Austria

Model distillation is a fundamental technique in building large language models (LLMs), transferring knowledge from a teacher model to a student model. However, distillation can lead to model homogenization, reducing diversity among models and impairing their ability to robustly handle complex or novel tasks. These limitations underscore the need to systematically quantify the distillation process and its impact. In this work, we propose a framework to evaluate and quantify model distillation. Our method addresses two key aspects: (1) Identifying identity cognition contradictions to assess discrepancies in how models perceive and represent identity-related information, and (2) Analyzing multi-granularity response similarities across models to measure the extent of homogenization. Experimental results demonstrate two key insights: (1) Well-known closed-source and open-source LLMs usually exhibit high distillation degrees, except for Claude, Doubao, and Gemini. (2) Base LLMs show higher distillation degrees compared to aligned LLMs. By offering a systematic approach to improve the transparency of LLM data distillation, we call for LLMs with more independent development and more transparent technical reports to improve LLMs&#39; robustness and safety. The code and data are available at https://github.com/Aegis1863/LLMs-Distillation-Quantification.

ACL 2025

Quantification of Large Language Model Distillation

jailbreak

quantification

distillation

### 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)

Model distillation is a fundamental technique in building large language models (LLMs), transferring knowledge from a teacher model to a student model. However, distillation can lead to model homogenization, reducing diversity among models and impairing their ability to robustly handle complex or novel tasks. These limitations underscore the need to systematically quantify the distillation process and its impact. In this work, we propose a framework to evaluate and quantify model distillation. Our method addresses two key aspects: (1) Identifying identity cognition contradictions to assess discrepancies in how models perceive and represent identity-related information, and (2) Analyzing multi-granularity response similarities across models to measure the extent of homogenization. Experimental results demonstrate two key insights: (1) Well-known closed-source and open-source LLMs usually exhibit high distillation degrees, except for Claude, Doubao, and Gemini. (2) Base LLMs show higher distillation degrees compared to aligned LLMs. By offering a systematic approach to improve the transparency of LLM data distillation, we call for LLMs with more independent development and more transparent technical reports to improve LLMs' robustness and safety. The code and data are available at https://github.com/Aegis1863/LLMs-Distillation-Quantification.

poster

Recent advancements in Large Language Models (LLMs) have demonstrated enhanced reasoning capabilities, evolving from Chain-of-Thought (CoT) prompting to advanced, product-oriented solutions like OpenAI o1. During our re-implementation of this model, we noticed that in multimodal tasks requiring visual input (e.g., geometry problems), Multimodal LLMs (MLLMs) struggle to maintain focus on the visual information, in other words, MLLMs suffer from a gradual decline in attention to visual information as reasoning progresses, causing text-over-relied outputs. To investigate this, we ablate image inputs during long-chain reasoning. Concretely, we truncate the reasoning process midway, then re-complete the reasoning process with the input image removed. We observe only a ~2 accuracy drop on MathVista’s test-hard subset, revealing the model's textual outputs dominate the following reasoning process. Motivated by this, we propose Take-along Visual Conditioning (TVC), a strategy that shifts image input to critical reasoning stages and compresses redundant visual tokens via dynamic pruning. This methodology helps the model retain attention to the visual components throughout the reasoning. Our approach achieves state-of-the-art performance on average across five mathematical reasoning benchmarks (+3.4% vs previous sota), demonstrating the effectiveness of TVC in enhancing multimodal reasoning systems. The project page is available at \url{https://sun-hailong.github.io/projects/TVC}.

Mitigating Visual Forgetting via Take-along Visual Conditioning for Multi-modal Long CoT Reasoning

Large language models (LLMs) often exhibit systematic preferences for certain answer choices when responding to multiple-choice questions—a behavior known as selection bias. This bias reduces the accuracy and reliability of LLM outputs, limiting their usefulness in decision-critical applications. While prior work has focused on adjusting model inputs or outputs to mitigate this issue, our work takes a fundamentally different approach by identifying and removing the internal sources of bias. We introduce two methods: Bias Node Pruning (BNP), which prunes parameters that contribute to selection bias, and Auxiliary Option Injection (AOI), which introduces an additional answer choice to reduce bias in both white-box and black-box settings. To address the shortcomings of existing evaluation metrics, we propose Choice Kullback-Leibler Divergence (CKLD), a new metric that captures distributional imbalances in model predictions. Experiments on three LLMs across multiple datasets demonstrate that our methods consistently improve answer accuracy while reducing selection bias, providing a robust solution for both open- and closed-source models.

