China

Reasoning models have demonstrated remarkable performance on complex tasks by generating long reasoning traces prior to producing final answers. However, previous research on long-context scaling in language models has generally focused on managing lengthy input prompts instead of producing long outputs. To leverage the strong long context understanding abilities of current models, we introduce Understanding-to-Reasoning Transition (URT) fine-tuning, a sequence-level curriculum learning framework that gradually shifts a model’s focus from interpreting long chain-of-thoughts to generating them. By incorporating partial reasoning steps in the input context, URT naturally exposes the model to diverse prompt lengths during training, preserving its performance on long-context comprehension while developing advanced reasoning capabilities. Experiments on rigorous reasoning benchmarks, including AIME24 and GPQA Diamond, reveal that our approach surpasses standard fine-tuning by over 10%, while maintaining robust performance on the understanding tasks in RULER.

EMNLP 2025

Long Chain-of-Thought Fine-tuning via Understanding-to-Reasoning Transition

reasoning models

long context

llms

## 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."<br>

*Message from the General Chair, Dirk Hovy*





<br>

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

technical paper

Standard Byte-Pair Encoding (BPE) tokeniza- tion compresses text by pairing a learned token vocabulary with a detailed merge list. Recent work has shown that this merge list exposes a potential attack surface for extracting informa- tion about language model’s training data. In this paper, we explore the downstream impact of BPE inference algorithms that do not rely on this merge list at all, and hence differ from the encoding process during the BPE training. To address this question, we investigate two broad classes of BPE inference schemes that differ from BPE appliction during training: a) targetted deviation from merge-lists including random merge orders, and various corruptions of merge list involving deletion/truncation, and b) non-targetted BPE inference algorithms that do not depend on the merge list but focus on compressing the text either greedily or exactly. Extensive experiments across diverse language modeling tasks like accuracy-based QA bench- marks, machine translation, and open-ended generation reveal that while the targetted devi- ation from the merge lists exhibit significant degradation in language model performance, the non-targetted merge-list free inference algo- rithms result in minimal impact on downstream performance that is often much smaller than expected. These findings pave way for simpler and potentially more privacy-preserving tok- enization schemes that do not catastrophically compromise model performance.

Train It and Forget It: Merge Lists are Unnecessary for BPE Inference in Language Models

Growing literature explores toxicity and polarization in discourse, with comparatively less work on characterizing what makes dialogue prosocial and constructive. We explore conversational discourse and investigate a method for characterizing its quality built upon the notion of ``responsivity''---whether one person's conversational turn is responding to a preceding turn. We develop and evaluate methods for quantifying responsivity---first through semantic similarity of speaker turns, and second by leveraging state-of-the-art large language models (LLMs) to identify the relation between two speaker turns. We evaluate both methods against a ground truth set of human-annotated conversations. Furthermore, selecting the better performing LLM-based approach, we characterize the nature of the response---whether it responded to that preceding turn in a substantive way or not. We view these responsivity links as a fundamental aspect of dialogue but note that conversations can exhibit significantly different responsivity structures. Accordingly, we then develop conversation-level derived metrics to address various aspects of conversational discourse. We use these derived metrics to explore other conversations and show that they support meaningful characterizations and differentiations across a diverse collection of conversations.

Computational Analysis of Conversation Dynamics through Participant Responsivity

Word-level quality estimation (QE) methods aim to detect erroneous spans in machine translations, which can direct and facilitate human post-editing.
While the accuracy of word-level QE systems has been assessed extensively, their usability and downstream influence on the speed, quality and editing choices of human post-editing remain understudied. In this study, we investigate the impact of word-level QE on machine translation (MT) post-editing in a realistic setting involving 42 professional post-editors across two translation directions. We compare four error-span highlight modalities, including supervised and uncertainty-based word-level QE methods, for identifying potential errors in the outputs of a state-of-the-art neural MT model. Post-editing effort and productivity are estimated from behavioral logs, while quality improvements are assessed by word- and segment-level human annotation. We find that domain, language and editors' speed are critical factors in determining highlights' effectiveness, with modest differences between human-made and automated QE highlights underlining a gap between accuracy and usability in professional workflows.

QE4PE: Word-level Quality Estimation for Human Post-Editing

We introduce MultiBLiMP 1.0, a massively multilingual benchmark of linguistic minimal pairs, covering 101 languages and 2 types of subject-verb agreement, containing more than 128,000 minimal pairs. Our minimal pairs are created using a fully automated pipeline, leveraging the large-scale linguistic resources of Universal Dependencies and UniMorph. MultiBLiMP 1.0 evaluates abilities of LLMs at an unprecedented multilingual scale, and highlights the shortcomings of the current state-of-the-art in modelling low-resource languages.

MultiBLiMP 1.0: A Massively Multilingual Benchmark of Linguistic Minimal Pairs

Text-to-SQL is an important task that helps access databases by generating SQL queries. Currently, correcting the generated SQL based on large language models (LLMs) automatically is an effective method to enhance the quality of the generated SQL. However, previous research shows that it is hard for LLMs to detect mistakes in SQL directly, leading to poor performance. Therefore, in this paper, we propose to employ the decomposed correction to enhance text-to-SQL performance. We first demonstrate that detecting and fixing mistakes based on the decomposed sub-tasks is easier than using SQL directly. Then, we introduce Decomposed Automation Correction (DAC), which first generates the entities and skeleton corresponding to the question, and then compares the differences between the initial SQL and the generated entities and skeleton as feedback for correction. Experimental results show that, compared with the previous automation correction method, DAC improves performance by 1.4% of Spider, Bird, and KaggleDBQA on average, demonstrating the effectiveness of DAC.

DAC: Decomposed Automation Correction for Text-to-SQL

Multimodal Retrieval-Augmented Generation (MM-RAG) has emerged as a promising approach for enhancing the reliability and factuality of large vision-language models (LVLMs). While end-to-end loss backpropagation is infeasible due to non-differentiable operations during the forward process, current methods primarily focus on component-level optimizations, necessitate extensive component-specific training datasets and suffer from a gap between local and global optimization objectives. In this paper, we propose a new paradigm that backpropagates global rewards from the system output to each component and then transforms these rewards into specific local losses, enabling each component to perform gradient descent and thus ensuring end-to-end optimization. Specifically, we first insert two lightweight multimodal components, a query translator and an adaptive reranker, to address the heterogeneity of multimodal knowledge and the varying knowledge demands for different questions, and then tune only these inserted components using our proposed paradigm to integrate the entire system. Our method achieves SOTA performance on multiple knowledge-intensive multimodal benchmarks with high training efficiency, relying exclusively on supervised signals from an external reward model. Experimental results and our detailed analysis of the evolution of components during training collectively reveal the advantages and considerable potential of this paradigm as a promising direction for MM-RAG research.

End-to-End Optimization for Multimodal Retrieval-Augmented Generation via Reward Backpropagation

In recent years, driven by advancements in the diffusion process, Text-to-Image (T2I) models have rapidly developed. However, evaluating T2I models remains a significant challenge. While previous research has thoroughly assessed the quality of generated images and image-text alignment, there has been little study on the creativity of these models. In this work, we defined the creativity of T2I models, inspired by previous definitions of machine creativity. We also proposed corresponding metrics and designed a method to test the reliability of the metric. Additionally, we developed a fully automated pipeline capable of transforming existing image-text datasets into benchmarks tailored for evaluating creativity, specifically through text vector retrieval and the text generation capabilities of large language models (LLMs). Finally, we conducted a series of tests and analyses on the evaluation methods for T2I model creativity and the factors influencing the creativity of the models, revealing that current T2I models demonstrate a lack of creativity. The code and benchmark will be released.

Evaluation of Text-to-Image Generation from a Creativity Perspective

Conversational explainable artificial intelligence (ConvXAI) systems based on large language models (LLMs) have garnered considerable attention for their ability to enhance user comprehension through dialogue-based explanations. Current ConvXAI systems often are based on intent recognition to accurately identify the user's desired intention and map it to an explainability method. While such methods offer great precision and reliability in discerning users' underlying intentions for English, a significant challenge in the scarcity of training data persists, which impedes multilingual generalization. Besides, the support for free-form custom inputs, which are user-defined data distinct from pre-configured dataset instances, remains largely limited. To bridge these gaps, we first introduce MultiCoXQL, a multilingual extension of the CoXQL dataset spanning five typologically diverse languages, including one low-resource language. Subsequently, we propose a new parsing approach aimed at enhancing multilingual parsing performance, and evaluate three LLMs on MultiCoXQL using various parsing strategies. Furthermore, we present Compass, a new multilingual dataset designed for custom input extraction in ConvXAI systems, encompassing 11 intents across the same five languages as MultiCoXQL. We conduct monolingual, cross-lingual, and multilingual evaluations on Compass, employing three LLMs of varying sizes alongside BERT-type models.

Multilingual Datasets for Custom Input Extraction and Explanation Requests Parsing in Conversational XAI Systems

Fine-tuning large language models (LLMs) is computationally expensive, and Low-Rank Adaptation (LoRA) provides a cost-effective solution by approximating weight updates through low-rank matrices. In real-world scenarios, LLMs are fine-tuned on data from multiple domains to perform tasks across various fields, embodying multi-task learning (MTL). LoRA often underperforms in such complex scenarios. To enhance LoRA's capability in multi-task learning, we propose R-LoRA, which incorporates Multi-Head Randomization. Multi-Head Randomization diversifies the head matrices through Multi-Head Dropout and Multi-Head Random Initialization, enabling more efficient learning of task-specific features while maintaining shared knowledge representation. Our approach not only improves performance in MTL but also reduces GPU memory usage and training time. Experiments show that R-LoRA's gains stem from increased diversity in the head matrices, demonstrating its effectiveness for multi-task learning. The code will be open-sourced after the revision.

R-LoRA: Randomized Multi-Head LoRA for Efficient Multi-task Learning

Language confusion---where large language models (LLMs) generate unintended languages against the user's need---remains a critical challenge, especially for English-centric models. We present the first mechanistic interpretability (MI) study of language confusion, combining behavioral benchmarking with neuron-level analysis. Using the Language Confusion Benchmark (LCB), we show that confusion points (CPs)---specific positions where language switches occur---are central to this phenomenon. Through layer-wise analysis with TunedLens and targeted neuron attribution, we reveal that transition failures in the final layers drive confusion. We further demonstrate that editing a small set of critical neurons, identified via comparative analysis with multilingual-tuned models, substantially mitigates confusion without harming general competence or fluency. Our approach matches multilingual alignment in confusion reduction for most languages and yields cleaner, higher-quality outputs. These findings provide new insights into the internal dynamics of LLMs and highlight neuron-level interventions as a promising direction for robust, interpretable multilingual language modeling.

Downloads

Next from EMNLP 2025

Train It and Forget It: Merge Lists are Unnecessary for BPE Inference in Language Models

Similar lecture

Facilitating Long Context Understanding via Supervised Chain-of-Thought Reasoning