China

We present Branch-Train-Stitch (BTS), an efficient and flexible training algorithm for combining independently trained large language model (LLM) experts into a single, capable generalist model. Following Li et al., we start with a single seed language model which is branched into domain-specific (e.g., coding or math) experts with continual pretraining. BTS combines experts into a generalist model using lightweight stitch layers, which are inserted between frozen experts and the seed LLM, and trained on a small datamix of the expert domains. Stitch layers enable the seed LLM to integrate representations from any number of experts during the forward pass, allowing it to generalize to new domains, despite remaining frozen. Because BTS does not alter the constituent LLMs, BTS provides a modular and flexible approach: experts can be easily removed and new experts can be added with only a small amount of training. Compared to alternative model merging approaches, BTS yields the best generalist performance on a variety of downstream tasks, retaining the specialized capabilities of each of the experts.

EMNLP 2025

BTS: Harmonizing Specialized Experts into a Generalist LLM

modular training

continual pretraining

model merging

llms

Welcome to the 18th Conference of the European Chapter of the Association for Computational Linguistics (EACL). Continuing its mission of expanding and involving the science community of all European countries, EACL has selected the Malta community for the 18th EACL. Considering the importance of physical interaction among researchers, the conference will be held at the Hotel Radisson Blu, St. Julians, in Malta, from 17 to 22 of March, 2024. The conference will also feature the possibility to participate virtually. As the flagship European conference in the field of computational linguistics, EACL welcomes European and international researchers covering a broad spectrum of research areas that are concerned with computational approaches to natural language. 
#### [Conference Handbook](https://drive.google.com/file/d/13Wn38q0yev6U3RTgnf4Tn5T9GuAuZ-xW/view?usp=sharing) *downloadable*

To access the EACL 2024 event page you need to register [**here**](https://acl.swoogo.com/EACL2024). 

Please register!

EACL 2024

As the flagship European conference in the field of computational linguistics, EACL welcomes European and international researchers covering a broad spectrum of research areas that are concerned with computational approaches to natural language.

poster

## 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*

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

Knowledge distillation (KD) is crucial for compressing large text embedding models, but faces challenges when teacher and student models use different tokenizers (Cross-Tokenizer KD - CTKD). Vocabulary mismatches impede the transfer of relational knowledge encoded in deep representations, such as hidden states and attention matrices, which are vital for producing high-quality embeddings. Existing CTKD methods often focus on direct output alignment, neglecting this crucial structural information. We propose a novel framework tailored for CTKD embedding model distillation. We first map tokens one-to-one via Minimum Edit Distance (MinED). Then, we distill intra-model relational knowledge by aligning attention matrix patterns using Centered Kernel Alignment, focusing on the top-m most important tokens of the directly mapped tokens. Simultaneously, we align final hidden states via Optimal Transport with Importance-Scored Mass Assignment, which emphasizes semantically important token representations, based on importance scores derived from attention weights. We evaluate distillation from state-of-the-art embedding models (e.g., LLM2Vec, BGE) to a Bert-base-uncased model on embedding-reliant tasks such as text classification, sentence pair classification, and semantic textual similarity. Our proposed framework significantly outperforms existing CTKD baselines. By preserving attention structure and prioritizing key representations, our approach yields smaller, high-fidelity embedding models despite tokenizer differences.

EMO: Embedding Model Distillation via Intra-Model Relation and Optimal Transport Alignments

While reinforcement learning (RL) has demonstrated remarkable success in enhancing large language models (LLMs), it has primarily focused on single-turn tasks such as solving math problems. Training effective web agents for multi-turn interactions remains challenging due to the complexity of long-horizon decision-making across dynamic web interfaces. In this work, we present WebAgent-R1, a simple yet effective end-to-end multi-turn RL framework for training web agents. It learns directly from online interactions with web environments by asynchronously generating diverse trajectories, entirely guided by binary rewards depending on task success. Experiments on the WebArena-Lite benchmark demonstrate the effectiveness of WebAgent-R1, boosting the task success rate of Qwen-2.5-3B from 6.1% to 33.9% and LLaMA-3.1-8B from 8.5% to 44.8%, significantly outperforming existing state-of-the-art methods and strong proprietary models such as OpenAI o3. In-depth analyses reveal the effectiveness of the thinking-based prompting strategy and test-time scaling through increased interactions for web tasks. We further investigate different RL initialization policies by introducing two variants, namely WebAgent-R1-Zero and WebAgent-R1-CoT, which highlight the importance of the warm-up training stage (i.e., behavior cloning) and provide insights on incorporating long chain-of-thought (CoT) reasoning in web agents.

