China

The ability to critique is vital for models to self-improve and serve as reliable AI assistants. While extensively studied in language-only settings, multimodal critique of Large Multimodal Models (LMMs) remains underexplored despite their growing capabilities in tasks like captioning and visual reasoning. In this work, we introduce MM-Critic, a holistic benchmark for evaluating the critique ability of LMMs across three dimensions: basic, correlation, and comparison. Covering 8 task types and over 500 tasks, MM-Critic collects responses from LMMs with various model sizes. To enhance the evaluation reliability, we design expert-informed scoring rubrics that guide GPT-4o in annotating responses and generating reference critiques, which serve as anchors for trustworthy judgments. Extensive experiments validate the effectiveness of MM-Critic and provide a comprehensive assessment of leading LMMs’ critique capabilities. Further analysis reveals key insights, including the correlation between response quality and critique, and varying critique difficulty across evaluation dimensions.

EMNLP 2025

MM-CRITIC: A Holistic Evaluation of Large Multimodal Models as Multimodal Critique

poster

## 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*

[**Link to Conference Handbook**](https://drive.google.com/file/d/1johU5QqVVYO4RfH7QcIORr7qrVBdzdwC/view?usp=sharing)





<br>

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

In this paper, we investigate the relationship between Information Content (IC) and the squared norm of embeddings at the text level, focusing on the mechanisms and composition functions used to combine token embeddings. i) We formally derive two sufficient conditions for this correspondence to hold in embedding models. ii) We empirically examine the correspondence and the validity of these conditions at word level for both static and contextual embeddings and different subword token composition mechanisms. iii) Building on Shannon’s Constant Entropy Rate principle, we explore whether embedding mechanisms exhibit a linearly monotonic increase in information content as text length increases. Our formal analysis and experiments reveal that: i) At the word embedding level, models satisfy the sufficient conditions and show a strong correspondence when certain subword composition functions are applied. ii) Only scaled embedding averages proposed in this paper and certain information-theoretic composition functions preserve the correspondence. Some non-compositional representations—such as the CLS token in BERT or the EOS token in LLaMA—tend to converge toward a fixed point. The CLS token in ModernBERT, however, exhibits behavior that aligns more closely with the CER hypothesis.

On the Correspondence between the Squared Norm and Information Content in Text Embeddings

Retrieval-augmented generation (RAG) systems have advanced large language models (LLMs) in complex deep search scenarios requiring multi-step reasoning and iterative information retrieval. However, existing approaches face critical limitations that lack high-quality training trajectories or suffer from the distributional mismatches in simulated environments and prohibitive computational costs for real-world deployment. This paper introduces SimpleDeepSearcher, a lightweight yet effective framework that bridges this gap through strategic data engineering rather than complex training paradigms. Our approach synthesizes high-quality training data by simulating realistic user interactions in live web search environments, coupled with a multi-criteria curation strategy that optimizes the diversity and quality of input and output side. Experiments on five benchmarks across diverse domains demonstrate that SFT on only 871 curated samples yields significant improvements over RL-based baselines. Our work establishes SFT as a viable pathway by systematically addressing the data-scarce bottleneck, offering practical insights for efficient deep search systems. Our anonymous code is available at https://anonymous.4open.science/r/SimpleDeepSearcher.

SimpleDeepSearcher: Deep Information Seeking via Web-Powered Reasoning Trajectory Synthesis

As the utilization of language models in interdisciplinary, human-centered studies grow, the expectation of model capabilities continues to evolve. Beyond excelling at conventional tasks, models are recently expected to perform well on user-centric measurements involving confidence and human (dis)agreement- factors that reflect subjective preferences. While modeling of subjectivity plays an essential role in cognitive science and has been extensively studied, it remains under-explored within the NLP community. In light of this gap, we explore how language models can quantify subjectivity by conducting comprehensive experiments and analysis across various scenarios using both fine-tuned models and prompt-based large language models (LLMs). Our quantitative and qualitative experimental results demonstrate that personality traits and demographical information are critical for measuring subjectivity. However, our findings reveal that existing post-hoc calibration approaches often fail to produce satisfactory results. Furthermore, our in-depth analysis offers valuable insights for future research and development in the interdisciplinary studies of NLP and cognitive science.

