China

Recent advances in Large Language Model (LLM)-based Role-Playing Language Agents (RPLAs) have attracted broad attention in various applications. While chain-of-thought reasoning has shown importance in many tasks for LLMs, the internal thinking processes of RPLAs remain unexplored. Understanding characters’ inner thoughts is crucial for developing advanced RPLAs. In this paper, we introduce ROLETHINK, a novel benchmark constructed from literature for evaluating character thought generation. We propose the task of inner thought reasoning, constructing 6,058 data entries from 76 books, which includes two sets: the gold set that compares generated thoughts with original character monologues, and the silver set that uses expert-synthesized character analyses as references. To address this challenge, we propose MIRROR, a chain-of-thought approach that generates character thoughts by retrieving memories, predicting character reactions, and synthesizing motivations. Through extensive experiments, we demonstrate the importance of inner thought reasoning for RPLAs, and MIRROR consistently outperforms existing methods.

EMNLP 2025

Guess What I am Thinking: A Benchmark for Inner Thought Reasoning of Role-Playing Language Agents

inner thought

role-playing language agents

mirror

## 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

Long-form egocentric video understanding provides rich contextual information and unique insights into long-term human behaviors, holding significant potential for applications in embodied intelligence, long-term activity analysis, and personalized assistive technologies. However, existing benchmark datasets primarily focus on single, short (\eg, minutes to tens of minutes) to moderately long videos, leaving a substantial gap in evaluating extensive, ultra-long egocentric video recordings. To address this, we introduce X-LeBench, a novel benchmark dataset meticulously designed to fill this gap by focusing on tasks requiring a comprehensive understanding of extremely long egocentric video recordings. Our X-LeBench develops a life-logging simulation pipeline that produces realistic, coherent daily plans aligned with real-world video data. This approach enables the flexible integration of synthetic daily plans with real-world footage from Ego4D—a massive-scale egocentric video dataset covers a wide range of daily life scenarios—resulting in 432 simulated video life logs spanning from 23 minutes to 16.4 hours. The evaluations of several baseline systems and multimodal large language models (MLLMs) reveal their poor performance across the board, highlighting the inherent challenges of long-form egocentric video understanding, such as temporal localization and reasoning, context aggregation, and memory retention, and underscoring the need for more advanced models. Our dataset is available at X-LeBench.

X-LeBench: A Benchmark for Extremely Long Egocentric Video Understanding

Many vision-language models (VLMs) that prove very effective at a range of multimodal task, build on CLIP-based vision encoders, which are known to have various limitations. We investigate the hypothesis that the strong language backbone in VLMs compensates for possibly weak visual features by contextualizing or enriching them. Using three CLIP-based VLMs, we perform controlled self-attention ablations on a carefully designed probing task. Our findings show that despite known limitations, CLIP visual representations offer ready-to-read semantic information to the language decoder. However, in scenarios of reduced contextualization in the visual representations, the language decoder can largely compensate for the deficiency and recover performance. This suggests a dynamic division of labor in VLMs and motivates future architectures that offload more visual processing to the language decoder.

LLMs Can Compensate for Deficiencies in Visual Representations

Annotated datasets for supervised learning tasks often contain incorrect gold annotations, i.e. label noise. To address this issue, many noisy label learning approaches incorporate metrics to filter out unreliable samples, for example using heuristics such as high loss or low confidence. However, when these metrics are integrated into larger pipelines, it becomes difficult to compare their effectiveness, and understand their individual contribution to reducing label noise. This paper directly compares popular sample metrics for detecting incorrect annotations in named entity recognition (NER). NER is commonly approached as token classification, so the metrics are calculated for each training token and we flag the incorrect ones by defining metrics thresholds. We compare the metrics based on (i) their accuracy in detecting the incorrect labels and (ii) the test scores when retraining a model using the cleaned dataset. We show that training dynamics metrics work the best overall. The best metrics effectively reduce the label noise across different noise types. The errors that the model has not yet memorized are more feasible to detect, and relabeling these tokens is a more effective strategy than excluding them from training.

Token-Level Metrics for Detecting Incorrect Gold Annotations in Named Entity Recognition

The growing emotional stress in modern society has increased the demand for Emotional Support Conversations (ESC). While Large Language Models (LLMs) show promise for ESC, they face two key challenges: (1) low strategy selection accuracy, and (2) preference bias, limiting their adaptability to users’ emotional needs. Existing supervised fine-tuning (SFT) struggles to address these issues, as it rigidly trains models on single gold-standard responses without modeling nuanced strategy trade-offs. To overcome these limitations, we propose a novel two-stage framework that optimizes strategy selection preferences at each dialogue turn. We first leverage Monte Carlo Tree Search to construct ESC-Pro, a high-quality preference dataset with turn-level strategy-response pairs. Then training on ESC-Pro with Chain-of-Strategy Optimization (CSO) improves both strategy accuracy and bias mitigation, enabling LLMs to generate more empathetic and contextually appropriate responses. Experiments on LLaMA-3.1-8B, Gemma-2-9B, and Qwen2.5-7B demonstrate that CSO outperforms standard SFT, highlighting the efficacy of fine-grained, turn-level preference modeling in ESC.

Chain of Strategy Optimization Makes Large Language Models Better Emotional Supporter

The frequency distribution of words in human-written texts roughly follows a simple mathematical form known as Zipf's law. Somewhat less well known is the related Heaps' law, which describes a sublinear power-law growth of vocabulary size with document size. We study the applicability of Zipf's and Heaps' laws to texts generated by Large Language Models (LLMs). We empirically show that Heaps’ and Zipf's laws only hold for LLM-generated texts in a narrow model-dependent temperature range. These temperatures have an optimal value close to t=1 for all the base models except the large Llama models, are higher for instruction-finetuned models and do not depend on the model size or prompting. This independently confirms the recent discovery of phase transitions in LLM-generated texts.

Zipf’s and Heaps’ Laws for Tokens and LLM-generated Texts

Understanding and solving complex reasoning tasks is vital for addressing the information needs of a user. Although dense neural models learn contextualised embeddings, they still underperform on queries containing negation. To understand this phenomenon, we study negation in both traditional neural information retrieval and LLM-based models. We (1) introduce a taxonomy of negation that derives from philosophical, linguistic, and logical definitions; (2) generate two benchmark datasets that can be used to evaluate the performance of neural information retrieval models and to fine-tune models for a more robust performance on negation; and (3) propose a logic-based classification mechanism that can be used to analyze the performance of retrieval models on existing datasets. Our taxonomy produces a balanced data distribution over negation types, providing a better training setup that leads to faster convergence on the NevIR dataset. Moreover, we propose a classification schema that reveals the coverage of negation types in existing datasets, offering insights into the factors that might affect the generalisation of fine-tuned models on negation.

A Comprehensive Taxonomy of Negation for NLP and Neural Retrievers

Language models built from various sources are the foundation of today’s NLP progress. However, for many low-resource languages, the diversity of domains is often limited, more biased to a religious domain, which impacts their performance when evaluated on distant and rapidly evolving domains such as social media. Domain adaptive pre-training (DAPT) and task-adaptive pre-training (TAPT) are popular techniques to reduce this bias through continual pre-training for BERT-based models, but they have not been explored for African multilingual gual encoders. In this paper, we explore DAPT and TAPT continual pertaining approaches for the African languages social media domain. We introduce AfriSocial - a large-scale social media and news domain corpus for continual pre-training on several African languages. Leveraging AfriSocial, we show that DAPT consistently improves performance on three subjective tasks: sentiment analysis, multi-label emotion, and hate speech classification, covering 19 languages from 1% to 30% F1 score. Similarly, leveraging TAPT on one task data improves performance on other related tasks. For example, training with unlabeled sentiment data (source) for a fine-grained emotion classification task (target) improves the baseline results by an F1 score ranging from 0.55% to 15.11%. Combining these two methods (i.e. DAPT + TAPT) further improves the overall performance.

AfroXLMR-Social: Adapting Pre-trained Language Models for African Languages Social Media Text

As Large Language Models (LLMs) gain expertise across diverse domains and modalities, scalable oversight becomes increasingly challenging, particularly when their capabilities may surpass human evaluators. Debate has emerged as a promising mechanism for enabling such oversight. In this work, we extend the debate paradigm to a multimodal setting, exploring its potential for weaker models to supervise and enhance the performance of stronger models. We focus on visual question answering (VQA), where two "sighted" expert vision-language models debate an answer, while a "blind" (text-only) judge adjudicates based solely on the quality of the arguments. In our framework, the experts defend only answers aligned with their beliefs, thereby obviating the need for explicit role-playing and concentrating the debate on instances of expert disagreement. Experiments on several multimodal tasks demonstrate that the debate framework consistently outperforms individual expert models. Moreover, judgments from weaker LLMs can help instill reasoning capabilities in vision-language models through finetuning.

Debating for Better Reasoning in Vision-Language Models

Large Vision-Language Models (LVLMs) have shown strong performance across multimodal tasks. However, they often produce hallucinations—text that is inconsistent with visual input, due to the limited ability to verify information in different regions of the image. To address this, we propose **Multi-Region Fusion Decoding (MRFD)**, a training-free decoding method that improves factual grounding by modeling inter-region consistency. MRFD identifies salient regions using cross-attention, generates initial responses for each, and computes reliability weights based on Jensen-Shannon Divergence (JSD) among the responses. These weights guide a consistency-aware fusion of per-region predictions, using region-aware prompts inspired by Chain-of-Thought reasoning. Experiments across multiple LVLMs and benchmarks show that MRFD significantly reduces hallucinations and improves response factuality without requiring model updates.

MRFD: Multi-Region Fusion Decoding with Self-Consistency for Mitigating Hallucinations in LVLMs

LLMs have been widely adopted to tackle many traditional NLP tasks. Their effectiveness remains uncertain in scenarios where pre-trained models have limited prior knowledge of a language. In this work, we examine LLMs' generalization in under-resourced settings through the task of orthographic normalization across Otomi language variants. We develop two approaches: a rule-based method using a finite-state transducer (FST) and an in-context learning (ICL) method that provides the model with string transduction examples. We compare the performance of FSTs and neural approaches in low-resource scenarios, providing insights into their potential and limitations. Our results show that while FSTs outperform LLMs in zero-shot settings, ICL enables LLMs to surpass FSTs, stressing the importance of combining linguistic expertise with machine learning in current approaches for low-resource scenarios

Downloads

Next from EMNLP 2025

X-LeBench: A Benchmark for Extremely Long Egocentric Video Understanding

Similar lecture

Character is Destiny: Can Persona-assigned Language Models Make Personal Choices?