China

LLM-as-a-Judge employs large language models (LLMs), such as GPT-4, to evaluate the quality of LLM-generated responses, gaining popularity for its cost-effectiveness and strong alignment with human evaluations. However, training proxy judge models using evaluation data generated by powerful teacher models introduces a critical yet previously overlooked issue: teacher preference bias, where the proxy judge model learns a biased preference for responses from the teacher model. To tackle this problem, we propose a novel setting that incorporates an additional assistant model, which is not biased toward the teacher model&#39;s responses, to complement the training data. Building on this setup, we introduce AGDe-Judge, a three-stage framework designed to debias from both the labels and feedbacks in the training data. Extensive experiments demonstrate that AGDe-Judge effectively reduces teacher preference bias while maintaining strong performance across six evaluation benchmarks. \footnote{Code is available at \url{https://anonymous.4open.science/r/AGDe-Judge-E352 }.}.

EMNLP 2025

Assistant-Guided Mitigation of Teacher Preference Bias in LLM-as-a-Judge

llm-as-a-judge

bias

evaluation

## 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.

LLM-as-a-Judge employs large language models (LLMs), such as GPT-4, to evaluate the quality of LLM-generated responses, gaining popularity for its cost-effectiveness and strong alignment with human evaluations. However, training proxy judge models using evaluation data generated by powerful teacher models introduces a critical yet previously overlooked issue: teacher preference bias, where the proxy judge model learns a biased preference for responses from the teacher model. To tackle this problem, we propose a novel setting that incorporates an additional assistant model, which is not biased toward the teacher model's responses, to complement the training data. Building on this setup, we introduce AGDe-Judge, a three-stage framework designed to debias from both the labels and feedbacks in the training data. Extensive experiments demonstrate that AGDe-Judge effectively reduces teacher preference bias while maintaining strong performance across six evaluation benchmarks. \footnote{Code is available at \url{https://anonymous.4open.science/r/AGDe-Judge-E352 }.}.

poster

The Key-Value (KV) cache reading latency increases significantly with context lengths, hindering the efficiency of long-context LLM inference. To address this, previous works propose retaining a small fraction of KV cache based on token importance. For example, KV eviction uses static heuristics to retain tokens, while KV retrieval dynamically selects query-relevant tokens for more adaptive cache management. However, we observe that important tokens are often sparsely distributed across the long context. This sparsity makes existing page-level KV retrieval inaccurate, as each page may include irrelevant tokens and miss critical ones. In this work, we propose Fier, a **Fi**ne-Grained and **E**fficient KV cache **R**etrieval method. Fier uses 1-bit quantized keys to estimate the importance of each token, resulting in efficient and precise retrieval. Experiments show that Fier matches full KV performance using only 11\% of the cache budget across various long-context tasks, reducing decoding latency by 1.2times to 1.5times.

FIER: Fine-Grained and Efficient KV Cache Retrieval for Long-context LLM Inference

Figure captions are crucial for helping readers understand and remember a figure's key message. Many models have been developed to generate these captions, helping authors compose better quality captions more easily. Yet, authors almost always need to revise generic AI-generated captions to match their writing style and the domain's style, highlighting the need for personalization. Despite language models' personalization (LaMP) advances, these technologies often focus on text-only settings and rarely address scenarios where both inputs and profiles are multimodal. This paper introduces LaMP-Cap, a dataset for personalized figure caption generation with multimodal figure profiles. For each target figure, LaMP-Cap provides not only the needed inputs, such as figure images, but also up to three other figures from the same document—each with its image, caption, and figure-mentioning paragraphs—as a profile to characterize the context. Experiments with four LLMs show that using profile information consistently helps generate captions closer to the original author-written ones. Ablation studies reveal that images in the profile are more helpful than figure-mentioning paragraphs, highlighting the advantage of using multimodal profiles over text-only ones.

