China

Multimodal large language models (MLLMs) have recently achieved state-of-the-art performance on tasks ranging from visual question answering to video understanding. However, existing studies have concentrated mainly on visual–textual misalignment, leaving largely unexplored the MLLMs&#39; ability to preserve an originally correct answer when confronted with misleading information. We reveal a response uncertainty phenomenon: across nine standard datasets, twelve state-of-the-art open-source MLLMs overturn a previously correct answer in 65% of cases after receiving a single deceptive cue. To systematically quantify this vulnerability, we propose a two-stage evaluation pipeline: (1) elicit each model’s original response on unperturbed inputs; (2) inject \emph{explicit} (false-answer hints) and \emph{implicit} (contextual contradictions) misleading instructions, and compute the \emph{misleading rate}—the fraction of correct-to-incorrect flips. Leveraging the most susceptible examples, we curate the Multimodal Uncertainty Benchmark (MUB), a collection of image–question pairs stratified into low, medium, and high difficulty based on how many of twelve state-of-the-art MLLMs they mislead. Extensive evaluation on twelve open-source and five closed-source models reveals a high uncertainty: average misleading rates exceed 86%, with explicit cues over 67.19% and implicit cues over 80.67%. To reduce the misleading rate, we then fine-tune all open-source MLLMs on a compact 2,000-sample mixed-instruction dataset, reducing misleading rates to 6.97% (explicit) and 32.77% (implicit), boosting consistency by nearly 29.37% on highly deceptive inputs, and slightly improving accuracy on standard benchmarks.

EMNLP 2025

Exploring Response Uncertainty in MLLMs: An Empirical Evaluation under Misleading Scenarios

misleading instructions

response uncertainty

multimodal large language models

benchmark

evaluation

robustness

## 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.

Multimodal large language models (MLLMs) have recently achieved state-of-the-art performance on tasks ranging from visual question answering to video understanding. However, existing studies have concentrated mainly on visual–textual misalignment, leaving largely unexplored the MLLMs' ability to preserve an originally correct answer when confronted with misleading information. We reveal a response uncertainty phenomenon: across nine standard datasets, twelve state-of-the-art open-source MLLMs overturn a previously correct answer in 65% of cases after receiving a single deceptive cue. To systematically quantify this vulnerability, we propose a two-stage evaluation pipeline: (1) elicit each model’s original response on unperturbed inputs; (2) inject \emph{explicit} (false-answer hints) and \emph{implicit} (contextual contradictions) misleading instructions, and compute the \emph{misleading rate}—the fraction of correct-to-incorrect flips. Leveraging the most susceptible examples, we curate the Multimodal Uncertainty Benchmark (MUB), a collection of image–question pairs stratified into low, medium, and high difficulty based on how many of twelve state-of-the-art MLLMs they mislead. Extensive evaluation on twelve open-source and five closed-source models reveals a high uncertainty: average misleading rates exceed 86%, with explicit cues over 67.19% and implicit cues over 80.67%. To reduce the misleading rate, we then fine-tune all open-source MLLMs on a compact 2,000-sample mixed-instruction dataset, reducing misleading rates to 6.97% (explicit) and 32.77% (implicit), boosting consistency by nearly 29.37% on highly deceptive inputs, and slightly improving accuracy on standard benchmarks.

poster

Recent advances in retrieval-augmented generation (RAG) have substantially improved question-answering systems, particularly for factoid '5Ws' questions. However, significant challenges remain when addressing '1H' questions, specifically how-to questions, which are integral for decision-making and require dynamic, step-by-step responses. The key limitation lies in the prevalent data organization paradigm, chunk, which commonly divides documents into fixed-size segments, and disrupts the logical coherence and connections within the context. To address this, we propose THREAD, a novel data organization paradigm enabling systems to handle how-to questions more effectively. Specifically, we introduce a new knowledge granularity, 'logic unit' (LU), where large language models transform documents into more structured and loosely interconnected LUs. Extensive experiments across both open-domain and industrial settings show that THREAD outperforms existing paradigms significantly, improving the success rate of handling how-to questions by 21% to 33%. Additionally, THREAD demonstrates high adaptability across diverse document formats, reducing retrieval information by up to 75% compared to chunk, and also shows better generalizability to '5Ws' questions, such as multi-hop questions, outperforming other paradigms.

