Austria

Long-CoT reasoning combined with reinforcement learning for large language models demonstrates remarkable performance and scalability. However, we observe that the initial policy model could significantly influence the final performance as well as the token efficiency. Additionally, there is a lack of systematic guidelines for obtaining a better initial policy model. To bridge this gap, we initiate a comprehensive investigation by activating the initial model using a variety of datasets with different data volumes and reasoning patterns. Then, we conduct a thorough analysis and comparison of the RL process for different initial models from the perspectives of upper bounds, diversity, and token efficiency, providing a deeper understanding and insight into the long-CoT RL. Based on our empirical results, we propose a systematic guideline and a novel Re-RFT method for constructing a better RL start point. Our experiment results based on the 14B model surpass the DeepSeek-R1-Distill-Qwen-14B by an average of 4.6%, demonstrating our approach&#39;s effectiveness and superiority.

ACL 2025

Towards A Better Initial Policy Model For Scalable Long-CoT Reinforcement Learning

llm; reasoning; reinforcement learning;

### 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)

Long-CoT reasoning combined with reinforcement learning for large language models demonstrates remarkable performance and scalability. However, we observe that the initial policy model could significantly influence the final performance as well as the token efficiency. Additionally, there is a lack of systematic guidelines for obtaining a better initial policy model. To bridge this gap, we initiate a comprehensive investigation by activating the initial model using a variety of datasets with different data volumes and reasoning patterns. Then, we conduct a thorough analysis and comparison of the RL process for different initial models from the perspectives of upper bounds, diversity, and token efficiency, providing a deeper understanding and insight into the long-CoT RL. Based on our empirical results, we propose a systematic guideline and a novel Re-RFT method for constructing a better RL start point. Our experiment results based on the 14B model surpass the DeepSeek-R1-Distill-Qwen-14B by an average of 4.6%, demonstrating our approach's effectiveness and superiority.

poster

Fine-tuning transformer-based models for part-of-speech tagging of unedited Greek text has outperformed traditional systems. However, when applied to lemmatisation or morphological analysis, fine-tuning has not yet achieved competitive results. This paper explores various approaches to combine morphological features to both reduce label complexity and enhance multi-task training. Specifically, we group three nominal features into a single label, and combine the three most distinctive features of verbs into another unified label. These combined labels are used to fine-tune DBBERT, a BERT model pre-trained on both ancient and modern Greek. Additionally, we experiment with joint training -- both among these labels and in combination with POS tagging -- within a multi-task framework to improve performance by transferring parameters. To evaluate our models, we use a manually annotated gold standard from the Database of Byzantine Book Epigrams. Our results show a nearly 9 pp. improvement, demonstrating that multi-task learning is a promising approach for linguistic annotation in less standardised corpora.

Lemmatisation & Morphological Analysis of Unedited Greek: Do Simple Tasks Need Complex Solutions?

The ranked list truncation task involves determining a truncation point to retrieve the relevant items from a ranked list. Despite current advancements, truncation methods struggle with limited capacity, unstable training and inconsistency of selected threshold. To address these problems we introduce TMP Adapter, a novel approach that builds upon the improved adapter model and incorporates the Threshold Margin Penalty (TMP) as an additive loss function to calibrate ranking model relevance scores for ranked list truncation. We evaluate TMP Adapter’s performance on various retrieval datasets and observe that TMP Adapter is a promising advancement in the calibration methods, which offers both theoretical and practical benefits for ranked list truncation.

Relevance Scores Calibration for Ranked List Truncation via TMP Adapter

Large language models (LLMs) have made remarkable progress in various domains, yet they often suffer from repetitive text generation, a phenomenon we refer to as the ''Repeat Curse''. While previous studies have proposed decoding strategies to mitigate repetition, the underlying mechanism behind this issue remains insufficiently explored. In this work, we investigate the root causes of repetition in LLMs through the lens of mechanistic interpretability. Inspired by recent advances in Sparse Autoencoders (SAEs), which enable monosemantic feature extraction, we propose a novel approach—''Duplicatus Charm''—to induce and analyze the Repeat Curse. Our method systematically identifies ``Repetition Features'' -the key model activations responsible for generating repetitive outputs. First, we locate the layers most involved in repetition through logit analysis. Next, we extract and stimulate relevant features using SAE-based activation manipulation. To validate our approach, we construct a repetition dataset covering token and paragraph level repetitions and introduce an evaluation pipeline to quantify the influence of identified repetition features. Furthermore, by deactivating these features, we have effectively mitigated the Repeat Curse.

Understanding the Repeat Curse in Large Language Models from a Feature Perspective

Large Reasoning Models (LRMs) have significantly advanced beyond traditional Large Language Models (LLMs) with their exceptional logical reasoning capabilities, yet these improvements introduce heightened safety risks. When subjected to jailbreak attacks, their ability to generate more targeted and organized content can lead to greater harm. Although some studies claim that reasoning enables safer LRMs against existing LLM attacks, they overlook the inherent flaws within the reasoning process itself. To address this gap, we propose the first jailbreak attack targeting LRMs, exploiting their unique vulnerabilities stemming from the advanced reasoning capabilities. Specifically, we introduce a Chaos Machine, a novel component to transform attack prompts with diverse one-to-one mappings. The chaos mappings iteratively generated by the machine are embedded into the reasoning chain, which strengthens the variability and complexity and also promotes a more robust attack. Based on this, we construct the Mousetrap framework, which makes attacks projected into nonlinear-like low sample spaces with mismatched generalization enhanced. Also, due to the more competing objectives, LRMs gradually maintain the inertia of unpredictable iterative reasoning and fall into our trap. Success rates of the Mousetrap attacking o1-mini, Claude-Sonnet and Gemini-Thinking are as high as 96%, 86% and 98% respectively on our toxic dataset Trotter. On benchmarks such as AdvBench, StrongREJECT, and HarmBench, attacking Claude-Sonnet, well-known for its safety, Mousetrap can astonishingly achieve success rates of 87.5%, 86.58% and 93.13% respectively. Attention: This paper contains inappropriate, offensive and harmful content.

A Mousetrap: Fooling Large Reasoning Models for Jailbreak with Chain of Iterative Chaos

A significant portion of the textual data used in the field of Natural Language Processing (NLP) exhibits gender biases, particularly due to the use of masculine generics (masculine words that are supposed to refer to mixed groups of men and women), which can perpetuate and amplify stereotypes. Gender rewriting, a NLP task that involves automatically detecting and replacing gendered forms with neutral or opposite forms (e.g., from masculine to feminine), can be employed to mitigate these biases. While such systems have been developed in a number of languages (English, Arabic, Portuguese, German, French), automatic use of gender neutralization techniques (as opposed to inclusive or gender-switching techniques) has only been studied for English. This paper presents GeNRe, the very first French gender-neutral rewriting system using collective nouns, which are gender-fixed in French. We introduce a rule-based system (RBS) tailored for the French language alongside two fine-tuned large language models trained on data generated by our RBS. We also explore the use of instruction models to enhance the performance of our other systems and find that Claude 3 Opus combined with our dictionary achieves results close to our RBS. Through this contribution, we hope to promote the advancement of gender bias mitigation techniques in NLP for French.

GeNRe: a French Gender-Neutral Rewriting System Using Collective Nouns

The scientific literature is growing rapidly, making it hard to keep track of the state-of-the-art. Systematic literature reviews (SLRs) aim to identify and evaluate all relevant papers on a topic. After retrieving a set of candidate papers, the abstract screening phase determines initial relevance. To date, abstract screening methods using large language models (LLMs) focus on binary classification settings; existing question answering (QA) based ranking approaches suffer from error propagation. LLMs offer a unique opportunity to evaluate the SLR’s inclusion and exclusion criteria, yet, existing benchmarks do not provide them exhaustively. We manually extract these criteria as well as research questions for 57 SLRs, enabling principled comparisons between approaches. Moreover, we propose LGAR, a zero-shot LLM Guided Abstract Ranker composed of an LLM based graded relevance scorer and a dense re-ranker. Our extensive experiments show that LGAR outperforms existing QA-based methods by 5-10 pp. in mean average precision. Our code and data is publicly available.

