Austria

Safety reasoning is a recent paradigm where LLMs reason over safety policies before generating responses, thereby mitigating limitations in existing safety measures such as over-refusal and jailbreak vulnerabilities. However, implementing this paradigm is challenging due to the resource-intensive process of creating high-quality policy-embedded chain-of-thought (CoT) datasets while ensuring reasoning remains accurate and free from hallucinations or policy conflicts. To tackle this, we propose AIDSAFE: Agentic Iterative Deliberation for Safety Reasoning, a novel data generation recipe that leverages multi-agent deliberation to iteratively expand reasoning on safety policies. A data refiner stage in AIDSAFE ensures high-quality outputs by eliminating repetitive, redundant, and deceptive thoughts. AIDSAFE-generated CoTs provide a strong foundation for supervised fine-tuning (SFT)-based safety training. Additionally, to address the need of preference data in alignment stages, such as DPO training, we introduce a supplemental recipe that uses belief augmentation to create distinct selected and rejected CoT samples. Our evaluations demonstrate that AIDSAFE-generated CoTs achieve superior policy adherence and reasoning quality. Consequently, we show that fine-tuning open-source LLMs on these CoTs can significantly improve safety generalization and jailbreak robustness while maintaining acceptable utility and over-refusal accuracy.

ACL 2025

Towards Safety Reasoning in LLMs: AI-agentic Deliberation for Policy-embedded CoT Data Creation

cot data

agentic deliberation

safety reasoning

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

poster

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Improvements in large language models have led to increasing optimism that they can serve as reliable evaluators of natural language generation outputs. In this paper, we challenge this optimism by thoroughly re-evaluating five state-of-the-art factuality metrics on a collection of 11 datasets for summarization, retrieval-augmented generation, and question answering. We find that these evaluators are inconsistent with each other and often misestimate system-level performance, both of which can lead to a variety of pitfalls. We further show that these metrics exhibit biases against highly paraphrased outputs and outputs that draw upon faraway parts of the source documents. We urge users of these factuality metrics to proceed with caution and manually validate the reliability of these metrics in their domain of interest before proceeding.

Verify with Caution: The Pitfalls of Relying on Imperfect Factuality Metrics

Evaluating reference tables qualitatively and quantitatively is a challenging problem, as traditional metrics often fail to capture the nuanced discrepancies in both structure and content. To mitigate this, we first introduce a methodical rubric that combines multi-level structural descriptors with fine-grained contextual quantification, laying a robust foundation for comprehensive table comparisons. Building on this, we propose **TabXEval**, an eXhaustive and eXplainable two-phase evaluation framework that initially aligns reference tables structurally via TabAlign and subsequently conducts a systematic semantic and syntactic comparison using TabCompare. This approach not only clarifies the evaluation process but also pinpoints subtle discrepancies that conventional methods overlook. To validate our framework, we present **TabXBench**, a diverse, multi-domain benchmark featuring realistic table perturbations and human-annotated assessments. Finally, we provide a systematic analysis of existing evaluation methods through sensitivity-specificity trade-offs and demonstrate the qualitative and quantitative effectiveness of **TabXEval** on text-to-table generation tasks, paving the way for future innovations in explainable table evaluation.

TabXEval: Why this is a Bad Table? An eXhaustive Rubric for Table Evaluation

Large Language Models (LLMs) recently achieved great success in medical text summarization by simply using in-context learning. However, these recent efforts do not perform fine-grained evaluations under difficult settings where LLMs might fail. They typically report performance scores over the entire dataset. Through our benchmarking study, we show that LLMs show a significant performance drop for data points with high concentration of out-of-vocabulary (OOV) words or with high novelty. Vocabulary adaptation is an intuitive solution to this vocabulary mismatch issue where the LLM vocabulary gets updated with certain expert domain (here, medical) words or subwords. An interesting finding from our study is that Llama-3.1, even with a vocabulary size of around 128K tokens, still faces _over-fragmentation_ issue with medical words. To that end, we show vocabulary adaptation helps improve the LLM summarization performance even in difficult settings. Through extensive experimentation of multiple vocabulary adaptation strategies, two continual pretraining strategies, and three benchmark medical summarization datasets, we gain valuable insights into the role of vocabulary adaptation strategies for customizing LLMs to the medical domain. We also performed a human evaluation study with medical experts where they found that vocabulary adaptation results in more relevant and faithful summaries. Our codebase is made publicly available at https://github.com/gb-kgp/LLM-MedicalSummarization-Benchmark.

