ACL 2025 Papers

By virtue of linguistic compositionality, few syntactic rules and a finite lexicon can generate an unbounded number of sentences. That is, language, though seemingly high-dimensional, can be explained using relatively few degrees of freedom. An open question is whether contemporary language models (LMs) reflect the intrinsic simplicity of language that is enabled by compositionality. We take a geometric view of this problem by relating the degree of compositionality in a dataset to the intrinsic dimension (ID) of its representations under an LM, a measure of feature complexity. We find not only that the degree of dataset compositionality is reflected in representations’ ID, but that the relationship between compositionality and geometric complexity arises due to learned linguistic features over training. Finally, our analyses reveal a striking contrast between nonlinear and linear dimensionality, showing they respectively encode semantic and superficial aspects of linguistic composition.

Advancing LLM reasoning skills has captivated wide interest. However, current post-training techniques rely heavily on supervisory signals, such as outcome supervision or auxiliary reward models, which face the problem of scalability and high annotation costs. This motivates us to enhance LLM reasoning without the need for external supervision. Given the input query, the LLM seeks the globally optimal response by stepwise sampling and self-rewarding, and optimizes itself with the collected responses. Genius offers some technical solutions to address the following key challenges. To tackle the problem of how to determine the steps in the response via self-rewarding, Genius introduces a stepwise foresight re-sampling strategy to sample and estimate the step value by simulating future outcomes. Recognizing the intrinsic noise and uncertainty of self-supervision, we propose an advantage-calibrated optimization (ACO) loss function to mitigate estimation inconsistencies. In short, Genius provides an advanced initial step towards self-improve LLM reasoning with general queries and without supervision, revolutionizing reasoning scaling laws given the vast availability of general queries.

Generative Reward Models (GenRMs) leverage synthesized Chains of Thought (CoT) to reduce the need for massive labeled data, but this approach introduces risks of overoptimization due to the inability to guarantee the correctness of the CoTs. Identifying and optimizing unexpected behaviors within these synthesized CoT remains a challenge, as it heavily depends on precise annotations of intermediate behavior, similar to process supervision. In this work, we introduce a criteria-based preference tree for reward modeling, where each path in the tree represents a reasoning trajectory based on synthesized criteria. Crucially, each reasoning trajectory can be independently optimized through RL algorithm. These fine-grained process reward signals are derived from the inference-time computations and predefined rules, eliminating the need for human supervision. In experiments, SyncPL showed significant improvements over baselines on multiple human preference benchmarks. We further demonstrate that synthesized data can be learned using a long CoT format, analogous to an o1-like model, further enhancing performance while keeping stability and efficiency during training.

Values are core drivers of individual and collective perception, cognition, and behavior. Value systems, such as Schwartz’s Theory of Basic Human Values, delineate the hierarchy and interplay among these values, enabling cross-disciplinary investigations into decision-making and societal dynamics. Recently, the rise of Large Language Models (LLMs) has raised concerns regarding their elusive intrinsic values. Despite growing efforts in evaluating, understanding, and aligning LLM values, a psychologically grounded LLM value system remains underexplored. This study addresses the gap by introducing the Generative Psycho-Lexical Approach (GPLA), a scalable, adaptable, and theoretically informed method for constructing value systems. Leveraging GPLA, we propose a psychologically grounded five-factor value system tailored for LLMs. For systematic validation, we present three benchmarking tasks that integrate psychological principles with cutting-edge AI priorities. Our results reveal that the proposed value system meets standard psychological criteria, better captures LLM values, improves LLM safety prediction, and enhances LLM alignment, when compared to the canonical Schwartz’s values.

Recent years have witnessed rapid advancements in text-to-music generation using large language models, yielding notable outputs. A critical challenge is understanding users with diverse musical expertise and generating music that meets their expectations, an area that remains underexplored.To address this gap, we introduce the novel task of Professional and Amateur Description-to-Song Generation. This task focuses on aligning generated content with human expressions from varying musical proficiency levels, aiming to produce songs that accurately meet auditory expectations and adhere to musical structural conventions. We utilized the MuChin dataset, which contains annotations from both professionals and amateurs for identical songs, as the source for these distinct description types. We also collected a pre-train dataset of over 1.5 million songs; lyrics were included for some, while for others, lyrics were generated using Automatic Speech Recognition (ASR) models.Furthermore, we propose MuDiT/MuSiT, a single-stage framework designed to enhance human-machine alignment in song generation. This framework employs Chinese MuLan (ChinMu) for cross-modal comprehension between natural language descriptions and auditory musical attributes, thereby aligning generated songs with user-defined outcomes. Concurrently, a DiT/SiT model facilitates end-to-end generation of complete songs audio, encompassing both vocals and instrumentation. We proposed metrics to evaluate semantic and auditory discrepancies between generated content and target music. Experimental results demonstrate that MuDiT/MuSiT outperforms baseline models and exhibits superior alignment with both professional and amateur song descriptions.