WebAgent-R1: Training Web Agents via End-to-End Multi-Turn Reinforcement Learning

With the rapid development of large language models (LLMs), natural language explanations (NLEs) have become increasingly important for understanding model predictions. However, these explanations often fail to faithfully represent the model's internal reasoning process. While existing work has demonstrated that LLMs can self-critique and refine their initial outputs for various tasks, this capability remains unexplored for improving explanation faithfulness. To address this gap, we introduce Self-critique and Refinement for Natural Language Explanations (SR-NLE), a framework that enables models to improve the faithfulness of their own explanations -- specifically, post-hoc NLEs -- through an iterative critique and refinement process without external supervision. Our framework leverages different feedback mechanisms to guide the refinement process, including natural language self-feedback and, notably, a novel feedback approach based on feature attribution that highlights important input words. Our experiments across three datasets and four state-of-the-art LLMs demonstrate that SR-NLE significantly reduces unfaithfulness rates, with our best method achieving an average unfaithfulness rate of 36.02%, compared to 54.81% for baseline -- an absolute reduction of 18.79%. These findings reveal that the investigated LLMs can indeed refine their explanations to better reflect their internal reasoning process, requiring only appropriate guidance through feedback without additional training or fine-tuning.

Self-Critique and Refinement for Faithful Natural Language Explanations

When AI agents retrieve and reason over external documents, adversaries can manipulate the data they receive to subvert their behaviour. Previous research has studied indirect prompt injection, where the attacker injects malicious instructions. We argue that malicious instructions are not necessary to manipulate agents - attackers can accomplish the same by supplying biased, misleading, or false information. We term this attacks-by-persuasion. Existing defenses, which focus on detecting hidden commands, are ineffective against attacks-by-persuasion. To defend themselves and their users, agents must critically evaluate retrieved information, corroborating claims with external evidence and evaluating source trustworthiness. We argue that this is analogous to an existing NLP task, automated fact-checking, which we propose to repurpose as a cognitive self-defense tool for agents.

Attacks by Content: Automated Fact-checking is an AI Security Issue

Integrating large language models (LLMs) as action proposers in reinforcement learning (RL) significantly boosts performance in text-based environments but incurs prohibitive computational costs. We introduce a cache-efficient framework for Bayesian RL that leverages LLM-derived action suggestions, drastically reducing these costs while maintaining near-optimal performance. Our approach features an adaptive caching mechanism, optimized via meta-learning based on policy performance, to enable efficient inference across text-based games (e.g., TextWorld, ALFWorld) and robotic control tasks (e.g., MuJoCo, MetaWorld). This framework achieves a 3.8times–4.7times reduction in LLM queries and 4.0times–12.0times lower median latencies (85–93ms on consumer hardware), while retaining 96–98% of the uncached policy's performance. We provide theoretical guarantees on the reliability of cached decisions with Kullback-Leibler (KL) divergence bounds, which are validated empirically by high success rates (90.4–95.6%) in complex text environments. For offline RL, our proposed CQL-Prior variant improves performance by 14–29% and reduces training time by 38–40%. Evaluations across eight diverse tasks demonstrate the framework's generalizability and practicality for resource-constrained settings, making LLM-guided RL a viable and accessible approach for both text-based and robotic applications.

Cache-Efficient Posterior Sampling for Reinforcement Learning with LLM-Derived Priors Across Discrete and Continuous Domains

Characterizing the expressive power of the Transformer architecture is critical to understanding its capacity limits and scaling law. Recent works provide the circuit complexity bounds to Transformer-like architecture. On the other hand, Rotary Position Embedding (mathsfRoPE) has emerged as a crucial technique in modern large language models, offering superior performance in capturing positional information compared to traditional position embeddings, which shows great potential in application prospects, particularly for the long context scenario. Empirical evidence also suggests that mathsfRoPE-based Transformer architectures demonstrate greater generalization capabilities compared to conventional Transformer models. In this work, we establish a circuit complexity bound for Transformers with mathsfRoPE attention. Our key contribution is that we show that unless mathsfTC⁰ = mathsfNC¹, a mathsfRoPE-based Transformer with poly(n)-precision, O(1) layers, hidden dimension d leq O(n) cannot solve the arithmetic formula evaluation problem or the Boolean formula value problem. This result significantly demonstrates the fundamental limitation of the expressivity of mathsfRoPE-based Transformers, although they achieve giant empirical success. Our theoretical result not only establishes the complexity bound but also gives insights into designing novel Transformer architectures.