Evaluation of LLMs in Medical Text Summarization: The Role of Vocabulary Adaptation in High OOV Settings

Writing is inherently iterative, each revision enhancing information representation. One revision may contain many edits. Examination of the intentions behind edits provides valuable insights into an editor’s expertise, the dynamics of collaborative writing, and the evolution of a document. Current research on edit intentions lacks a comprehensive edit intention taxonomy (EIT) that spans multiple application domains. As a result, researchers often create new EITs tailored to specific needs, a process that is both time-consuming and costly. To address this gap, we propose UniT, a Unified edit intention Taxonomy that integrates existing EITs encompassing a wide range of edit intentions. We examine the lineage relationship and the construction of 24 EITs. They together have 232 categories across various domains. During the literature survey and integration process, we identify challenges such as one-to-many category matches, incomplete definitions, and varying hierarchical structures. We propose solutions for resolving these issues. Finally, our evaluation shows that our UniT achieves higher inter-annotator agreement scores compared to existing EITs and is applicable to a large set of application domains.

UniT: One Document, Many Revisions, Too Many Edit Intention Taxonomies

Detecting toxicity in online multimodal environments such as memes remains a challenging task due to the complex contextual connections across modalities (e.g., text and visual), demanding both common sense reasoning and contextual awareness. To bridge this gap, we propose a hybrid neurosymbolic framework that unifies (1) distillation of implicit contextual knowledge (e.g., sarcasm, cultural references) from Large Vision-Language Models (LVLMs) and (2) infusion of explicit relational semantics through sub-graphs from Knowledge Graphs (KGs). Experimental results on two benchmark datasets show the superior performance of our approach, Knowledge-Infused Distilled Vision-Language Model (KID-VLM), over the state-of-the-art baselines across AUC, and F1, with improvements of 0.5%, and 10.6%, respectively in HatefulMemes Benchmark across variants. Further, KID-VLM demonstrates better generalizability and achieves best performance across all baselines in HarMeme Dataset with an 6.3% and 3.2% in F1 and AUC. Given the contextual complexity of the toxicity detection task, KID-VLM showcases the significance of learning compact models (~500M parameters) from both explicit (i.e., KG) and implicit (i.e., LVLMs) contextual cues incorporated through a hybrid neurosymbolic approach. KID-VLM provides an accurate and scalable approach to recognize toxic content for safer online environments in real-world applications. Our code is available at Anonymous Github at: https://anonymous.4open.science/r/MultimodalToxicity-124F/README.md.

Just KIDDIN’ : Knowledge Infusion and Distillation for Detection of INdecent Memes

Open-world planning poses a significant challenge for general artificial intelligence due to environmental complexity and task diversity, especially in long-term tasks and lifelong learning. Inspired by cognitive theories, we propose M2PA, an open-world multi-memory planning agent. M2PA innovates by combining Large Language Models (LLMs) with human-like multi-memory systems, aiming to fully leverage the strengths of both while mitigating their respective limitations. By integrating the expansive world knowledge and language processing capabilities of LLMs with the perception and experience accumulation abilities of the human memory system, M2PA exhibits situation awareness, and experience generalization capabilities, as well as the potential for lifelong learning. In experiments, M2PA significantly outperforms current state-of-the-art agents across 50 Minecraft tasks in zero-shot learning. In exploratory lifelong learning experiments, M2PA demonstrates its continuous learning ability, achieving a \textbf{38.33%} success rate in the “ObtainDiamond” task. Our findings provide a novel paradigm for constructing more effective agents in open-world environments.

M2PA: A Multi-Memory Planning Agent for Open Worlds Inspired by Cognitive Theory

Decompilers are fundamental tools for critical security tasks, from vulnerability discovery to malware analysis, yet their evaluation remains fragmented. Existing approaches primarily focus on syntactic correctness through synthetic micro-benchmarks or subjective human ratings, failing to address real-world requirements for semantic fidelity and analyst usability. We present **DecompileBench**, the first comprehensive framework that enables effective evaluation of decompilers in reverse engineering workflows through three key components: real-world function extraction (comprising 23,400 functions from 130 real-world programs), runtime-aware validation, and automated human-centric assessment using LLM-as-Judge to quantify the effectiveness of decompilers in reverse engineering workflows. Through a systematic comparison between six industrial-strength decompilers and six recent LLM-powered approaches, we demonstrate that LLM-based methods surpass commercial tools in code understandability despite 52.2% lower functionality correctness. These findings highlight the potential of LLM-based approaches to transform human-centric reverse engineering. We open source **DecompileBench** to provide a framework to advance research on decompilers and assist security experts in making informed tool selections based on their specific requirements.