Despite recent advances in Video Large Language Models (VideoLLMs), effectively understanding long-form videos remains a significant challenge. Perceiving lengthy videos containing thousands of frames poses substantial computational burden. To mitigate this issue, this paper introduces Generative Frame Sampler (GenS), a plug-and-play module integrated with VideoLLMs to facilitate efficient lengthy video perception. Built upon a lightweight VideoLLM, GenS leverages its inherent vision-language capabilities to identify question-relevant frames. To facilitate effective retrieval, we construct GenS-Video-150K, a large-scale video instruction dataset with dense frame relevance annotations. Extensive experiments demonstrate that GenS consistently boosts the performance of various VideoLLMs, including open-source models (Qwen2-VL-7B, Aria-25B, LLaVA-Video-7B/72B) and proprietary assistants (GPT-4o, Gemini). When equipped with GenS, open-source VideoLLMs achieve impressive state-of-the-art results on long-form video benchmarks: LLaVA-Video-72B reaches 66.8 (+4.3) on LongVideoBench and 77.0 (+2.7) on MLVU, while Aria obtains 39.2 on HourVideo surpassing the Gemini-1.5-pro by 1.9 points.

This paper explores emotion-aware speech-to-text translation (ST) using generative error correction (GER) by large language models (LLMs). Despite recent advancements in ST, the impact of the emotional content has been overlooked. First, we enhance the translation of emotional speech by adopting the GER paradigm: Finetuned an LLM to generate the translation based on the decoded N-best hypotheses. Moreover, we combine the emotion and sentiment labels into the LLM finetuning process to enable the model to consider the emotion content. In addition, we project the ST model’s latent representation into the LLM embedding space to further improve emotion recognition and translation. Experiments on an English-Chinese dataset show the effectiveness of the combination of GER, emotion and sentiment labels, and the projector for emotion-aware ST. Our code is available at https://github.com/N-Orien/EmoST.

Video Question-Answer-Distractors (QADs) show promising values for assessing the performance of systems in perceiving and comprehending multimedia content. Given the significant cost and labor demands of manual annotation, existing large-scale Video QADs benchmarks are typically generated automatically using video captions. Since video captions are incomplete representations of visual content and susceptible to error propagation, direct generation of QADs from video is crucial. This work first leverages a large vision-language model for video QADs generation. To enhance the consistency and diversity of the generated QADs, we propose utilizing temporal motion to describe the video objects. In addition, We design a selection mechanism that chooses diverse temporal object motions to generate diverse QADs focusing on different objects and interactions, maximizing overall semantic uncertainty for a given video. Evaluation on the NExT-QA and Perception Test benchmarks demonstrates that the proposed approach significantly improves both the consistency and diversity of QADs generated by a range of large vision-language models, thus highlighting its effectiveness and generalizability.

In designing multiple-choice questions (MCQs) in education, creating plausible distractors is crucial for identifying students’ misconceptions and gaps in knowledge and accurately assessing their understanding. However, prior studies on distractor generation have not paid sufficient attention to enhancing the difficulty of distractors, resulting in reduced effectiveness of MCQs. This study presents a pipeline for training a model to generate distractors that are more likely to be selected by students. First, we train a pairwise ranker to reason about students’ misconceptions and assess the relative plausibility of two distractors. Using this model, we create a dataset of pairwise distractor ranks and then train a distractor generator via Direct Preference Optimization (DPO) to generate more plausible distractors. Experiments on computer science subjects (Python, DB, MLDL) demonstrate that our pairwise ranker effectively identifies students’ potential misunderstandings and achieves ranking accuracy comparable to human experts. Furthermore, our distractor generator outperforms several baselines in generating plausible distractors and produces questions with a higher item discrimination index (DI).

Visuals are valuable tools for teaching math word problems (MWPs), helping young learners interpret textual descriptions into mathematical expressions before solving them.However, creating such visuals is labor-intensive and there is a lack of automated methods to support this process. In this paper, we present Math2Visual, an automatic framework for generating pedagogically meaningful visuals from MWP text descriptions. Math2Visual leverages a pre-defined visual language and a design space grounded in interviews with math teachers, to illustrate the core mathematical relationships in MWPs.Using Math2Visual, we construct an annotated dataset of 1,903 visuals and evaluate Text-to-Image (TTI) models for their ability to generate visuals that align with our design. We further fine-tune several TTI models with our dataset, demonstrating improvements in educational visual generation. Our work establishes a new benchmark for automated generation of pedagogically meaningful visuals and offers insights into key challenges in producing multimodal educational content, such as the misrepresentation of mathematical relationships and the omission of essential visual elements.