Mitigating Selection Bias with Node Pruning and Auxiliary Options

While Retrieval-Augmented Generation (RAG) has exhibited promise in utilizing external knowledge, its generation process heavily depends on the quality and accuracy of the retrieved context. Large language models (LLMs) struggle to evaluate the correctness of non-parametric knowledge retrieved externally when it differs from internal memorization, leading to knowledge conflicts during response generation. To this end, we introduce the **R**etrieval **P**reference **O**ptimization (RPO), a lightweight and effective alignment method to adaptively leverage multi-source knowledge based on retrieval relevance. An implicit representation of retrieval relevance is derived and incorporated into the reward model to integrate retrieval evaluation and response generation into a single model, solving the problem that previous methods necessitate the additional procedure to assess the retrieval quality. Notably, RPO is a RAG-dedicated alignment approach that quantifies the awareness of retrieval relevance in training, first overcoming mathematical obstacles. Experiments on four datasets demonstrate that RPO outperforms RAG by 4-10% in accuracy without any extra component, exhibiting its robust generalization.

RPO: Retrieval Preference Optimization for Robust Retrieval-Augmented Generation

Large language models (LLMs) are renowned for their extensive linguistic knowledge and strong generalization capabilities, but their high computational demands make them unsuitable for resource-constrained environments. In contrast, small language models (SLMs) are computationally efficient but often lack the broad generalization capacity of LLMs. To bridge this gap, we propose PiFi, a novel framework that combines the strengths of both LLMs and SLMs to achieve high performance while maintaining efficiency. PiFi integrates a single frozen layer from an LLM into a SLM and fine-tunes the combined model for specific tasks, boosting performance without a significant increase in computational cost. We show that PiFi delivers consistent performance improvements across a range of natural language processing tasks, including both natural language understanding and generation. Moreover, our findings demonstrate PiFi’s ability to effectively leverage LLM knowledge, enhancing generalization to unseen domains and facilitating the transfer of linguistic abilities.

Plug-in and Fine-tuning: Bridging the Gap between Small Language Models and Large Language Models

Speculative decoding is a powerful technique that accelerates Large Language Model (LLM) inference by leveraging a lightweight speculative draft model. However, existing designs suffers in performance due to misalignment between training and inference. Recent methods have tried to solve this issue by adopting a multi-step training strategy, but the complex inputs of different training steps make it harder for the draft model to converge. To address this, we propose CORAL, a novel framework that improves both accuracy and efficiency in speculative drafting. CORAL introduces Cross-Step Representation Alignment, a method that enhances consistency across multiple training steps, significantly improving speculative drafting performance. Additionally, we identify the LM head as a major bottleneck in the inference speed of the draft model. We introduce a weight-grouping mechanism that selectively activates a subset of LM head parameters during inference, substantially reducing the latency of the draft model. We evaluate CORAL on three LLM families and three benchmark datasets, achieving speedup ratios of 2.50x-4.07x, outperforming state-of-the-art methods such as EAGLE-2 and HASS. Our results demonstrate that CORAL effectively mitigates training-inference misalignment and delivers significant speedup for modern LLMs with large vocabularies.

CORAL: Learning Consistent Representations across Multi-step Training with Lighter Speculative Drafter

In Reinforcement Learning from Human Feedback (RLHF), the reward model (RM) evaluates the response quality based on the given context and assigns a reward. It plays a crucial role in aligning RLHF with human preferences. Although the current RM training paradigm concatenates the context and response while amplifying the reward difference between good and bad response pairs, we demonstrate that the RM faces two significant issues: i) it often allocates only a small proportion of attention to the context, and ii) it frequently ignores segments of the context that are relevant for evaluating the response quality. These issues undermine the RM's effectiveness in modeling human preferences. To further address these challenges, we propose AttnRM, a novel optimization framework that enables the RM to concentrate on crucial segments of the context. Experimental results demonstrate that AttnRM significantly improves preference modeling by increasing attention to relevant information within the context. It also enhances the RM's generalizability and achieves better performance in aligning with human preferences.

Lost in the Context: Insufficient and Distracted Attention to Contexts in Preference Modeling

Cross-lingual transfer from related high-resource languages is a well-established strategy to enhance low-resource language technologies. Prior work has shown that adapters show promise for, e.g., improving low-resource machine translation (MT). In this work, we investigate an adapter souping method combined with cross-attention fine-tuning of a pre-trained MT model to leverage language transfer for three low-resource Creole languages, which exhibit relatedness to different language groups across distinct linguistic dimensions. Our approach improves performance substantially over baselines. However, we find that linguistic relatedness---or even a lack thereof---does not covary meaningfully with adapter performance. Surprisingly, our cross-attention fine-tuning approach appears equally effective with randomly initialized adapters, implying that the benefit of adapters in this setting lies in parameter regularization, and not in meaningful information transfer. We provide analysis supporting this regularization hypothesis. Our findings underscore the reality that neural language processing involves many success factors, and that not all neural methods leverage linguistic knowledge in intuitive ways.

Limited-Resource Adapters Are Regularizers, Not Linguists

Parameter Efficient Fine-Tuning (PEFT) methods have gained popularity and democratized the usage of Large Language Models (LLMs). Recent studies have shown that a small subset of weights significantly impacts performance. Based on this observation, we introduce a novel PEFT method, called Gaussian noise Injected Fine Tuning of Salient Weights (GIFT-SW). Our method updates only salient columns, while injecting Gaussian noise into non-salient ones. To identify these columns, we developed a generalized sensitivity metric that extends and unifies metrics from previous studies. Experiments with LLaMA models demonstrate that GIFT-SW outperforms full fine-tuning and modern PEFT methods under the same computational budget. Moreover, GIFT-SW offers practical advantages to recover performance of models subjected to mixed-precision quantization with keeping salient weights in full precision.

GIFT-SW: Gaussian noise Injected Fine-Tuning of Salient Weights for LLMs

Large language models (LLMs) have made exciting achievements across various domains, yet their deployment on resource-constrained personal devices remains hindered by the prohibitive computational and memory demands of task-specific fine-tuning. While quantization offers a pathway to efficiency, existing methods struggle to balance performance and overhead, either incurring high computational/memory costs or failing to address activation outliers—a critical bottleneck in quantized fine-tuning. To address these challenges, we propose the Outlier Spatial Stability Hypothesis (__OSSH__): _During fine-tuning, certain activation outlier channels retain stable spatial positions across training iterations._ Building on OSSH, we propose __Quaff__, a Quantized parameter-efficient fine-tuning framework for LLMs, optimizing low-precision activation representations through targeted momentum scaling. Quaff dynamically suppresses outliers exclusively in invariant channels using lightweight operations, eliminating full-precision weight storage and global rescaling while reducing quantization errors. Extensive experiments across ten benchmarks validate OSSH and demonstrate Quaff’s efficacy. Specifically, on the GPQA reasoning benchmark, Quaff achieves a $1.73\times$ latency reduction and 30% memory savings over full-precision fine-tuning while improving accuracy by 0.6% on the Phi-3 model, reconciling the triple trade-off between efficiency, performance, and deployability. By enabling consumer-grade GPU fine-tuning (e.g., RTX 2080 Super) without sacrificing model utility, Quaff democratizes personalized LLM deployment. The code is available at https://anonymous.4open.science/r/Quaff-B322/.

Quaff: Quantized Parameter-Efficient Fine-Tuning under Outlier Spatial Stability Hypothesis

Recent advancements in large language models (LLMs) have given rise to the LLM-as-a-judge paradigm, showcasing their potential to deliver human-like judgments. However, in the field of machine translation (MT) evaluation, current LLM-as-a-judge methods fall short of learned automatic metrics. In this paper, we propose Multidimensional Multi-Agent Debate (M-MAD), a systematic LLM-based multi-agent framework for advanced LLM-as-a-judge MT evaluation. Our findings demonstrate that M-MAD achieves significant advancements by (1) decoupling heuristic MQM criteria into distinct evaluation dimensions for fine-grained assessments; (2) employing multi-agent debates to harness the collaborative reasoning capabilities of LLMs; (3) synthesizing dimension-specific results into a final evaluation judgment to ensure robust and reliable outcomes. Comprehensive experiments show that M-MAD not only outperforms all existing LLM-as-a-judge methods but also competes with state-of-the-art reference-based automatic metrics, even when powered by a suboptimal model like GPT-4o mini. Detailed ablations and analysis highlight the superiority of our framework design, offering a fresh perspective for LLM-as-a-judge paradigm. Our code and data are publicly available at https://github.com/SU-JIAYUAN/M-MAD.

Downloads

Next from ACL 2025

Mitigating Visual Forgetting via Take-along Visual Conditioning for Multi-modal Long CoT Reasoning

Similar lecture

Pre-training Distillation for Large Language Models: A Design Space Exploration