Austria

Recent advances in mathematical problem-solving with language models (LMs) integrate chain-of-thought (CoT) reasoning and code execution to harness their complementary strengths. However, existing hybrid frameworks exhibit a critical limitation: they depend on externally dictated instructions or rigid code-integration templates, lacking metacognitive awareness—the capacity to dynamically evaluate intrinsic capabilities and autonomously determine when and how to integrate tools. This rigidity motivates our study of autonomous code integration, enabling models to adapt tool-usage strategies as their reasoning abilities evolve during training.

While reinforcement learning (RL) shows promise for boosting LLM reasoning at scale (e.g., DeepSeek-R1), we demonstrate its inefficiency in learning autonomous code integration due to inadequate exploration of the vast combinatorial space of CoT-code interleaving patterns. To address this challenge, we propose a novel Expectation-Maximization (EM) framework that synergizes structured exploration (E-step) with off-policy RL optimization (M-step), creating a self-reinforcing cycle between metacognitive tool-use decisions and evolving capabilities. Experiments reveal our method achieves superior results through improved exploration. Notably, our 7B model improves over 11% on MATH500 and 9.4% on AIME without o1-like CoT.

ACL 2025

Learning Autnomous Code Integration for Math Language Models

tool integration

math reasoning

adaptive

language models

### 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)

poster

Large Language Models (LLMs) often succumb to adversarial prompts, a phenomenon popularly known as ``jailbreaking.'' While jailbreaking primarily targets short-term noncompliance with predefined policies, we argue that a deeper vulnerability lies in altering an LLM's \emph{fundamental axiomatic beliefs}, such as mathematical or philosophical truths. In this work, we introduce \textsc{GoodLiar}, a reinforcement learning (RL)-based framework that generates deceptive contexts to systematically \emph{rewrite} an LLM's core logical or philosophical understandings. By incentivizing an RL agent to produce persuasive and coherent arguments, \textsc{GoodLiar} aims to induce \emph{persistent} belief shifts, rather than merely influencing immediate judgments of factual truthfulness. %rather than one-off policy breaches. Our approach introduces \emph{DA-ILQL}, a novel offline RL method that extends ILQL by integrating on-policy data and language exploration to enhance the language discovery and optimization. Through extensive evaluations on multiple LLMs, we show that deceptive contexts discovered by \textsc{GoodLiar} consistently outperform simple multi-turn prompting methods.

GOODLIAR: A Reinforcement Learning-Based Deceptive Agent for Disrupting LLM Beliefs on Foundational Principles

Large language models are widely used for long-form text generation, but factual errors existing in responses would undermine their reliability. In this work, we systematically investigate the impact of response length on factual precision. First, we introduce an automatic and bi-level long-form factuality evaluation method, which achieves high agreement with human annotations while being cost-effective. Using this framework, we conduct controlled experiments and find that longer responses exhibit lower factual precision, confirming the existence of length bias in long-form factuality. To understand the underlying causes, we empirically examine three hypotheses: error propagation, long context, and facts exhaustion. Our findings reveal that facts exhaustion, where the model gradually exhausts more reliable knowledge, is the primary cause of factual degradation, rather than the other two hypotheses. Our study provides new insights into LLM factuality degradation in extended outputs and suggests future directions for improving long-form factual accuracy.

How Does Generation Length Affect Long-Form Factuality

Stance detection, which aims to identify public opinion towards specific targets using social media data, is an important yet challenging task. With the increasing number of online debates among social media users, conversational stance detection has become a crucial research area. However, existing conversational stance detection datasets are restricted to a limited set of specific targets, which constrains the effectiveness of stance detection models when encountering a large number of unseen targets in real-world applications. To bridge this gap, we manually curate a large-scale, high-quality zero-shot conversational stance detection dataset, named ZS-CSD, comprising 280 targets across two distinct target types. Leveraging the ZS-CSD dataset, we propose SITPCL, a speaker interaction and target-aware prototypical contrastive learning model, and establish the benchmark performance in the zero-shot setting. Experimental results demonstrate that our proposed SITPCL model achieves state-of-the-art performance in zero-shot conversational stance detection. Notably, the SITPCL model attains only an F1-macro score of 43.81%, highlighting the persistent challenges in zero-shot conversational stance detection.

Zero-Shot Conversational Stance Detection: Dataset and Approaches

Claim verification is a critical task in natural language processing (NLP) that involves determining whether available evidence supports or refutes a given claim. While Large Language Models (LLMs) have shown promise in this domain, they still face significant challenges, especially when handling domain-specific knowledge. We propose to enhance LLM-based claim verification through a novel step-by-stepsynthetic data generation framework SynthVerify. Our key insight is that synthetic data generation, guided by domain-specific claim patterns and evidence plans, can effectively address LLMs' knowledge gaps in specialized fields. Using our proposed framework, we create a domain-diverse synthetic claim verification dataset for training zero-shot verification models. We evaluated the fine-tuned LLMs across multiple specialized domains and demonstrate substantial performance gains in verification accuracy compared to the original LLMs, achieving a 20.1 point improvement on the Llama-3-8B model. Our results show that systematically generated synthetic data can effectively address the limitations of current LLM-based verification systems, particularly in handling domain-specific knowledge tasks.

SYNTHVERIFY: Unleashing the Power of LLMs for Zero-Shot Claim Verification via Step-by-Step Synthetic Data Generation

Large Language Models (LLMs) are widely applied to downstream domains. However, current LLMs for high-stakes domain tasks, such as financial investment and legal QA, typically generate brief answers without reasoning processes and explanations. This limits users’ confidence in making decisions based on their responses. While original CoT shows promise, it lacks self-correction mechanisms during reasoning. This work introduces Domaino1s, which enhances LLMs’ reasoning capabilities on domain tasks through supervised fine-tuning and tree search. We construct CoT-stock-2k and CoT-legal-2k datasets for fine-tuning models that activate domain-specific reasoning steps based on their judgment. Additionally, we propose Selective Tree Exploration to spontaneously explore solution spaces and sample optimal reasoning paths to improve performance. We also introduce PROOF-Score, a new metric for evaluating domain models’ explainability, complementing traditional accuracy metrics with richer assessment dimensions. Extensive experiments on stock investment recommendation and legal reasoning QA tasks demonstrate Domaino1s’s leading performance and explainability. Our code is available at https://anonymous.4open.science/r/Domaino1s-006F/.

Domaino1s: Guiding LLM Reasoning for Explainable Answers in High-Stakes Domains

Evaluating generative models with open-ended generation is challenging due to inconsistencies in response formats. Multiple-choice (MC) evaluation mitigates this issue, but generating high-quality distractors is time-consuming and labor-intensive. We introduce D-GEN, the first open-source distractor generator model that transforms open-ended data into an MC format. To evaluate distractor quality, we propose two novel methods: 1) ranking alignment, ensuring generated distractors retain the discriminatory power of ground-truth distractors, and 2) entropy analysis, comparing model confidence distributions. Our results show that D-GEN preserves ranking consistency (Spearman’s 𝜌 0.99, Kendall’s 𝜏 0.94) and closely matches the entropy distribution of ground-truth distractors. Human evaluation further confirms the fluency, coherence, distractiveness, and incorrectness. Our work advances robust and efficient distractor generation with automated evaluation, setting a new standard for MC evaluation.

D-GEN: Automatic Distractor Generation and Evaluation for Reliable Assessment of Generative Models

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.

HammerBench: Fine-Grained Function-Calling Evaluation in Real Mobile Assistant Scenarios

Large Language Models have achieved impressive performance across various natural language generation tasks. However, their lack of inherent constraints limits their effectiveness in applications requiring strict adherence to predefined taxonomies, syntactic structures, and domain-specific rules. Existing approaches, such as fine-tuning and prompting, remain insufficient for ensuring compliance with these requirements, particularly in low-resource scenarios and structured text generation tasks.

To address these limitations, we introduce GRAMMAR-LLM, a novel framework that integrates formal grammatical constraints into the LLM decoding process. GRAMMAR-LLM enforces syntactic correctness in linear time while maintaining expressiveness in grammar rule definition. To achieve this, we propose a new class of grammars, which we name LL(prefix), specifically designed for use with LLMs. These grammars are expressive enough to support common tasks such as hierarchical classification, vocabulary restriction, and structured parsing. We formally prove that LL(prefix) grammars can be transformed into LL(1) grammars in linear time, ensuring efficient processing via deterministic pushdown automata. We evaluate GRAMMAR-LLM across diverse NLP tasks, including hierarchical classification, sign language translation, and Abstract Meaning Representation (AMR) parsing. Our experiments, conducted on models such as LLaMA 3 (for classification and translation) and AMRBART (for parsing), demonstrate that GRAMMAR-LLM consistently improves task performance across zero-shot, few-shot, and fine-tuned settings.

Grammar-Constrained Natural Language Generation

This study presents ***BanStereoSet***, a dataset designed to evaluate stereotypical social biases in multilingual LLMs for the Bangla language. In an effort to extend the focus of bias research beyond English-centric datasets, we have localized the content from the StereoSet, IndiBias, and kamruzzaman-etal's datasets, producing a resource tailored to capture biases prevalent within the Bangla-speaking community. Our BanStereoSet dataset consists of 1,194 sentences spanning 9 categories of bias: race, profession, gender, ageism, beauty, beauty in profession, region, caste, and religion. This dataset not only serves as a crucial tool for measuring bias in multilingual LLMs but also facilitates the exploration of stereotypical bias across different social categories, potentially guiding the development of more equitable language technologies in *Bangladeshi* contexts. Our analysis of several language models using this dataset indicates significant biases, reinforcing the necessity for culturally and linguistically adapted datasets to develop more equitable language technologies.

BanStereoSet: A Dataset to Measure Stereotypical Social Biases in LLMs for Bangla

Multimodal Large Language Models (mLLMs) are trained on a large amount of text-image data. While most mLLMs are trained on caption-like data only, Alayrac et al. (2022) showed that additionally training them on interleaved sequences of text and images can lead to the emergence of in-context learning capabilities. However, the dataset they used, M3W, is not public and is only in English. There have been attempts to reproduce their results but the released datasets are English-only. In contrast, current multilingual and multimodal datasets are either composed of caption-like only or medium-scale or fully private data. This limits mLLM research for the 7,000 other languages spoken in the world. We therefore introduce mOSCAR, to the best of our knowledge the first large-scale multilingual and multimodal document corpus crawled from the web. It covers 163 languages, 303M documents, 200B tokens and 1.15B images. We carefully conduct a set of filtering and evaluation steps to make sure mOSCAR is sufficiently safe, diverse and of good quality. We additionally train two types of multilingual model to prove the benefits of mOSCAR: (1) a model trained on a subset of mOSCAR and captioning data and (2) a model trained on captioning data only. The model additionally trained on mOSCAR shows a strong boost in few-shot learning performance across various multilingual image-text tasks and benchmarks, confirming previous findings for English-only mLLMs. The dataset will be made publicly accessible under the Creative Commons CC BY 4.0 license.

Downloads

Next from ACL 2025

GOODLIAR: A Reinforcement Learning-Based Deceptive Agent for Disrupting LLM Beliefs on Foundational Principles

Similar lecture

One Missing Piece for Open-Source Reasoning Models: A Dataset to Mitigate Cold-Starting Short CoT LLMs in RL