Knowledge graphs (KGs) have been a cornerstone of search and recommendation due to their ability to store factual knowledge about any domain in a structured form enabling easy search and retrieval. Large language models (LLMs) have shown impressive world knowledge across different benchmarks and domains but their knowledge is inconveniently scattered across their billions of parameters. In this paper, we propose a prompt-based method to construct domain-specific KGs by extracting knowledge solely from LLMs’ parameters. First, we use an LLM to create a schema for a specific domain, which contains a set of domain-representative entities and relations. After that, we use the schema to guide the LLM through an iterative data generation process equipped with Chain-of-Verification (CoVe) for increased data quality. Using this method, we construct KGs for two domains: books and landmarks, which we then evaluate against Wikidata, an open-source human-created KG. Our results show that LLMs can generate large domain-specific KGs containing tens of thousands of entities and relations. However, due to the increased hallucination rates as the procedure evolves, the utility of large-scale LLM-generated KGs in practical applications could remain limited.

Using Large Language Models (LLMs) to generate training data can potentially be a preferable way to improve zero or few-shot NLP tasks. However, many problems remain to be investigated for this direction. For the task of Relation Extraction (RE), we find that samples generated by directly prompting LLMs may easily have high structural similarities with each other. They tend to use a limited variety of phrasing while expressing the relation between a pair of entities. Therefore, in this paper, we study how to effectively improve the diversity of the training samples generated with LLMs for RE, while also maintaining their correctness. We first try to make the LLMs produce dissimilar samples by directly giving instructions in In-Context Learning (ICL) prompts. Then, we propose an approach to fine-tune LLMs for diversity training sample generation through Direct Preference Optimization (DPO). Our experiments on commonly used RE datasets show that both attempts can improve the quality of the generated training data. We also find that comparing with directly performing RE with an LLM, training a non-LLM RE model with its generated samples may lead to better performance.

Improving context faithfulness in large language models is essential for developing trustworthy retrieval augmented generation systems and mitigating hallucinations, especially in long-form question answering (LFQA) tasks or scenarios involving knowledge conflicts. Existing methods either intervene LLMs only at inference without addressing their inherent limitations or overlook the potential for self-improvement. In this paper, we introduce GenDiE(Generate, Discriminate, Evolve), a novel self-evolving framework that enhances context faithfulness through fine-grained sentence-level optimization. GenDiE combines both generative and discriminative training, equipping LLMs with self-generation and self-scoring capabilities to facilitate iterative self-evolution. This supports both data construction for model alignment and score-guided search during inference. Furthermore, by treating each sentence in a response as an independent optimization unit, GenDiE effectively addresses the limitations of previous approaches that optimize at the holistic answer level, which may miss unfaithful details. Experiments on ASQA (in-domain LFQA) and ConFiQA (out-of-domain counterfactual QA) datasets demonstrate that GenDiE surpasses various baselines in both faithfulness and correctness, and exhibits robust performance for domain adaptation.

The spread of fake news on online platforms has long been a pressing concern. Considering this, extensive efforts have been made to develop fake news detectors. However, a major drawback of these models is their relatively low performance—lagging by more than 20%—in identifying *fake* news compared to *real* news, making them less suitable for practical deployment. This gap is likely due to an imbalance in the dataset and the model’s inadequate understanding of data distribution on the targeted platform. In this work, we focus on improving the model’s effectiveness in detecting *fake* news. To achieve this, we **first** adopt an LLM to **generate** fake news in three different styles, which are later incorporated into the training set to augment the representation of fake news. **Then**, we apply Reinforcement Learning to dynamically **sample** fake news, allowing the model to learn the optimal real-to-fake news ratio for training an effective fake news detector on the targeted platform. This approach allows our model to perform effectively even with a limited amount of annotated news data and consistently improve detection accuracy across different platforms. Experimental results demonstrate that our approach achieves state-of-the-art performance on two benchmark datasets, improving *fake* news detection performance by 24.02% and 11.06% respectively.