Thread: A Logic-Based Data Organization Paradigm for How-To Question Answering with Retrieval Augmented Generation

Designing crystal materials with desired physicochemical properties remains a fundamental challenge in materials science. While large language models (LLMs) have demonstrated strong in-context learning (ICL) capabilities, existing LLM-based crystal generation approaches are limited to zero-shot scenarios and are unable to benefit from few-shot scenarios. In contrast, human experts typically design new materials by modifying relevant known structures which aligns closely with the few-shot ICL paradigm. Motivated by this, we propose CrystalICL, a novel model designed for few-shot crystal generation. Specifically, we introduce a space-group based crystal tokenization method, which effectively reduces the complexity of modeling crystal symmetry in LLMs. We further introduce a condition-structure aware hybrid instruction tuning framework and a multi-task instruction tuning strategy, enabling the model to better exploit ICL by capturing structure–property relationships from limited data. Extensive experiments on four crystal generation benchmarks demonstrate the superiority of CrystalICL over the leading baseline methods on conditional and unconditional generation tasks.

CrystalICL: Enabling In-Context Learning for Crystal Generation

Test-Time Scaling (TTS) is a promising new approach to progressively elicit the model’s intelligence during inference. Recently, training-based test-time scaling methods, such as continued reinforcement learning (RL), have further surged in popularity, while training-free methods have faded from prominence. However, the additional train-time computation amplifies the burden on test-time scaling. In this paper, we design a finer-grained sequential scaling method called Conditional Step-level Self-refinement with the support of process verification. On top of its effectiveness, we further combine it with other classical parallel scaling methods at the step level, to introduce a novel paradigm called Hybrid Test-Time Scaling. Extensive experiments on five instruction-tuned LLMs across different scales (3B-14B) and families demonstrate that this hybrid strategy incorporating multiple training-free test-time scaling methods at finer granularity has considerable potential for expanding the reasoning performance boundaries of LLMs.

Step-level Verifier-guided Hybrid Test-Time Scaling for Large Language Models

Vision-language models (VLMs) achieve promising results in medical reasoning but struggle with hallucinations, vague descriptions, Inconsistent logic and poor localization. To address this, we propose a agent framework named Medical Visual Reasoning Agent (\textbf{Med-VRAgent}). The approach is based on Visual Guidance and Self-Reward paradigms and Monte Carlo Tree Search (MCTS). By combining the Visual Guidance with tree search, Med-VRAgent improves the medical visual reasoning capabilities of VLMs. We use the trajectories collected by Med-RAgent as feedback to further improve the performance by fine-tuning the VLMs with the proximal policy optimization (PPO) objective. Experiments on multiple medical VQA benchmarks demonstrate that our method outperforms existing approaches.

Med-VRAgent: A Framework for Medical Visual Reasoning-Enhanced Agents

Recent advances in Video Large Language Models (VLLMs) have significantly enhanced their ability to understand video content. Nonetheless, processing long videos remains challenging due to high computational demands and the redundancy present in the visual data. In this work, we propose textbfMETok, a training-free, textbfMulti-stage textbfEvent-based textbfToken compression framework designed to accelerate VLLMs' inference while preserving accuracy. METok progressively eliminates redundant visual tokens across three critical stages: (1) event-aware compression during vision encoding, (2) hierarchical token pruning in the prefilling stage based on semantic alignment and event importance, and (3) a decoding-stage KV Cache optimization that further reduces memory consumption. Our experiments on diverse video benchmarks demonstrate that METok achieves an optimal trade-off between efficiency and accuracy by dynamically selecting informative visual tokens. For instance, equipping LongVA-7B with METok realizes an 80.6\% FLOPs reduction and 93.5\% KV Cache memory savings, all while maintaining comparable or even superior accuracy.

METok: Multi-Stage Event-based Token Compression for Efficient Long Video Understanding

Multimodal Large Language Models (MLLMs) have achieved strong performance across vision-language tasks, but suffer from significant computational overhead due to the quadratic growth of attention computations with the number of multimodal tokens. Though efforts have been made to prune tokens in MLLMs, *they lack a fundamental understanding of how MLLMs process and fuse multimodal information*. Through systematic analysis, we uncover a three-stage cross-modal interaction process: (1) Shallow layers recognize task intent, with visual tokens acting as passive attention sinks; (2) Cross-modal fusion occurs abruptly in middle layers, driven by a few critical visual tokens; (3) Deep layers discard vision tokens, focusing solely on linguistic refinement. Based on these findings, we propose *VisiPruner*, a training-free pruning framework that reduces **99.9%** of vision-related attention computations and **62.8%** of FLOPs while maintaining performance. It significantly outperforms existing token pruning methods and generalizes across diverse MLLMs. Beyond pruning, our insights further provide actionable guidelines for training efficient MLLMs by aligning model architecture with its intrinsic layer-wise processing dynamics.