DecompileBench: A Comprehensive Benchmark for Evaluating Decompilers in Real-World Scenarios

Code generation is crucial in software engineering for automating the coding process efficiently. 
While test-time computation methods show promise, they suffer from high latency due to multiple computation rounds.
To overcome this, we introduce \textbf{ThinkCoder}, a framework that combines thorough exploration with optimal refinement.
The exploration phase diversifies the solution space by searching for potential solutions, followed by a refinement phase that enhances precision.
This approach allows us to select the best solution through careful consideration before taking action, avoiding excessive trial and error.
To further minimize test-time computation overhead, we introduce preference-driven optimization with Reinforced Self-Training (ReST), which uses exploration trajectories from ThinkCoder to guide LLM's evolution.
This approach enhances LLM's exploration efficiency via preference learning, cutting costs while maintaining accuracy.
ThinkCoder boosts the performance with a single LLM, excelling on benchmarks like HumanEval and MBPP. 
Compared to SOTA models, it improves Pass@1 by 3.0\% over MapCoder with just 6.4\% of the computation cost.
Against AgentCoder, ThinkCoder achieves a 0.5\% higher Pass@1 after 2 rounds, outperforming AgentCoder's 5 rounds.
Additionally, ReST with success trajectories enhances efficiency, allowing models like LLaMA2-7B to achieve competitive results using only 20\% of the computational resources. 
These results highlight the framework's effectiveness and scalability. 
\footnote{The code is available at \url{https://github.com/xiaoqzhwhu/ThinkCoder}}

Thinking Before Running! Efficient Code Generation with Thorough Exploration and Optimal Refinement

Hallucination is a persistent challenge in large language models (LLMs), where even with rigorous quality control, models often generate distorted facts. This paradox, in which error generation continues despite high-quality training data, calls for a deeper understanding of the underlying LLM mechanisms. To address it, we propose a novel concept: knowledge overshadowing, where model’s dominant knowledge can obscure less prominent knowledge during text generation, causing the model to fabricate inaccurate details. Building on this idea, we introduce a novel framework to quantify factual hallucinations by modeling knowledge overshadowing. Central to our approach is the log-linear law, which predicts that the rate of factual hallucination increases linearly with the logarithmic scale of (1) Knowledge Popularity, (2) Knowledge Length, and (3) Model Size. The law provides a means to preemptively quantify hallucinations, offering foresight into their occurrence even before model training or inference. Built on overshadowing effect, we propose a new decoding strategy CoDa, to mitigate hallucinations, which notably enhance model factuality on Overshadow (27.9%), MemoTrap (13.1%) and NQ-Swap (18.3%). Our findings not only deepen understandings of the underlying mechanisms behind hallucinations but also provide actionable insights for developing more predictable and controllable language models.

The Law of Knowledge Overshadowing: Towards Understanding, Predicting and Preventing LLM Hallucination

Large Language Models (LLMs) have demonstrated impressive performances in tasks related to coreference resolution. However, previous studies mostly assessed LLM performance on coreference resolution with nouns and third person pronouns. This study evaluates LLM performance on coreference resolution with indexical like I, you, here and tomorrow} which come with unique challenges due to their linguistic properties. We present the first study examining how LLMs interpret indexicals in English, releasing the English Indexical Dataset with 1600 multiple-choice questions. We evaluate pioneering LLMs, including GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, and DeepSeek V3. Our results reveal that LLMs exhibit an impressive performance with some indexicals (I), while struggling with others (you, here, tomorrow), and that syntactic cues (e.g. quotation) contribute to LLM performance with some indexicals, while they reduce performance with others.

Downloads

Next from ACL 2025

Verify with Caution: The Pitfalls of Relying on Imperfect Factuality Metrics

Similar lecture

Reasoning-to-Defend: Safety-Aware Reasoning Can Defend Large Language Models from Jailbreaking