This paper introduces Generalized Attention Flow (GAF), a novel feature attribution method for Transformer-based models to address the limitations of current approaches. By extending Attention Flow and replacing attention weights with the generalized Information Tensor, GAF integrates attention weights, their gradients, the maximum flow problem, and the barrier method to enhance the performance of feature attributions. The proposed method exhibits key theoretical properties and mitigates the shortcomings of prior techniques that rely solely on simple aggregation of attention weights. Our comprehensive benchmarking on sequence classification tasks demonstrates that a specific variant of GAF consistently outperforms state-of-the-art feature attribution methods in most evaluation settings, providing a more reliable interpretation of Transformer model outputs.

Cognitive distortion is a critical issue in psychology, with most existing studies based on Burns’ cognitive distortion theory. However, differences in annotation standards lead to variations in building analysis tools, resulting in inconsistent analyses and limiting the generalizability of findings, especially in large-scale and cross-linguistic contexts. To address this issue, we collected all publicly available datasets (four in total) and conducted a series of experiments to evaluate the generalizability of various cross-linguistic models. The results indicate that models exhibit significant performance differences across datasets, highlighting the generalization problem. To mitigate this issue, we propose two solutions. First, we propose a multi-task learning model based on teacher student architecture solution, which demonstrates improved generalization performance in our experiments. Second, we introduce a new dataset (~5,000 samples) derived from reannotating existing open datasets to ensure standardized alignment. The annotation process we provided is interpretable and grounded in psychological principles. Based on this, we constructed large language models with cognitive reasoning chains, enhancing both generalizability and interpretability. This study identifies the generalization challenge in cognitive distortion research, and our experiments show that the proposed solutions significantly improve model performance. The dataset and code are publicly available at: https://github.com/HongzhiQ/CrossLinCD.

Large Language Models (LLMs) are prone to generating content that exhibits gender biases, raising significant ethical concerns. Alignment, the process of fine-tuning LLMs to better align with desired behaviors, is recognized as an effective approach to mitigate gender biases. Although proprietary LLMs have made significant strides in mitigating gender bias, their alignment datasets are not publicly available. The commonly used and publicly available alignment dataset, HH-RLHF, still exhibits gender bias to some extent. There is a lack of publicly available alignment datasets specifically designed to address gender bias. Hence, we developed a new dataset named GenderAlign, aiming at mitigating a comprehensive set of gender biases in LLMs. This dataset comprises 8k single-turn dialogues, each paired with a “chosen” and a “rejected” response. Compared to the “rejected” responses, the “chosen” responses demonstrate lower levels of gender bias and higher quality. Furthermore, we categorized the gender biases in the “rejected” responses of GenderAlign into 4 principal categories. The experimental results show the effectiveness of GenderAlign in reducing gender bias in LLMs.

We present a comprehensive evaluation of gender fairness in large language models (LLMs), focusing on their ability to handle both binary and non-binary genders. While previous studies primarily focus on binary gender distinctions, we introduce the Gender Inclusivity Fairness Index (GIFI), a novel and comprehensive metric that quantifies the diverse gender inclusivity of LLMs. GIFI consists of a wide range of evaluations at different levels, from simply probing the model with respect to provided gender pronouns to testing various aspects of model generation and cognitive behaviors under different gender assumptions, revealing biases associated with varying gender identifiers. We conduct extensive evaluations with GIFI on 22 prominent open-source and proprietary LLMs of varying sizes and capabilities, discovering significant variations in LLMs’ gender inclusivity. Our study highlights the importance of improving LLMs’ inclusivity, providing a critical benchmark for future advancements in gender fairness in generative models.

Large Language Models (LLMs) can enhance their capabilities as AI assistants by integrating external tools, allowing them to access a wider range of information. While recent LLMs are typically fine-tuned with tool usage examples during supervised fine-tuning (SFT), questions remain about their ability to develop robust tool-usage skills and can effectively generalize to unseen queries and tools. In this work, we present GenTool, a novel training framework that prepares LLMs for diverse generalization challenges in tool utilization. Our approach addresses two fundamental dimensions critical for real-world applications: Zero-to-One Generalization, enabling the model to address queries initially lacking a suitable tool by adopting and utilizing one when it becomes available, and Weak-to-Strong Generalization, allowing models to leverage enhanced versions of existing tools to solve queries. To achieve this, we develop synthetic training data simulating these two dimensions of tool usage and introduce a two-stage fine-tuning approach: optimizing tool ranking, then refining tool selection. Through extensive experiments across four generalization scenarios, we demonstrate that our method significantly enhances the tool-usage capabilities of LLMs ranging from 1B to 8B parameters, achieving performance that surpasses GPT-4o. Furthermore, our analysis also provides valuable insights into the challenges LLMs encounter in tool generalization.