ACL 2025 Papers

Recent advances in neural topic models (NTMs) have improved topic quality but still face challenges: weak document-topic alignment, high inference costs due to large pretrained language models (PLMs), and limited modeling of hierarchical topic structures. To address these issues, we introduce HiCOT (Hierarchical Clustering and Contrastive Learning with Optimal Transport for Neural Topic Modeling), a novel framework that enhances topic coherence and efficiency. HiCOT integrates Optimal Transport to refine document-topic relationships using compact PLM-based embeddings, captures semantic structure of the documents. Additionally, it employs hierarchical clustering combine with contrastive learning to disentangle topic-word and topic-topic relationships, ensuring clearer structure and better coherence. Experimental results on multiple benchmark datasets demonstrate HiCOT’s superior effectiveness over existing NTMs in topic coherence, topic performance, representation quality, and computational efficiency.

Large Language Model (LLM)-based agents exhibit significant potential across various domains, operating as interactive systems that process environmental observations to generate executable actions for target tasks. The effectiveness of these agents is significantly influenced by their memory mechanism, which records historical experiences as sequences of action-observation pairs. We categorize memory into two types: cross-trial memory, accumulated across multiple attempts, and in-trial memory (working memory), accumulated within a single attempt. While considerable research has optimized performance through cross-trial memory, the enhancement of agent performance through improved working memory utilization remains underexplored. Instead, existing approaches often involve directly inputting entire historical action-observation pairs into LLMs, leading to redundancy in long-horizon tasks. Inspired by human problem-solving strategies, this paper introduces HiAgent, a framework that leverages subgoals as memory chunks to manage the working memory of LLM-based agents hierarchically. Specifically, HiAgent prompts LLMs to formulate subgoals before generating executable actions and enables LLMs to decide proactively to replace previous subgoals with summarized observations, retaining only the action-observation pairs relevant to the current subgoal. Experimental results across five long-horizon tasks demonstrate that HiAgent achieves a twofold increase in success rate and reduces the average number of steps required by 3.8. Additionally, our analysis shows that HiAgent consistently improves performance across various steps, highlighting its robustness and generalizability. Code is available in this URL: https://github.com/HiAgent2024/HiAgent

Recent advances in Large Language Models(LLMs) have led to remarkable achievements across a variety of Natural Language Processing(NLP) tasks, making prompt engineering increasingly central to guiding model outputs. While manual methods (e.g., “chain-of-thought,” “step-by-step” prompts) can be effective, they typically rely on intuition and do not automatically refine prompts over time. In contrast, automatic prompt optimization employing heuristic-based search algorithms can systematically explore and improve prompts with minimal human oversight. This survey proposes a comprehensive taxonomy of these methods, categorizing them by where optimization occurs, what is optimized, what criteria drive the optimization, which operators generate new prompts, and which iterative search algorithms are applied. We further highlight specialized datasets and tools that support and accelerate automated prompt refinement. We conclude by discussing key open challenges, pointing toward future opportunities for more robust and versatile LLM applications.

Inference-Time Scaling has been critical to the success of recent models such as OpenAI o1 and DeepSeek R1. However, many techniques used to train models for inference-time scaling require tasks to have answers that can be verified, limiting their application to domains such as math, coding and logical reasoning. We take inspiration from how humans make first attempts, ask for detailed feedback from others and make improvements based on such feedback across a wide spectrum of open-ended endeavors. To this end, we collect HelpSteer3 data to train dedicated Feedback and Edit Models that are capable of performing inference-time scaling for open-ended general-domain tasks. In our setup, one model generates an initial response, which are given feedback by a second model, that are then used by a third model to edit the response. We show that performance on Arena Hard, a benchmark strongly predictive of Chatbot Arena Elo can be boosted by scaling the number of initial response drafts, effective feedback and edited responses. When scaled optimally, our setup based on 70B models from the Llama 3 family can reach SoTA performance on Arena Hard at 92.7 as of 5 Mar 2025, surpassing OpenAI o1-preview-2024-09-12 with 90.4 and DeepSeek R1 with 92.3.

Can LLMs provide support to creative writers by giving meaningful writing feedback? In this paper, we explore the challenges and limitations of model-generated writing feedback by defining a new task, dataset, and evaluation frameworks. To study model performance in a controlled manner, we present a novel test set of 1,300 stories that we corrupted to intentionally introduce writing issues. We study the performance of commonly used LLMs in this task with both automatic and human evaluation metrics. Our analysis shows that current models have strong out-of-the-box behavior in many respects—providing specific and mostly accurate writing feedback. However, models often fail to identify the biggest writing issue in the story and to correctly decide when to offer critical vs. positive feedback.

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.

Large language models (LLMs) have demonstrated exceptional performance in text generation within current NLP research. However, the lack of factual accuracy is still a dark cloud hanging over the LLM skyscraper. Structural knowledge prompting (SKP) is a prominent paradigm to integrate external knowledge into LLMs by incorporating structural representations, achieving state-of-the-art results in many knowledge-intensive tasks. However, existing methods often focus on specific problems, lacking a comprehensive exploration of the generalization and capability boundaries of SKP. This paper aims to evaluate and rethink the generalization capability of the SKP paradigm from four perspectives including Granularity, Transferability, Scalability, and Universality. To provide a thorough evaluation, we introduce a novel multi-granular, multi-level benchmark called SUBARU, consisting of 9 different tasks with varying levels of granularity and difficulty. Through extensive experiments, we draw key conclusions regarding the generalization of SKP, offering insights to guide the future development and extension of the SKP paradigm.

Language changes over time, including in the hate speech domain, which evolves quickly following social dynamics and cultural shifts. While NLP research has investigated the impact of language evolution on model training and has proposed several solutions for it, its impact on model benchmarking remains under-explored. Yet, hate speech benchmarks play a crucial role to ensure model safety. In this paper, we empirically evaluate the robustness of 20 language models across two evolving hate speech experiments, and we show the temporal misalignment between static and time-sensitive evaluations. Our findings call for time-sensitive linguistic benchmarks in order to correctly and reliably evaluate language models in the hate speech domain.

Despite regulations imposed by nations and social media platforms, e.g. (Government of India, 2021; European Parliament and Council of the European Union, 2022), inter alia, hateful content persists as a significant challenge. Existing approaches primarily rely on reactive measures such as blocking or suspending offensive messages, with emerging strategies focusing on proactive measurements like detoxification and counterspeech. In our work, which we call HATEPRISM, we conduct a comprehensive examination of hate speech regulations and strategies from three perspectives: country regulations, social platform policies, and NLP research datasets. Our findings reveal significant inconsistencies in hate speech definitions and moderation practices across jurisdictions and platforms, alongside a lack of alignment with research efforts. Based on these insights, we suggest ideas and research direction for further exploration of a unified framework for automated hate speech moderation incorporating diverse strategies.

To address the global challenge of online hate speech, prior research has developed detection models to flag such content on social media. However, due to systematic biases in evaluation datasets, the real-world effectiveness of these models remains unclear, particularly across geographies. We introduce HateDay, the first global hate speech dataset representative of social media settings, constructed from a random sample of all tweets posted on September 21, 2022 and covering eight languages and four English-speaking countries. Using HateDay, we uncover substantial variation in the prevalence and composition of hate speech across languages and regions. We show that evaluations on academic datasets greatly overestimate real-world detection performance, which we find is very low, especially for non-European languages. Our analysis identifies key drivers of this gap, including models’ difficulty to distinguish hate from offensive speech and a mismatch between the target groups emphasized in academic datasets and those most frequently targeted in real-world settings. We argue that poor model performance makes public models ill-suited for automatic hate speech moderation and find that high moderation rates are only achievable with substantial human oversight. Our results underscore the need to evaluate detection systems on data that reflects the complexity and diversity of real-world social media.

In Machine Translation (MT) evaluation, metric performance is assessed based on agreement with human judgments. In recent years, automatic metrics have demonstrated increasingly high levels of agreement with humans. To gain a clearer understanding of metric performance and establish an upper bound, we incorporate human baselines in the MT meta-evaluation, that is, the assessment of MT metrics’ capabilities. Our results show that human annotators are not consistently superior to automatic metrics, with state-of-the-art metrics often ranking on par with or higher than human baselines. Despite these findings suggesting human parity, we discuss several reasons for caution. Finally, we explore the broader implications of our results for the research field, asking: Can we still reliably measure improvements in MT evaluation? With this work, we aim to shed light on the limits of our ability to measure progress in the field, fostering discussion on an issue that we believe is crucial to the entire MT evaluation community.

Prosody affects how people produce and understand language, yet studies of how it does so have been hindered by the lack of efficient tools for analyzing prosodic stress. We fine-tune OpenAI Whisper large-v2, a state-of-the-art speech recognition model, to recognize phrasal, lexical, and contrastive stress using a small, carefully annotated dataset. Our results show that Whisper can learn distinct, gender-specific stress patterns to achieve near-human and super-human accuracy in stress classification and transfer its learning from one type of stress to another, surpassing traditional machine learning models. Furthermore, we explore how acoustic context influences its performance and propose a novel black-box evaluation method for characterizing the decision boundaries used by Whisper for prosodic stress interpretation. These findings open new avenues for large-scale, automated prosody research. Models can be found at github.com/SSSohn/ProsodyBench.