Circuit Complexity Bounds for RoPE-based Transformer Architecture

Prompting and context-based fine-tuning methods, which we call Prefix Learning, have been proposed to enhance the performance of language models on various downstream tasks. They are empirically efficient and effective, matching the performance of full parameter fine-tuning, but the theoretical understandings are limited. In this paper, we aim to address this limitation by studying their ability from the perspective of prefix length. In particular, we provide a convergence guarantee for training an ultra-long prefix in a stylized setting using the Neural Tangent Kernel (NTK) framework. Based on this strong theoretical guarantee, we design and implement an algorithm that only needs to introduce and fine-tune a few extra trainable parameters instead of an infinite-long prefix in each layer of a transformer, and can approximate the prefix attention to a guaranteed polynomial-small error. Preliminary experimental results on vision, natural language, and math data show that our method achieves superior or competitive performance compared to existing methods like full parameters fine-tuning, P-Tuning V2, and LoRA. This demonstrates our method is promising for parameter-efficient fine-tuning.

Towards Infinite-Long Prefix in Transformer

This paper presents AlphaOne (α1), a universal framework for modulating reasoning progress in large reasoning models (LRMs) at test time. α1 first introduces α moment, which represents the scaled thinking phase with a universal parameter α. Within this scaled pre-α moment phase, it dynamically schedules slow thinking transitions by modeling the insertion of reasoning transition tokens as a Bernoulli stochastic process. After the α moment, α1 deterministically terminates slow thinking with the end-of-thinking token, thereby fostering fast reasoning and efficient answer generation. This approach unifies and generalizes existing monotonic scaling methods by enabling flexible and dense slow-to-fast reasoning modulation. Extensive empirical studies on various challenging benchmarks across mathematical, coding, and scientific domains demonstrate α1's superior reasoning capability and efficiency. Anonymous project page: https://alphaone-paper.github.io/.

AlphaOne: Reasoning Models Thinking Slow and Fast at Test Time

Video large language models (Vid-LLMs), which excel in diverse video-language tasks, can be effectively constructed by adapting image-pretrained vision-language models (VLMs). However, this adaptation remains challenging, as it requires processing dense and temporally extended visual inputs that exceed the capacity of image-based models. This paper identifies the perception bottleneck and token overload as key challenges in extending image-based VLMs to the video domain. To address these issues, we propose D-CoDe, a training-free adaptation framework that incorporates dynamic compression and question decomposition. Specifically, dynamic compression alleviates the perception bottleneck through adaptive selection of representative frames and content-aware aggregation of spatial tokens, thereby reducing redundancy while preserving informative content. In parallel, question decomposition mitigates token overload by reformulating the original query into sub-questions, guiding the model to focus on distinct aspects of the video and enabling more comprehensive understanding. Experiments demonstrate that D-CoDe effectively improves video understanding across various benchmarks. Furthermore, strong performance on the challenging long-video benchmark highlights the potential of D-CoDe in handling complex video-language tasks.

D-CoDe: Scaling Image-Pretrained VLMs to Video via Dynamic Compression and Question Decomposition

While significant progress has been made with dual- and bi-encoder dense retrievers, they often struggle on queries with logical connectives, a use case that is often overlooked yet important in downstream applications. Current dense retrievers struggle with such queries, such that the retrieved results do not respect the logical constraints implied in the queries. To address this challenge, we introduce LogiCoL, a logically-informed contrastive learning objective for dense retrievers. LogiCoL builds upon in-batch supervised contrastive learning, and learns dense retrievers to respect the subset and mutually-exclusive set relation between query results via two sets of soft constraints expressed via t-norm in the learning objective. We evaluate the effectiveness of LogiCoL on the task of entity retrieval, where the model is expected to retrieve a set of entities in Wikipedia that satisfy the implicit logical constraints in the query. We show that models trained with LogiCoL yield improvement both in terms of retrieval performance and logical consistency in the results. We provide detailed analysis and insights to uncover why queries with logical connectives are challenging for dense retrievers and why LogiCoL is most effective.

Downloads

Next from EMNLP 2025

EMO: Embedding Model Distillation via Intra-Model Relation and Optimal Transport Alignments

Similar lecture

Hyper-BTS Dataset: Scalability and Enhanced Analysis of Back TranScription (BTS) for ASR Post-Processing

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES