China

Large Language Models (LLMs) perform well on reasoning benchmarks but often fail when inputs alter slightly, raising concerns about the extent to which their success relies on memorization. This issue is especially acute in Chain-of-Thought (CoT) reasoning, where spurious memorized patterns can trigger intermediate errors that cascade into incorrect final answers. We introduce STIM, a novel framework for Source-aware Token-level Identification of Memorization, which attributes each token in a reasoning chain to one of multiple memorization sources – local, mid-range, or long-range – based on their statistical co-occurrence with the token in the pretraining corpus. Our token-level analysis across tasks and distributional settings reveals that models rely more on memorization in complex or long-tail cases, and that local memorization is often the dominant driver of errors, leading to up to 67% of wrong tokens. We also show that memorization scores from STIM can be effective in predicting the wrong tokens in the wrong reasoning step. STIM offers a powerful tool for diagnosing and improving model reasoning and can generalize to other structured step-wise generation tasks.

EMNLP 2025

Diagnosing Memorization in Chain-of-Thought Reasoning, One Token at a Time

hardness of samples

knowledge tracing/discovering/inducing

data influence

## 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.

poster

Existing text-to-audio (TTA) generation methods have neither systematically explored audio event relation modeling, nor proposed any new framework to enhance this capability. In this work, we systematically study audio event relation modeling in TTA generation models. We first establish a benchmark for this task by: (1) proposing a comprehensive relation corpus covering all potential relations in real-world scenarios; (2) introducing a new audio event corpus encompassing commonly heard audios; and (3) proposing new evaluation metrics to assess audio event relation modeling from various perspectives. Furthermore, we propose a gated prompt tuning strategy that significantly existing TTA models' performance in relation modeling by introducing negligible parameters.

RiTTA: Modeling Event Relations in Text-to-Audio Generation

CLIP is one of the most popular foundational models and is heavily used for many vision-language tasks. However, little is known about the inner workings of CLIP. While recent work has proposed decomposition-based interpretability methods for identifying textual descriptions of attention heads in CLIP, the implications of conceptual consistency in these text labels on interpretability and model performance has not been explored. To bridge this gap, we study the conceptual consistency of text descriptions for attention heads in CLIP-like models. We conduct extensive experiments on six different models from OpenAI and OpenCLIP which vary by size, type of pre-training data and patch size. We propose Concept Consistency Score (CCS), a novel interpretability metric that measures how consistently individual attention heads in CLIP models align with specific concepts. To assign concept labels to heads, we use in-context learning with ChatGPT, guided by a few manually-curated examples, and validate these labels using an LLM-as-a-judge approach. Our soft-pruning experiments reveal that high CCS heads are critical for preserving model performance, as pruning them leads to a significantly larger performance drop than pruning random or low CCS heads. Notably, we find that high CCS heads capture essential concepts and play a key role in out-of-domain detection, concept-specific reasoning, and video-language understanding. Moreover, we prove that high CCS heads learn spurious correlations amplifying social biases. These results position CCS as a powerful interpretability metric exposing the paradox of performance and social biases in CLIP models.

Pruning the Paradox: How CLIP’s Most Informative Heads Enhance Performance While Amplifying Bias

Guesstimation—the task of making approximate quantitative estimates about objects or events—is a common real-world skill, yet remains underexplored in large language model (LLM) research. We introduce three guesstimation datasets: MARBLES, FUTURE, and ELECPRED, spanning physical estimation (e.g., how many marbles fit in a cup) to abstract predictions (e.g., the 2024 U.S. presidential election). Inspired by the social science concept of Wisdom of Crowds (WOC)—where the median of multiple estimates improves accuracy—we propose WOC decoding for LLMs. We replicate WOC effects in human participants and find that LLMs exhibit similar benefits: median aggregation across sampled responses consistently improves accuracy over greedy and self-consistency decoding. This suggests that LLMs encode a world model that supports approximate reasoning. Our results position guesstimation as a useful probe of LLM world knowledge and highlight WOC decoding as a strategy for enhancing LLM guesstimation performance on real-world tasks.

Probing LLM World Models: Enhancing Guesstimation with Wisdom of Crowds Decoding

Large language model (LLM) agents have demonstrated remarkable potential in advancing scientific discovery. However, their capability in the fundamental yet crucial task of reproducing code from research papers, especially in the NLP domain, remains underexplored. This task includes unique complex reasoning challenges in the intellectual synthesis of abstract concepts and the comprehension of code repositories with interdependent files. Motivated by this gap, we present LMR-Bench, a comprehensive benchmark designed to systematically evaluate the capability of LLM agents on code reproduction from NLP research papers. It consists of 28 code reproduction tasks derived from 23 research papers published in top-tier NLP venues over the past five years, spanning nine fundamental categories. Models are provided with a research paper, a code repository containing one or more masked methods, and instructions for implementing these methods. We conduct extensive experiments in standalone and agent-based settings on state-of-the-art LLMs, evaluating the accuracy of unit tests and performing both LLM and human evaluation of code correctness. Experimental results reveal that even the most advanced models still exhibit persistent limitations in scientific reasoning and code synthesis, highlighting critical gaps in LLMs’ ability to autonomously reproduce scientific research. We will release our benchmark and code after publication.

LMR-BENCH: Evaluating LLM Agent's Ability on Reproducing Language Modeling Research

As LLMs are increasingly applied in socially impactful settings, concerns about gender bias have prompted growing efforts both to measure and mitigate such bias. These efforts often rely on evaluation tasks that differ from natural language distributions, as they typically involve carefully constructed task prompts that overtly or covertly signal the presence of bias-related content. In this paper, we examine how signaling the evaluative purpose of a task impacts measured gender bias in LLMs. Concretely, we test models under prompt conditions that (1) make the testing context salient, and (2) make gender-focused content salient. We then assess prompt sensitivity across four task formats with both token-probability and discrete-choice metrics. We find that even minor prompt changes can substantially alter bias outcomes, sometimes reversing their direction entirely. Discrete-choice metrics further tend to amplify bias relative to probabilistic measures. These findings do not only highlight the brittleness of LLM bias evaluations but open a new puzzle for the NLP benchmarking and development community: To what extent can well-controlled testing designs trigger testing-environment performance, and how do we construct fine-tuning data that minimizes this inference behavior towards more robust bias assessment protocols.

Measuring Bias or Measuring the Task: Understanding the Brittle Nature of LLM Gender Biases

Process-supervised reward models serve as a fine-grained function that provides detailed step-wise feedback to model responses, facilitating effective selection of reasoning trajectories for complex tasks. Despite its advantages, evaluation on PRMs remains less explored, especially in the multimodal domain. To address this gap, this paper first benchmarks current vision large language models (VLLMs) as two types of reward models: output reward models (ORMs) and process reward models (PRMs) on multiple vision-language benchmarks, which reveal that neither ORM nor PRM consistently outperforms across all tasks, and superior VLLMs do not necessarily yield better rewarding performance. To further advance evaluation, we introduce \textsc{ViLBench}, a vision-language benchmark designed to require intensive process reward signals. Notably, OpenAI's GPT-4o with Chain-of-Thought (CoT) achieves only 27.3% accuracy, indicating the benchmark's challenge for current VLLMs. Lastly, we preliminarily showcase a promising pathway towards bridging the gap between general VLLMs and reward models --- by collecting 73.6K vision-language process reward data using an enhanced tree-search algorithm, our 3B model is able to achieve an average improvement of 3.3% over standard CoT and up to 2.5% compared to its untrained counterpart on \textsc{ViLBench} by selecting OpenAI o1’s generations.

ViLBench: A Suite for Vision-Language Process Reward Modeling

Faithful generation in large language models (LLMs) is challenged by knowledge conflicts between parametric memory and external context. Existing contrastive decoding methods tuned specifically to handle conflict often lack adaptability and can degrade performance in low conflict settings. We introduce CoCoA (Confidence- and Context-Aware Adaptive Decoding), a novel token-level algorithm for principled conflict resolution and enhanced faithfulness. CoCoA resolves conflict by utilizing confidence-aware measures (entropy gap and contextual peakedness) and the generalized divergence between the parametric and contextual distributions. Crucially, CoCoA maintains strong performance even in low conflict settings. Extensive experiments across multiple LLMs on diverse Question Answering (QA), Summarization, and Long-Form Question Answering (LFQA) benchmarks demonstrate CoCoA's state-of-the-art performance over strong baselines like AdaCAD. It yields significant gains in QA accuracy, up to 9.2 points on average compared to the strong baseline AdaCAD, and improves factuality in summarization and LFQA by up to 2.5 points on average across key benchmarks. Additionally, it demonstrates superior sensitivity to conflict variations. CoCoA enables more informed, context-aware, and ultimately more faithful token generation.

CoCoA: Confidence- and Context-Aware Adaptive Decoding for Resolving Knowledge Conflicts in Large Language Models

Medical visual question answering (VQA) and federated learning (FL) have emerged as vital approaches for enabling privacy-preserving, collaborative learning across clinical institutions. However, both these approaches face significant challenges in cross-modal FL scenarios, where each client possesses unpaired images from only one modality. To address this limitation, we propose X-FLoRA, a cross-modal FL framework that uses modality-expert low-rank adaptation (LoRA) for medical VQA. Specifically, X-FLoRA enables the synthesis of images from one modality to another without requiring data sharing between clients. This is achieved by training a backward translation model within a federated asymmetric translation scheme that integrates clinical semantics from textual data. Additionally, X-FLoRA introduces modality-expert LoRA, which fine-tunes separate LoRA modules to strengthen modality-specific representations in the VQA task. The server aggregates the trained backward translation models and fine-tuned LoRA modules using discriminator quality scores and expert-aware weighting, which regulate the relative contributions from different clients. Experiments were conducted on VQA datasets encompassing different medical modalities, and the results demonstrate that X-FLoRA outperforms existing FL methods in terms of VQA performance.

X-FLoRA: Cross-modal Federated Learning with Modality-expert LoRA for Medical VQA

Understanding information from visually rich documents remains a significant challenge for traditional Retrieval-Augmented Generation (RAG) methods. Existing benchmarks predominantly focus on image-based question answering (QA), overlooking the fundamental challenges of efficient retrieval, comprehension, and reasoning within dense visual documents. To bridge this gap, we introduce \textbf{ViDoSeek}, a novel dataset designed to evaluate RAG performance on visually rich documents requiring complex reasoning. Based on it, we identify key limitations in current RAG approaches: (i) purely visual retrieval methods struggle to effectively integrate both textual and visual features, and (ii) previous approaches often allocate insufficient reasoning tokens, limiting their effectiveness. To address these challenges, we propose \textbf{ViDoRAG}, a novel multi-agent RAG framework tailored for complex reasoning across visual documents. ViDoRAG employs a Gaussian Mixture Model (GMM)-based hybrid strategy to effectively handle multi-modal retrieval. To further elicit the model's reasoning capabilities, we introduce an iterative agent workflow incorporating exploration, summarization, and reflection, providing a framework for investigating test-time scaling in RAG domains. Extensive experiments on ViDoSeek validate the effectiveness and generalization of our approach. Notably, ViDoRAG outperforms existing methods by over 10% on the competitive ViDoSeek benchmark. The code will be available.

ViDoRAG: Visual Document Retrieval-Augmented Generation via Dynamic Iterative Reasoning Agents

Long-range tasks demand reasoning over long inputs. However, existing solutions are limited, e.g., long-context models require large compute budgets, parameter-efficient fine-tuning (PEFT) needs training data, and retrieval-augmented generation (RAG) entails complex task-specific designs. Though in-context approaches overcome many of these issues, methods with short-context LLMs are inefficient, trading context for processing more tokens. We introduce **PRISM**, a highly token-efficient in-context method based on structured schemas that outperforms baselines on diverse tasks with **4x shorter contexts**. This approach produces concise outputs and efficiently leverages key-value (KV) caches to **reduce costs by up to 54%**. PRISM scales down to tiny contexts without increasing costs or sacrificing quality, and generalizes to new tasks with minimal effort by generating schemas from task descriptions.

Downloads

Next from EMNLP 2025

RiTTA: Modeling Event Relations in Text-to-Audio Generation

Similar lecture

Reason to Rote: Rethinking Memorization in Reasoning