Modeling Subjectivity in Cognitive Appraisal with Language Models

We explore Large Language Models (LLMs)' human motion knowledge through 3D avatar control. Given a motion instruction, we prompt LLMs to first generate a high-level movement plan with consecutive steps (**High-level Planning**), then specify body part positions in each step (**Low-level Planning**), which we linearly interpolate into avatar animations as a clear verification lens for human evaluators. Through carefully designed 20 representative motion instructions with full coverage of basic movement primitives and balanced body part usage, we conduct comprehensive evaluations including human assessment of both generated animations and high-level movement plans, as well as automatic comparison with oracle positions in low-level planning. We find that LLMs are strong at interpreting the high-level body movements but struggle with precise body part positioning. While breaking down motion queries into atomic components improves planning performance, LLMs have difficulty with multi-step movements involving high-degree-of-freedom body parts. Furthermore, LLMs provide reasonable approximation for general spatial descriptions, but fail to handle precise spatial specifications in text, and the precise spatial-temporal parameters needed for avatar control. Notably, LLMs show promise in conceptualizing creative motions and distinguishing culturally-specific motion patterns.

How Much Do Large Language Models Know about Human Motion? A Case Study in 3D Avatar Control

Few-shot intent detection (FSID) targets the classification of user queries into in-scope intent categories or detecting them as out-of-scope, with only a few or even zero labeled examples per class. Existing PLM-based methods struggle in low-resource situations; while LLM-based methods face high inference cost and label interference. To harness their complementary strengths, we propose the textbfFCSLM, a framework that collaborates a small prediction model with a large language model for the FSID task. During training, we leverage LLMs for data augmentation in self-supervised pretraining and supervised fine-tuning a task-specific prediction model. During inference, a multi-round reasoning process first applies the small prediction model to output candidate intents with uncertainty estimations, then invokes an LLM with enriched intent descriptions for refined prediction and OOS detection. Extensive experiments on three benchmark datasets demonstrate that our FCSLM outperforms strong competitors, achieving the new state-of-the-art performance in both intent classification and OOS detection.

On Collaborating Small and Large Models For Few-shot Intent Detection

Formulating information retrieval as a variant of generative modeling, specifically using autoregressive models to generate relevant identifiers for a given query, has recently attracted considerable attention. However, its application to personalized sticker retrieval remains largely unexplored and presents unique challenges: existing relevance-based generative retrieval methods typically lack personalization, leading to a mismatch between diverse user expectations and the retrieved results. To address this gap, we propose PEARL, a novel generative framework for personalized sticker retrieval, and make two key contributions: (i) To encode user-specific sticker preferences, we design a representation learning model to learn discriminative user representations. It is trained on three prediction tasks that leverage personal information and click history; and (ii) To generate stickers aligned with a user's query intent, we propose a novel intent-aware learning objective that prioritizes stickers associated with higher-ranked intents. Empirical results from both offline evaluations and online tests demonstrate that PEARL significantly outperforms state-of-the-art methods.

A Generative Framework for Personalized Sticker Retrieval

We introduce BiMediX2, a bilingual (Arabic-English) Bio-Medical EXpert Large Multimodal Model that supports text-based and image-based medical interactions. It enables multi-turn conversation in Arabic and English and supports diverse medical imaging modalities, including radiology, CT, and histology. To train BiMediX2, we curate BiMed-V, an extensive Arabic-English bilingual healthcare dataset consisting of 1.6M samples of diverse medical interactions. This dataset supports a range of medical Large Language Model (LLM) and Large Multimodal Model (LMM) tasks, including multi-turn medical conversations, report generation, and visual question answering (VQA). We also introduce BiMed-MBench, the first Arabic-English medical LMM evaluation benchmark, verified by medical experts. BiMediX2 demonstrates excellent performance across multiple medical LLM and LMM benchmarks, achieving state-of-the-art results in medical VQA, report generation, and summarization. On BiMed-MBench, BiMediX2 outperforms existing methods by over 9% in English and more than 20% in Arabic evaluations. Additionally, it surpasses GPT-4 by approximately 9% in UPHILL factual accuracy evaluations and excels in various medical VQA, report generation, and report summarization tasks. Our trained models, instruction set, and source code will be made publicly available.