LaMP-Cap: Personalized Figure Caption Generation With Multimodal Figure Profiles

Retrieval-Augmented Generation (RAG) has emerged as a widely adopted approach for knowledge injection during large language model (LLM) inference in recent years. However, due to their limited ability to exploit fine-grained inter-document relationships, current RAG implementations face challenges in effectively addressing the retrieved noise and redundancy content, which may cause error in the generation results. To address these limitations, we propose an **E**fficient **D**ynamic **C**lustering-based document **C**ompression framework (**EDC²-RAG**) that utilizes latent inter-document relationships while simultaneously removing irrelevant information and redundant content. We validate our approach, built upon GPT-3.5-Turbo and GPT-4o-mini, on widely used knowledge-QA and Hallucination-Detection datasets. Experimental results show that our method achieves consistent performance improvements across various scenarios and experimental settings, demonstrating strong robustness and applicability. Our code and datasets are available at \url{https://anonymous.4open.science/r/EDC-2-RAG-5F54}.

Efficient Dynamic Clustering-Based Document Compression for Retrieval-Augmented-Generation

This paper introduces a cross-lingual dubbing system that translates speech from one language to another while preserving key characteristics such as duration, speaker identity, and speaking speed. Despite the strong translation quality of existing speech translation approaches, they often overlook the transfer of speech patterns, leading to mismatches with source speech and limiting their suitability for dubbing applications. To address this, we propose a discrete diffusion-based speech-to-unit translation model with explicit duration control, enabling time-aligned translation. We then synthesize speech based on the predicted units and source identity with a conditional flow matching model. Additionally, we introduce a unit-based speed adaptation mechanism that guides the translation model to produce speech at a rate consistent with the source, without relying on any text. Extensive experiments demonstrate that our framework generates natural and fluent translations that align with the original speech's duration and speaking pace, while achieving competitive translation performance.

Dub-S2ST: Textless Speech-to-Speech Translation for Seamless Dubbing

Large language models (LLMs) are increasingly used for data generation. However, creating evaluation benchmarks raises the bar for this emerging paradigm. Benchmarks must target specific phenomena, penalize exploiting shortcuts, and be challenging. Through two case studies, we ask whether LLMs are ready to meet these demands---by generating reasoning-over-text benchmarks and comparing them to those that were created through careful crowdsourcing. Specifically, we evaluate both the *validity* and *difficulty* of LLM-generated versions of two high-quality reading comprehension datasets: CondaQA, which evaluates reasoning about negation, and DROP, which targets reasoning about quantities. We find that prompting LLMs can produce variants of these datasets that are often valid according to the annotation guidelines, at a fraction of the cost of the original crowdsourcing effort. However, we show that they are *less challenging for LLMs* than their human-authored counterparts. This finding sheds light on what may have been lost by generating evaluation data with LLMs, and calls for critically reassessing the immediate use of this increasingly prevalent approach to benchmark creation.

What Has Been Lost with Synthetic Evaluation?

As large language models (LLMs) become increasingly integrated into personal writing tools, a critical question arises: Can they faithfully imitate an individual’s writing style from just a few examples? Personal style is often subtle and implicit, making it difficult to specify through prompts yet essential for user-aligned generation. This work presents a comprehensive evaluation of state-of-the-art LLMs’ ability to mimic personal writing styles via in-context learning from a small number of user-authored samples. We introduce an ensemble of complementary metrics—including authorship attribution, style matching, and AI detection—to robustly assess style imitation. Our evaluation spans over 40,000 generations per model across domains such as news, email, forums, and blogs, covering writing samples from more than 400 real-world authors. Results show that while LLMs can approximate user style in structured formats like news and email, they struggle with nuanced, informal writing in blogs and forums. Further analysis on various prompting strategies such as number of demonstrations reveal key limitations in effective personalization. Our findings highlight a fundamental gap in personalized LLM adaptation and the need for improved techniques to support implicit, style-consistent generation.

Catch Me If You Can? Not Yet: LLMs Still Struggle to Imitate the Implicit Writing Styles of Everyday Authors

With the rapid development of large language models (LLMs) in math reasoning, the accuracy of models on existing math benchmarks has gradually approached 90\% or even higher. More challenging math benchmarks are hence urgently in need to satisfy the increasing evaluation demands. To bridge this gap, we propose HighMATH. Problems in HighMATH are collected according to 3 criteria: problem complexity, knowledge domain diversity and fine-grained annotations. We collect 5,293 problems from Chinese senior high school mathematics exams published in 2024, covering 8 subjects and 7 levels of difficulty, with each problem involving an average of more than 2.4 knowledge points. We conduct a thorough evaluation of latest LLMs on the curated HighMATH, including o1-like models. Evaluation results demonstrate that the accuracy of advanced LLMs on HighMATH is significantly lower than that on previous math reasoning benchmarks. This gap even exceeds 30\%. Our results also suggest that properly trained smaller LLMs may have great potential in math reasoning. Our data is available at https://anonymous.4open.science/r/data-VSDDDGvs/.

