Austria

Large Language Models (LLMs) have demonstrated exceptional capabilities across diverse natural language processing (NLP) tasks. The release of open-source LLMs like LLaMA and Qwen has triggered the development of numerous fine-tuned models tailored for various tasks and languages. In this paper, we explore an important question: is it possible to combine these specialized models to create a unified model with multi-task capabilities. We introduces **H**ierarchical **I**terative **Merging** (Hi-Merging), a training-free method for unifying different specialized LLMs into a single model. Specifically, Hi-Merging employs model-wise and layer-wise pruning and scaling, guided by contribution analysis, to mitigate parameter conflicts. Extensive experiments on multiple-choice and question-answering tasks in both Chinese and English validate Hi-Merging&#39;s ability for multi-task learning. The results demonstrate that Hi-Merging consistently outperforms existing merging techniques and surpasses the performance of models fine-tuned on combined datasets in most scenarios. Code is available at [Applied-Machine-Learning-Lab/Hi-Merging](https://github.com/Applied-Machine-Learning-Lab/Hi-Merging).

ACL 2025

Training-free LLM Merging for Multi-task Learning

training-free

model merging

large language models

multi-task 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)

Large Language Models (LLMs) have demonstrated exceptional capabilities across diverse natural language processing (NLP) tasks. The release of open-source LLMs like LLaMA and Qwen has triggered the development of numerous fine-tuned models tailored for various tasks and languages. In this paper, we explore an important question: is it possible to combine these specialized models to create a unified model with multi-task capabilities. We introduces **H**ierarchical **I**terative **Merging** (Hi-Merging), a training-free method for unifying different specialized LLMs into a single model. Specifically, Hi-Merging employs model-wise and layer-wise pruning and scaling, guided by contribution analysis, to mitigate parameter conflicts. Extensive experiments on multiple-choice and question-answering tasks in both Chinese and English validate Hi-Merging's ability for multi-task learning. The results demonstrate that Hi-Merging consistently outperforms existing merging techniques and surpasses the performance of models fine-tuned on combined datasets in most scenarios. Code is available at [Applied-Machine-Learning-Lab/Hi-Merging](https://github.com/Applied-Machine-Learning-Lab/Hi-Merging).

technical paper

Large language models (LLMs) have demonstrated exceptional performance across various applications, but their conversational abilities decline sharply as model size decreases, presenting a barrier to their deployment in resource-constrained environments. Knowledge distillation (KD) with Direct Preference Optimization (DPO) has emerged as a promising approach to enhance the conversational abilities of smaller models using a larger teacher model. However, current methods primarily focus on ``black-box'' KD, which only uses the teacher’s responses, overlooking the rich distributional information within the teacher’s probability distribution. This paper addresses this gap by introducing daDPO (Distillation-Aware DPO), a novel framework that integrates the teacher’s distributional information into DPO distillation while preserving theoretical guarantees. Our framework offers a unified objective that enhances both preference optimization and distribution-based distillation. We provide rigorous theoretical analysis and empirical validation, showing that daDPO outperforms existing methods in restoring performance for pruned models and enhancing smaller models within the same LLM family. Notably, in in-domain evaluation, our method enables a 20% pruned Vicuna1.5-7B to achieve near-teacher performance (-7.3% preference rate), and allows Qwen2.5-1.5B to occasionally outperform its 7b teacher model (14.0% win rate).

daDPO: Distribution-Aware DPO for Distilling Conversational Abilities

Understanding human preferences is crucial for improving foundation models and building personalized AI systems. However, preferences are inherently diverse and complex, making it difficult for traditional reward models to capture their full range. While fine-grained preference data can help, collecting it is expensive and hard to scale. In this paper, we introduce Decomposed Reward Models (DRMs), a novel approach that extracts diverse human preferences from binary comparisons without requiring fine-grained annotations. Our key insight is to represent human preferences as vectors and analyze them using Principal Component Analysis (PCA). By constructing a dataset of embedding differences between preferred and rejected responses, DRMs identify orthogonal basis vectors that capture distinct aspects of preference. These decomposed rewards can be flexibly combined to align with different user needs, offering an interpretable and scalable alternative to traditional reward models. We demonstrate that DRMs effectively extract meaningful preference dimensions (e.g., helpfulness, safety, humor) and adapt to new users without additional training. Our results highlight DRMs as a powerful framework for personalized and interpretable LLM alignment.

