China

We introduce VLM-Lens, a toolkit designed to enable systematic benchmarking, analysis, and interpretation of vision-language models (VLMs) by supporting the extraction of intermediate outputs from any layer during the forward pass of open-source VLMs. VLM-Lens provides a unified, YAML-configurable interface that abstracts away model-specific complexities and supports user-friendly operation across diverse VLMs. It currently supports 10 state-of-the-art base VLMs and their variants, and is extensible to accommodate new models without changing the core logic. The toolkit integrates easily with various interpretability and analysis methods. We demonstrate the utility of VLM-Lens through prob- ing experiments, revealing systematic differences in the hidden representations of VLMs across layers, and we report its performance in terms of time and memory efficiency. VLM-Lens is released as an open-sourced library to accelerate community efforts in understanding and improving VLMs. Github: https://github.com/compling-wat/vlm-lens

EMNLP 2025

From Behavioral Performance to Internal Competence: Interpreting Vision-Language Models with VLM-Lens

open-sourced projects

vision-language models

toolkit

interpretability

## 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.

demo

Graphical User Interface (GUI) agents have gained substantial attention due to their impressive capabilities to complete tasks through multiple interactions within GUI environments. However, existing agents primarily focus on enhancing the accuracy of individual actions and often lack effective mechanisms for detecting and recovering from errors. To address these shortcomings, we propose the BacktrackAgent, a robust framework that incorporates a backtracking mechanism to improve task completion efficiency. BacktrackAgent includes verifier, judger, and reflector components as modules for error detection and recovery, while also applying judgment rewards to further enhance the agent's performance. Additionally, we develop a training dataset specifically designed for the backtracking mechanism, which considers the outcome pages after action executions. Experimental results show that BacktrackAgent has achieved performance improvements in both task success rate and step accuracy on Mobile3M and Auto-UI benchmarks. Our data and code will be released upon acceptance.

BacktrackAgent: Enhancing GUI Agent with Error Detection and Backtracking Mechanism

The full-size MLPs and the projection layers in attention introduce tremendous model sizes of large language models (LLMs), imposing extremely demanding needs of computational resources in the pre-training stage. However, we empirically observe that the activations of pre-trained LLMs exhibit low-rank property. Motivated by such observations, we propose **CoLA** and its memory-efficient implementation, **CoLA-M**, to replace these full-size layers with compute-efficient auto-encoders that naturally enforce low-rank activations throughout training. This fundamental architectural change eliminates the activation redundancy and significantly boosts model capacity and training efficiency. Experiments on LLaMA models with 60 million to 7 billion parameters show that CoLA reduces the computing cost by bf 2pmbtimes and improves training throughput by bf 1.86pmbtimes while maintaining full-rank level performance. CoLA-M further squeezes memory cost without sacrificing throughput, offering a pre-training approach with collectively superior parameter, computing, and memory efficiency. The LLMs produced are also bf 2pmbtimes smaller, enabling faster inference with lower memory cost on resource-constrained platforms.

CoLA: Compute-Efficient Pre-Training of LLMs via Low-Rank Activation

The rapid development of Multimodal Large Reasoning Models (MLRMs) has demonstrated broad application potential, yet their safety and reliability remain critical concerns that require systematic exploration. To address this gap, we conduct a comprehensive and systematic safety evaluation of 11 MLRMs across 5 benchmarks and unveil prevalent safety degradation phenomena in most advanced models. Moreover, our analysis reveals distinct safety patterns across different benchmarks: significant safety degradation is observed across jailbreak robustness benchmarks, whereas safety-awareness benchmarks demonstrate less pronounced degradation. In particular, a long thought process in some scenarios even enhances safety performance. Therefore, it is a potential approach to address safety issues in MLRMs by leveraging the intrinsic reasoning capabilities of the model to detect unsafe intent. To operationalize this insight, we construct a multimodal tuning dataset that incorporates a safety-oriented thought process. Experimental results from fine-tuning existing MLRMs with this dataset effectively enhances the safety on both jailbreak robustness and safety-awareness benchmarks. This study provides a new perspective for developing safe MLRMs. Our dataset is available at \url{https://anonymous.4open.science/r/safety-70A4}.

Think in Safety: Unveiling and Mitigating Safety Alignment Collapse in Multimodal Large Reasoning Model

Counter-arguments (CAs) are a good means to improve the critical-thinking skills of learners, especially given that one has to thoroughly consider the logic of initial arguments (IA) when composing their CA. Although several tasks have been created for identifying the logical structure of CAs, no prior work has focused on capturing multiple interpretations of logical structures due to their complexity. In this work, we create CALSA+, a dataset consisting of 134 CAs annotated with 13 logical predicate questions. CALSA+ contains 1,742 instances annotated by 3 expert annotators (5,226 total annotations) with good agreement (Krippendorff α=0.46). Using CALSA+, we train a model with Reinforcement Learning with Verifiable Rewards (RLVR) to identify multiple logical interpretations and show that models trained with RLVR can perform on par with much bigger proprietary models. Our work is the first to attempt to annotate all the interpretations of logical structure on top of CAs. We publicly release our dataset to facilitate research in CA logical structure detection.

Identification of Multiple Logical Interpretations in Counter-Arguments

We revisit hierarchical bracketing encodings from a practical perspective in the context of dependency graph parsing. The approach encodes graphs as sequences, enabling linear-time parsing with n tagging actions, and still representing reentrancies, cycles, and empty nodes. Compared to existing graph linearizations, this representation substantially reduces the label space while preserving structural information. We evaluate it on a multilingual and multi-formalism benchmark, showing competitive results and consistent improvements over other methods in exact match accuracy.