Large Multimodal Models (LMMs) have demonstrated strong performance in English, but their effectiveness in Japanese remains limited due to the lack of high-quality training data. Current Japanese LMMs often rely on translated English datasets, restricting their ability to capture Japan-specific cultural knowledge. To address this, we explore the potential of Japanese PDF data as a training resource, an area that remains largely underutilized. We introduce a fully automated pipeline that leverages pretrained models to extract image-text pairs from PDFs through layout analysis, OCR, and vision-language pairing, removing the need for manual annotation. Additionally, we construct instruction data from extracted image-text pairs to enrich the training data. To evaluate the effectiveness of PDF-derived data, we train Japanese LMMs and assess their performance on the Japanese LMM Benchmark. Our results demonstrate substantial improvements, with performance gains ranging from 2.1% to 13.8% on Heron-Bench. Further analysis highlights the impact of PDF-derived data on various factors, such as model size and language models, reinforcing its value as a multimodal resource for Japanese LMMs.

Large Language Models (LLMs) have demonstrated remarkable capabilities across various domains, including their emerging role in mitigating threats to human life, infrastructure, and the environment during natural disasters. Despite increasing research on disaster-focused LLMs, there remains a lack of systematic reviews and in-depth analyses of their applications in natural disaster management. To address this gap, this paper presents a comprehensive survey of LLMs in disaster response, introducing a taxonomy that categorizes existing works based on disaster phases and application scenarios. By compiling public datasets and identifying key challenges and opportunities, this study aims to provide valuable insights for the research community and practitioners in developing advanced LLM-driven solutions to enhance resilience against natural disasters.

During a conversation, there can come certain moments where its outcome hangs in the balance. In these pivotal moments, how one responds can put the conversation on substantially different trajectories leading to significantly different outcomes. Systems that can detect when such moments arise could assist conversationalists in domains with highly consequential outcomes, such as mental health crisis counseling.In this work, we introduce an unsupervised computational method for detecting such pivotal moments as they happen. The intuition is that a moment is pivotal if our expectation of the conversation’s outcome varies widely depending on what might be said next. By applying our method to crisis counseling conversations, we first validate it by showing that it aligns with human perception—counselors take significantly longer to respond during moments detected by our method—and with the eventual conversational trajectory—which is more likely to change course at these times. We then use our framework to explore the relation between the counselor’s response during pivotal moments and the eventual outcome of the session.

Evaluating the performance of LLMs in multi-turn human-agent interactions presents significant challenges, particularly due to the complexity and variability of user behavior. In this paper, we introduce HammerBench, a novel benchmark framework for assessing LLMs’ function-calling capabilities in real-world, multi-turn dialogues. HammerBench simulates diverse mobile assistant use cases, incorporating imperfect instructions, dynamic question-answer trajectories, intent and argument shifts, and the indirect use of external information through pronouns. To construct this benchmark, we curate a comprehensive dataset derived from popular mobile app functionalities and anonymized user logs, complemented by a cost-effective data generation pipeline leveraging open-source models. HammerBench is further augmented with fine-grained interaction snapshots and metrics, enabling detailed evaluation of function-calling performance across individual conversational turns. We demonstrate the effectiveness of HammerBench by evaluating several leading LLMs and uncovering key performance trends. Our experiments reveal that different types of parameter name errors are a significant source of failure across different interaction scenarios, highlighting critical areas for further improvement in LLM robustness for mobile assistant applications.

As large language models (LLMs) become increasingly prevalent, concerns about their reliability, particularly due to hallucinations - factually inaccurate or irrelevant outputs - have grown. Our research investigates the relationship between the uncertainty in training dynamics and the emergence of hallucinations. Using models from the Pythia suite and several hallucination detection metrics, we analyze hallucination trends and identify significant variance during training. To address this, we propose Sensitivity Dropout (SenD), a novel training protocol designed to reduce hallucination variance during training by deterministically dropping embedding indices with significant variability. In addition, we develop an unsupervised hallucination detection metric, Efficient EigenScore (EES), which approximates the traditional EigenScore in 2x speed. This metric is integrated into our training protocol, allowing SenD to be both computationally scalable and effective at reducing hallucination variance. SenD improves test-time reliability of Pythia and Meta’s Llama models by up to 17% and enhances factual accuracy in Wikipedia, Medical, Legal, and Coding domains without affecting downstream task performance.

Large language models (LLMs) often generate responses that deviate from user input or training data, a phenomenon known as “hallucination.” These hallucinations undermine user trust and hinder the adoption of generative AI systems. Addressing hallucinations is important for the advancement of LLMs. This paper introduces a comprehensive hallucination benchmark HalluLens, incorporating both extrinsic and intrinsic evaluation tasks, built upon a clear taxonomy of hallucination. A major challenge in benchmarking hallucinations is the lack of a unified framework due to inconsistent definitions and categorizations. We disentangle LLM hallucination from “factuality” and propose a taxonomy distinguishing extrinsic and intrinsic hallucinations to promote consistency and facilitate research. We emphasize extrinsic hallucinations – where generated content deviates from training data – as they become increasingly relevant with LLM advancements. However, no benchmark is solely dedicated to extrinsic hallucinations. To address this gap, HalluLens introduces three new extrinsic tasks with dynamic test set generation to mitigate data leakage and ensure robustness. We release codebase for extrinsic hallucination benchmark.

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.