Rethinking Diverse Human Preference Learning through Principal Component Analysis

As Large Language Models (LLMs) scale up and gain powerful Chain-of-Thoughts (CoTs) reasoning abilities, practical resource constraints drive efforts to distill these capabilities into more compact Smaller Language Models (SLMs). We find that CoTs consist mainly of simple reasoning forms, with a small proportion (\~4.7%) of key reasoning steps that truly impact conclusions. However, previous distillation methods typically involve supervised fine-tuning student SLMs only on correct CoTs data produced by teacher LLMs, resulting in students struggling to learn the key, instead imitating the teacher's reasoning forms and making errors or omissions in reasoning. To address these issues, drawing an analogy to human learning, where analyzing mistakes according to correct solutions often reveals the crucial steps leading to successes or failures, we propose mistak\textbf{E}-\textbf{D}riven key reason\textbf{I}ng step distilla\textbf{T}ion (\textbf{EDIT}), a novel method that further aids SLMs learning key reasoning steps rather than mere simple fine-tuning. Firstly, to expose the crucial steps in CoTs, we carefully design specific prompts to generate dual CoTs data with similar reasoning paths but divergent conclusions. Then, we apply the minimum edit distance algorithm on the dual CoTs data to locate these key steps and optimize the likelihood on these tokens. Extensive experiments and analysis validate the effectiveness of EDIT across both in-domain(IND) and out-of-domain(OOD) benchmark reasoning datasets\footnote{Code can be found at \url{https://github.com/C-W-D/EDIT}}.

Capture the Key in Reasoning to Enhance CoT Distillation Generalization

Prompt optimization for language models faces challenges due to the large discrete search space, the reliance on continuous gradient updates, and the need to round continuous representations into discrete prompts, which causes inflexibility and instability. Existing methods attempt to address these by constraining the search space and adopting greedy, incremental improvements, but they often fail to fully leverage historical gradient information. In this paper, we model the prompt optimization problem by the probability distribution of the prompt and present a novel approach that integrates greedy strategies into optimization with continuous representations. This approach can exploit historical gradient information to address the instability caused by rounding in existing methods. Our study indicates that using continuous representations can improve prompt optimization performance on both text classification and attack tasks, as well as models, including GPT-2, OPT, Vicuna, and LLaMA-2, and also be adaptable to models of different sizes.

Direct Prompt Optimization with Continuous Representations

Deep Reinforcement Learning (DRL) is widely used in task-oriented dialogue systems to optimize dialogue policy, but it struggles to balance exploration and exploitation due to the high dimensionality of state and action spaces. This challenge often results in local optima or poor convergence. Evolutionary Algorithms (EAs) have been proven to effectively explore the solution space of neural networks by maintaining population diversity. Inspired by this, we innovatively combine the global search capabilities of EA with the local optimization of DRL to achieve a balance between exploration and exploitation. Nevertheless, the inherent flexibility of natural language in dialogue tasks complicates this direct integration, leading to prolonged evolutionary times. Thus, we further propose an elite individual injection mechanism to enhance EA's search efficiency by adaptively introducing best-performing individuals into the population. Experiments across four datasets show that our approach significantly improves the balance between exploration and exploitation, boosting performance. Moreover, the effectiveness of the EII mechanism in reducing exploration time has been demonstrated, achieving an efficient integration of EA and DRL on task-oriented dialogue policy tasks.

An Efficient Task-Oriented Dialogue Policy: Evolutionary Reinforcement Learning Injected by Elite Individuals.

Despite existing multimodal language models showing impressive performance on the video understanding task, extremely long videos still pose significant challenges to language model's context length, memory consumption, and computational complexity. To address these issues, we propose a vision-language model named Sophia for long video understanding, which can efficiently handle hour-scale long videos. First, we employ a Shot-adaptive Frame Pruning technique, which naturally segments long videos into multiple camera shots, to more sharply identify and focus on the frames relevant to the query. Additionally, we introduce a Hierarchical Attention mechanism to effectively model the long-term temporal dependencies between video frames, which achieves a time and space complexity of O(N) w.r.t. the input sequence length N while theoretically maintaining the global modeling efficiency. Experimentally, our Sophia exhibits competitive performance compared to existing video understanding baselines across various benchmarks for long video understanding with reduced time and memory consumption. The model code and weights are available at https://huggingface.co/Tao-tse/Sophia.

Sharper and Faster mean Better: Towards More Efficient Vision-Language Model for Hour-scale Long Video Understanding

Large language models (LLMs) combined with instruction tuning have shown significant progress in information extraction (IE) tasks, exhibiting strong generalization capabilities to unseen datasets by following annotation guidelines. However, their applicability to low-resource languages remains limited due to lack of both labeled data for fine-tuning, and unlabeled text for pre-training. In this paper, we propose TransFusion, a framework in which models are fine-tuned to use English translations of low-resource language data, enabling more precise predictions through annotation fusion. Based on TransFusion, we introduce GoLLIE-TF, a cross-lingual instruction-tuned LLM for IE tasks, designed to close the performance gap between high and low-resource languages. Our experiments across twelve multilingual IE datasets spanning 50 languages demonstrate that GoLLIE-TF achieves better cross-lingual transfer over the base model. In addition, we show that TransFusion significantly improves low-resource language named entity recognition when applied to proprietary models such as GPT-4 (+5 F1) with a prompting approach, or fine-tuning different language models including decoder-only (+14 F1) and encoder-only (+13 F1) architectures.

Translation and Fusion Improves Cross-lingual Information Extraction

The rapid advancements in Large Language Models (LLMs) have enabled their deployment as autonomous agents for handling complex tasks in dynamic environments. These LLMs demonstrate strong problem-solving capabilities and adaptability to multifaceted scenarios. However, their use as agents also introduces significant risks, including task-specific risks, which are identified by the agent administrator based on the specific task requirements and constraints, and systemic risks, which stem from vulnerabilities in their design or interactions, potentially compromising confidentiality, integrity, or availability (CIA) of information and triggering security risks. Existing defense agencies fail to adaptively and effectively mitigate these risks. In this paper, we propose AGrail, a lifelong agent guardrail to enhance LLM agent safety, which features adaptive safety check generation, effective safety check optimization, and tool compatibility & flexibility. Extensive experiments demonstrate that AGrail not only achieves strong performance against task-specific and system risks but also exhibits transferability across different LLM agents' tasks.

AGrail: A Lifelong Agent Guardrail with Effective and Adaptive Safety Detection

Content analysis breaks down complex and unstructured texts into theory-informed numerical categories. Particularly, in social science, this process usually relies on multiple rounds of manual annotation, domain expert discussion, and rule-based refinement. In this paper, we introduce SCALE, a novel multi-agent framework that effectively $\underline{\textbf{S}}$imulates $\underline{\textbf{C}}$ontent $\underline{\textbf{A}}$nalysis via $\underline{\textbf{L}}$arge language model (LLM) ag$\underline{\textbf{E}}$nts. SCALE imitates key phases of content analysis, including text coding, collaborative discussion, and dynamic codebook evolution, capturing the reflective depth and adaptive discussions of human researchers. Furthermore, by integrating diverse modes of human intervention, SCALE is augmented with expert input to further enhance its performance. Extensive evaluations on real-world datasets demonstrate that SCALE achieves human-approximated performance across various complex content analysis tasks, offering an innovative potential for future social science research.

SCALE: Towards Collaborative Content Analysis in Social Science with Large Language Model Agents and Human Intervention

Automating benefit verification phone calls saves time in healthcare and helps patients receive treatment faster. It is critical to obtain highly accurate information in these phone calls, as it can affect a patient's healthcare journey. Given the noise in phone call transcripts, we have a two-stage system that involves a post-call review phase for potentially noisy fields, where human reviewers manually verify the extracted data—a labor-intensive task. To automate this stage, we introduce Auto Review, which significantly reduces manual effort while maintaining a high bar for accuracy. This system, being highly reliant on call transcripts, suffers a performance bottleneck due to automatic speech recognition (ASR) issues. This problem is further exacerbated by the use of domain-specific jargon in the calls. In this work, we propose a second-stage postprocessing pipeline for accurate information extraction. We improve accuracy by using multiple ASR alternatives and a pseudo-labeling approach that does not require manually corrected transcripts. Experiments with general-purpose large language models and feature-based model pipelines demonstrate substantial improvements in the quality of corrected call transcripts, thereby enhancing the efficiency of Auto Review.

Downloads

Next from ACL 2025

daDPO: Distribution-Aware DPO for Distilling Conversational Abilities

Similar lecture

Unraveling LoRA Interference: Orthogonal Subspaces for Robust Model Merging