LGAR: Zero-Shot LLM-Guided Neural Ranking for Abstract Screening in Systematic Literature Reviews

Natural Language Inference (NLI) has gained significant attention recently due to its importance in understanding how machines comprehend and reason about language. While English has received tremendous interest, Morphologically Rich Languages (MRLs) like Hebrew, require more research. In this paper, we address the evaluation of Hebrew NLI models by introducing LCHAIM, a dataset designed to evaluate these models on tasks involving long premises and complex reasoning. The dataset, created by translating and validating the English ConTRoL dataset, consists of 8,325 context-hypothesis pairs that require coreferential, temporal, logical and analytical reasoning. Our experiments show the difficulty of contextual reasoning in Hebrew, as evidenced by the performance of different models. Fine-tuning the LongHero model on both the shorter premise Hebrew NLI and the LCHAIM datasets yielded a mean accuracy of 52%, that is 35% less than human performance. Similarly, Large language Models (LLMs) like Gemma-9B, Dicta-LM-2.0-7B, and GPT-4o achieved a top mean accuracy of 60.12% in few-shot setting.

LCHAIM - Investigating Long Context Reasoning in Hebrew

As large language models (LLMs) continue to scale up, knowledge editing techniques that modify models' internal knowledge without full retraining have gained significant attention. MEMIT, a prominent batch editing algorithm, stands out for its capability to perform mass knowledge modifications. However, we uncovers a critical limitation that MEMIT's editing efficacy significantly deteriorates when processing batches containing multiple edits sharing the same subject. Our analysis reveals the root cause lies in MEMIT’s key-value modeling framework: when multiple facts with the same subject in a batch are modeled through MEMIT's key-value mechanism, identical keys (derived from the shared subject) are forced to represent different values (corresponding to distinct knowledge), resulting in update conflicts during editing. Addressing this issue, we propose MEMIT-Merge, an enhanced approach that merges value computation processes for facts sharing the same subject, effectively resolving the performance degradation in same-subject batch editing scenarios. Experimental results demonstrate that at a batch size of 5, while the original MEMIT's success rate drops to 46%, MEMIT-Merge maintains a 98% editing success rate, showcasing remarkable robustness to subject entity collisions.

MEMIT-Merge: Addressing MEMIT's Key-Value Conflicts in Same-Subject Batch Editing for LLMs

Knowledge Distillation (KD) has emerged as a prominent technique for model compression. However, conventional KD approaches primarily focus on homogeneous architectures with identical tokenizers, constraining their applicability in cross-architecture scenarios. As for the cross-tokenizer KD, the differences in the tokenizers give rise to two fundamental challenges: (1) sequence misalignment caused by divergent tokenization strategies, and (2) mismatched vocabulary size and composition. While existing probability-matching methods attempt to address these issues, their efficacy remains limited due to suboptimal alignment in both the sequence and vocabulary aspects. To overcome these limitations, we propose Contextual Dynamic Mapping (CDM), a novel cross-tokenizer distillation framework that employs contextual information to enhance sequence alignment precision and dynamically improves vocabulary mapping. We evaluated the effectiveness of our approach across five advanced and widely-used model families ({\em i.e, }LLama3, Phi3, Gemma2, OPT and Qwen2), which were configured into three distinct teacher-student pairs. Our method shows significant advantages over existing cross-tokenizer distillation baselines across diverse benchmarks, including instruction-following, code generation and math. Notably, our analysis reveals that combining conventional same-tokenizer distillation and cross-tokenizer distillation through CDM yields further performance improvements.

Enhancing Cross-Tokenizer Knowledge Distillation with Contextual Dynamical Mapping

Finetuning language models (LMs) is crucial for adapting the models to downstream data and tasks. However, full finetuning is usually costly. Existing work, such as parameter-efficient finetuning (PEFT), often focuses on how to finetune but neglects the issue of where to finetune. As a pioneering work on reducing the cost of backpropagation (at the layer level) by answering where to finetune, we conduct a semantic analysis of the LM inference process. We first propose using transition traces of the latent representation to compute deviations (or loss). Then, using a derived formula of scaling law, we estimate the gain of each layer in reducing deviation (or loss). Further, we narrow down the scope for finetuning, and also, study the cost-benefit balance of LM finetuning. We perform extensive experiments across well-known LMs and datasets. The results show that our approach is effective and efficient, and outperforms the existing baselines. Our approach is orthogonal to other techniques on improving finetuning efficiency, such as PEFT methods, offering practical values on LM finetuning.

Downloads

Next from ACL 2025

Lemmatisation & Morphological Analysis of Unedited Greek: Do Simple Tasks Need Complex Solutions?

Similar lecture

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