Hierarchical Bracketing Encodings Work for Dependency Graphs

LLM-based translation agents have achieved highly human-like translation results and are capable of handling longer and more complex contexts with greater efficiency. However, they are typically limited to text-only inputs. In this paper, we introduce ViDove, a translation agent system designed for multimodal input. Inspired by the workflow of human translators, ViDove leverages visual and contextual background information to enhance the translation process. Additionally, we integrate a multimodal memory system and long-short term memory modules enriched with domain-specific knowledge, enabling the agent to perform more accurately and adaptively in real-world scenarios. As a result, ViDove achieves significantly higher translation quality in both subtitle generation and general translation tasks, with a 28\% improvement in BLEU scores and a 15\% improvement in SubER compared to previous state-of-the-art baselines. Moreover, we introduce DoveBench, a new benchmark for long-form automatic video subtitling and translation, featuring 17 hours of high-quality, human-annotated data. Our demo is available here: https://vidove.willbe03.com/

ViDove: A Translation Agent System with Multimodal Context and Memory-Augmented Reasoning

High-quality schematic diagrams, which provide a conceptual overview of the research, play a crucial role in summarizing and clarifying a study’s core ideas. However, creating these diagrams is time-consuming for authors and remains challenging for current AI systems, as it requires both a deep understanding of the paper’s content and a strong sense of visual design. To address this, we introduce SCISKETCH, an open-source framework that supports two automated workflows for schematic diagram generation using foundation models, shown in Figure 1. 1) In the graphic-code-based workflow, SCISKETCH follows a two-stage pipeline: it first produces a layout plan expressed in a graphical code language with a self-refinement and self-verification mechanism. It then integrates empirical images and symbolic icons to create a visually coherent, informative diagram. 2) In the image-based workflow, SCISKETCH directly synthesizes the diagram image through image generation with a self-refinement mechanism. Through both automatic and human evaluations, we show that SCISKETCH outperforms several state-of-the-art foundation models, including GPT-4o, and Gemini-2.5-Pro, in generating schematic diagrams for scientific papers. We make SCISKETCH fully open-sourced, providing researchers with an accessible, extensible tool for high-quality schematic diagram generation in scientific fields.

SciSketch: An Open-source Framework for Automated Schematic Diagram Generation in Scientific Papers

The rapid adoption of LLM-based conversational systems is already transforming the landscape of educational technology. However, the current state-of-the-art learning models do not take into account the student's affective states. Multiple studies in educational psychology support the claim that positive or negative emotional states can impact a student's learning capabilities. To bridge this gap, we present MathBuddy, an emotionally aware LLM-powered Math Tutor, which dynamically models the student's emotions and maps them to relevant pedagogical strategies, making the tutor-student conversation a more empathetic one. The student's emotions are captured from the conversational text as well as from their facial expressions. The student's emotions are aggregated from both modalities to confidently prompt our LLM Tutor for an emotionally-aware response. We have effectively evaluated our model using automatic evaluation metrics across eight pedagogical dimensions (Maurya et al., 2025) and user studies. We report a massive 23 point performance gain using the win rate (Macina et al., 2025) and a 3 point gain at an overall level using DAMR scores which strongly supports our hypothesis of improving LLM-based tutor's pedagogical abilities by modeling students' emotions.

MathBuddy: A Multimodal System for Affective Math Tutoring

Retrieval-augmented generation (RAG) often falls short when retrieved context includes confusing semi-relevant passages, or when answering questions require deep contextual understanding and reasoning. We propose an efficient fine-tuning framework, called PrismRAG, that (i) trains the model with distractor-aware QA pairs mixing gold evidence with subtle distractor passages, and (ii) instills reasoning-centric habits that make the LLM plan, rationalize, and synthesize without relying on extensive human engineered instructions. Evaluated across 12 open-book RAG QA benchmarks spanning diverse application domains and scenarios, PrismRAG improves average factuality by 5.4%, outperforming state-of-the-art solutions. Our method is being deployed in production.

PrismRAG: Boosting RAG Factuality with Distractor Resilience and Strategized Reasoning

Retrieval-augmented generation (RAG) aims to reduce hallucinations by grounding responses in external context, yet LLMs still frequently introduce unsupported information or contradictions even when provided with relevant context. In this paper, we present two complementary efforts at CompanyX to measure and benchmark LLM faithfulness in RAG. First, we describe our public LeaderboardX, which has tracked hallucination rates for over 140 LLMs since 2023 using our ModelX hallucination-detection model. Motivated by limitations observed in current detection methods, we introduce JudgeX, an LLM-as-a-judge framework that leverages a pool of diverse human-annotated hallucination examples to substantially improve the automated hallucination evaluation of LLMs. We introduce an enhanced hallucination leaderboard centered on JudgeX that benchmarks LLMs on RAG faithfulness in summarization, question-answering, and data-to-text generation tasks, alongside our current hallucination leaderboard. Together, these resources enable a more reliable benchmarking of LLM hallucinations in RAG and support the development of more trustworthy generative AI systems.

Downloads

Next from EMNLP 2025

BacktrackAgent: Enhancing GUI Agent with Error Detection and Backtracking Mechanism

Similar lecture

VIBE: Can a VLM Read the Room?

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from EMNLP 2025

BacktrackAgent: Enhancing GUI Agent with Error Detection and Backtracking Mechanism

Similar lecture

VIBE: Can a VLM Read the Room?

Downloads