HighMATH: Evaluating Math Reasoning of Large Language Models in Breadth and Depth

Digital communication has become the cornerstone of modern interaction, enabling rapid, accessible, and interactive exchanges. However, individuals with lower academic literacy often face significant barriers, exacerbating the "digital divide." In this work, we introduce a novel, universal ideographic metalanguage designed as an innovative communication framework that transcends academic, linguistic, and cultural boundaries. Our approach leverages principles of Neuro-symbolic AI, combining neural-based large language models (LLMs) enriched with world knowledge and symbolic knowledge heuristics grounded in the linguistic theory of Natural Semantic Metalanguage (NSM). This enables the semantic decomposition of complex ideas into simpler, atomic concepts. Adopting a human-centric, collaborative methodology, we engaged over 200 semi-literate participants in defining the problem, selecting ideographs, and validating the system. With over 80% semantic comprehensibility, an accessible learning curve, and universal adaptability, our system effectively serves underprivileged populations with limited formal education.

NIM: Neuro-symbolic Ideographic Metalanguage for Inclusive Communication

Retrieval-Augmented Generation (RAG) integrates external knowledge with Large Language Models (LLMs) to enhance factual correctness and mitigate hallucination. However, dense retrievers often become the bottleneck of RAG systems due to their limited parameters compared to LLMs and their inability to perform step-by-step reasoning. While prompt-based iterative RAG attempts to address these limitations, it is constrained by human-designed workflows. To address these limitations, we propose textbfR3-RAG, which uses textbfReinforcement learning to make the LLM learn how to textbfReason and textbfRetrieve step by step, thus retrieving comprehensive external knowledge and leading to correct answers. R3-RAG is divided into two stages. We first use cold start to make the model learn the manner of iteratively interleaving reasoning and retrieval. Then we use reinforcement learning to further harness its ability to better explore the external retrieval environment. Specifically, we propose two rewards for R3-RAG: 1) answer correctness for outcome reward, which judges whether the trajectory leads to a correct answer; 2) relevance-based document verification for process reward, encouraging the model to retrieve documents that are relevant to the user question, through which we can let the model learn how to iteratively reason and retrieve relevant documents to get the correct answer. Experimental results show that R3-RAG significantly outperforms baselines and can transfer well to different retrievers.

R3-RAG: Learning Step-by-Step Reasoning and Retrieval for LLMs via Reinforcement Learning

Large-scale models have achieved state-of-the-art performance in automatic speech recognition (ASR), but their high memory and computation demands pose significant challenges for deployment. To address these challenges, weight-only quantization is widely adopted in large-scale models, where weights dominate memory usage, as it enables efficient compression with minimal accuracy degradation compared to activation quantization. Accordingly, most prior quantization studies for ASR models have focused on weights and employed quantization-aware training (QAT) to restore accuracy. However, QAT incurs substantial additional training costs, posing clear limitations for practical application to large-scale models. Moreover, despite the varying quantization sensitivity across layers, mixed-precision quantization (MPQ) remains underexplored in ASR. In this paper, we propose GenPTQ, a mixed-precision post-training quantization method that optimizes the trade-off among accuracy, model size, and optimization cost by leveraging gradient-based sensitivity measurement and transforming the search space into a continuous domain for efficient numerical optimization. Applied to Whisper and Conformer models across multiple speech datasets, GenPTQ achieves up to 89.1% model size reduction (2.5-bit average precision) with only a 0.8% increase in WER, and completes optimization in just 15 seconds. These results demonstrate its effectiveness for low-resource ASR deployment.

Downloads

Next from EMNLP 2025

FIER: Fine-Grained and Efficient KV Cache Retrieval for Long-context LLM Inference

Similar lecture

From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge