resulting in smoother interactions. Furthermore, we find that personas with weak or neutral sentiment generally produce lower-quality dialogues. Motivated by these findings, we propose a dialogue generation approach that explicitly accounts for persona polarity by combining a turn-based generation strategy with a profile ordering mechanism and sentiment-aware prompting. Our study provides new insights into the sensitivity of LLMs to persona sentiment and offers guidance for developing more robust and nuanced personalized dialogue systems.

Relation Extraction (RE) is a key task in table understanding, aiming to extract semantic relations between columns. However, complex tables with hierarchical headers are hard to obtain high-quality textual formats (e.g., Markdown) for input under practical scenarios like webpage screenshots and scanned documents, while table images are more accessible and intuitive. Besides, existing works overlook the need of mining relations among multiple columns rather than just the semantic relation between two specific columns in real-world practice. In this work, we explore utilizing Multimodal Large Language Models (MLLMs) to address RE in tables with complex structures. We creatively extend the concept of RE to include calculational relations, enabling multi-task learning of both semantic and calculational RE for mutual reinforcement. Specifically, we reconstruct table images into graph structure based on neighboring nodes to extract graph-level visual features. Such feature enhancement alleviates the insensitivity of MLLMs to the positional information within table images. We then propose a Chain-of-Thought distillation framework with self-correction mechanism to enhance MLLMs’ reasoning capabilities without increasing parameter scale. Our method significantly outperforms most baselines on wide datasets. Additionally, we release a benchmark dataset for calculational RE in complex tables.

Detecting toxic content using language models is important but challenging. While large language models (LLMs) have demonstrated strong performance in understanding Chinese, recent studies show that simple character substitutions in toxic Chinese text can easily confuse the state-of-the-art (SOTA) LLMs. In this paper, we highlight the multimodal nature of Chinese language as a key challenge for deploying LLMs in toxic Chinese detection. First, we propose a taxonomy of 3 perturbation strategies and 8 specific approaches in toxic Chinese content. Then, we curate a dataset based on this taxonomy, and benchmark 9 SOTA LLMs (from both the US and China) to assess if they can detect perturbed toxic Chinese text. Additionally, we explore cost-effective enhancement solutions like in-context learning (ICL) and supervised fine-tuning (SFT). Our results reveal two important findings. (1) LLMs are less capable of detecting perturbed multimodal Chinese toxic contents. (2) ICL or SFT with a small number of perturbed examples may cause the LLMs “overcorrect”: misidentify many normal Chinese contents as toxic.