VisiPruner: Decoding Discontinuous Cross-Modal Dynamics for Efficient Multimodal LLMs

With the emergence of new topics on social media as sources of rumor propagation, addressing the domain shift between the source and target domain and the target domain samples scarcity remains a crucial task in cross-domain rumor detection. Traditional deep learning-based methods and LLM-based methods are mostly focused on the in-domain condition, thus having poor performance in cross-domain setting. Existing domain adaptation rumor detection approaches ignore the data generalization differences and rely on a large amount of unlabeled target domain samples to achieve domain adaptation, resulting in less effective on emerging topic rumor detection. In this paper, we propose a Gradient Coherence guided Meta-Learning approach (GCML) for emerging topics rumor detection. Firstly, we calculate the task generalization score of each source task (sampled from source domain) from a gradient coherence perspective, and selectively learn more ''generalizable'' tasks that are more beneficial in adapting to the target domain. Secondly, we leverage meta-learning to alleviate the target domain samples scarcity, which utilizes task generalization scores to re-weight meta-test gradients and adaptively updates learning rate. Extensive experimental results on real-world datasets show that our method substantially outperforms SOTA baselines.

GCML: Gradient Coherence Guided Meta-Learning for Cross-Domain Emerging Topic Rumor Detection

Processing long contexts is increasingly important for Large Language Models (LLMs) in tasks like multi-turn dialogues, code generation, and document summarization. This paper addresses the challenges of achieving high long-context performance, low computational complexity, and compatibility with pretrained models -- collectively termed the ``impossible triangle''. We introduce E2LLM (Encoder Elongated Large Language Models), a novel approach that effectively navigates this paradox. E2LLM divides long contexts into chunks, compresses each into soft prompts using a pretrained text encoder, and aligns these representations with a decoder-only LLM via an adapter. To enhance the LLM's reasoning with these soft prompts, we employ two training objectives: encoder output reconstruction and long-context instruction fine-tuning. Extensive experiments reveal that E2LLM not only outperforms 8 state-of-the-art (SOTA) methods in effectiveness and efficiency for document summarization and question answering, but also achieves the best performance on LongBench v2 among models of comparable size. Code will be available upon publication.

E2LLM: Encoder Elongated Large Language Models for Long-Context Understanding and Reasoning

Detecting LLM-generated text in specialized and high-stakes domains like medicine and law is crucial for combating misinformation and ensuring authenticity. However, current zero-shot detectors, while effective on general text, often fail when applied to specialized content due to domain shift. We provide a theoretical analysis showing this failure is fundamentally linked to the KL divergence between human, detector, and source text distributions. To address this, we propose DivScore, a zero-shot detection framework using normalized entropy-based scoring and domain knowledge distillation to robustly identify LLM-generated text in specialized domains. Experiments on medical and legal datasets show that DivScore consistently outperforms state-of-the-art detectors, with 14.4% higher AUROC and 64.0% higher recall at 0.1% false positive rate threshold. In adversarial settings, DivScore demonstrates superior robustness than other baselines, achieving on average 22.8% advantage in AUROC and 29.5% in recall.

DivScore: Zero-Shot Detection of LLM-Generated Text in Specialized Domains

Knowledge Tracing (KT) aims to model a student’s learning state over time and predict their future performance. However, traditional KT methods often face challenges in explainability, scalability, and effective modeling of complex knowledge dependencies. While Large Language Models (LLMs) present new avenues for KT, their direct application often struggles with generating structured, explainable student representations and lacks mechanisms for continuous, task-specific refinement. To address these gaps, we propose Collaborative Iterative Knowledge Tracing (CIKT), a framework that harnesses LLMs to enhance both prediction accuracy and explainability. CIKT employs a dual-component architecture: an Analyst generates dynamic, explainable user profiles from student historical responses, and a Predictor utilizes these profiles to forecast future performance. The core of CIKT is a synergistic optimization loop. In this loop, the Analyst is iteratively refined based on the predictive accuracy of the Predictor, which conditions on the generated profiles, and the Predictor is subsequently retrained using these enhanced profiles. Evaluated on multiple educational datasets, CIKT demonstrates significant improvements in prediction accuracy, offers enhanced explainability through its dynamically updated user profiles, and exhibits improved scalability. Our work presents a robust and explainable solution for advancing knowledge tracing systems, effectively bridging the gap between predictive performance and model transparency.

Downloads

Next from EMNLP 2025

Thread: A Logic-Based Data Organization Paradigm for How-To Question Answering with Retrieval Augmented Generation

Similar lecture

Unmasking Deceptive Visuals: Benchmarking Multimodal Large Language Models on Misleading Chart Question Answering