BiMediX2 : Bio-Medical EXpert LMM for Diverse Medical Modalities

Augmenting neural machine translation with external memory at decoding time, in the form of k-nearest neighbors machine translation (k-NN MT), is a well-established strategy for increasing translation performance. k-NN MT retrieves, at the token level, a set of tokens that occurred in the most similar contexts recorded in a prepared data store, using hidden state representations of translation contexts as vector lookup keys. One of the main disadvantages of this method is the high computational cost and memory requirements. Since an exhaustive search is not feasible in large data stores practitioners commonly use approximate k-NN lookup, yet even such algorithms are a bottleneck. In contrast to research directions seeking to accelerate k-NN MT by reducing data store size or the number of lookup calls, we pursue an orthogonal direction based on the performance properties of approximate k-NN lookup data structures. In particular, we propose encouraging angular dispersion of the neural hidden representations of contexts. We show that improving dispersion leads to better balance in the retrieval data structures, accelerating retrieval and slightly improving translations.

Angular Dispersion Accelerates k-Nearest Neighbors Machine Translation

Recent advancements in large language models (LLMs) have revolutionized natural language processing through their remarkable capabilities in understanding and executing diverse tasks. While supervised fine-tuning, particularly in Retrieval-Augmented Generation (RAG) scenarios, has proven effective for enhancing task-specific performance, it often leads to catastrophic forgetting, where models lose their previously acquired knowledge and general capabilities. Existing solutions either require access to general instruction data or face limitations in preserving the model's original distribution. To overcome these limitations, we propose SelfAug, a novel self-distribution alignment method. By aligning distributions through the logits of input sequences, SelfAug preserves the model’s semantic distribution, thereby simultaneously mitigating catastrophic forgetting and improving downstream task performance. Through extensive experiments, we show that SelfAug achieves a better balance between downstream task learning and the retention of general capabilities compared to existing methods. Our comprehensive empirical analysis reveals a direct correlation between distribution shifts and the severity of catastrophic forgetting in RAG scenarios, particularly highlighting how the absence of RAG capabilities in general instruction tuning leads to significant distribution shifts during fine-tuning.

SelfAug: Mitigating Catastrophic Forgetting in Retrieval-Augmented Generation via Distribution Self-Alignment

Keyword extraction involves identifying the most descriptive words in a document, allowing automatic categorisation and summarisation of large quantities of diverse textual data. Relying on the insight that real-world keyword detection often requires handling of diverse content, we propose a novel supervised keyword extraction approach based on the mixture of experts (MoE) technique. MoE uses a learnable routing sub-network to direct information to specialised experts, allowing them to specialise in distinct regions of the input space. SEKE, a mixture of Specialised Experts for supervised Keyword Extraction, uses DeBERTa as the backbone model and builds on the MoE framework, where experts attend to each token, by integrating it with a Long short-term memory (LSTM) network, to allow successful extraction even on smaller corpora, where specialisation is harder due to lack of training data. The MoE framework also provides an insight into inner workings of individual experts, enhancing the explainability of the approach. We benchmark SEKE on multiple English datasets, achieving state-of-the-art performance compared to strong supervised and unsupervised baselines. Our analysis reveals that depending on data size and type, experts specialise in distinct syntactic and semantic components, such as punctuation, stopwords, parts-of-speech, or named entities. Code is available at https://anonymous.4open.science/r/SEKE_keyword_extraction-79FC.

Downloads

Next from EMNLP 2025

On the Correspondence between the Squared Norm and Information Content in Text Embeddings

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES