China

As Natural Language Generation (NLG) continues to be widely adopted, properly assessing it has become quite difficult. Lately, using large language models (LLMs) for evaluating these generations has gained traction, as they tend to align more closely with human preferences than conventional n-gram or embedding-based metrics. In our experiments, we show that LLM judges have low intra-rater reliability in their assigned scores across different runs. This variance makes their ratings inconsistent, almost arbitrary in the worst case, making it difficult to measure how good their judgments actually are. We quantify this inconsistency across different NLG tasks and benchmarks and see if judicious use of LLM judges can still be useful following proper guidelines.

EMNLP 2025

Rating Roulette: Self-Inconsistency in LLM-As-A-Judge Frameworks

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.

Workload traces are essential to understand complex computer systems' behavior and manage processing and memory resources. Since real-world traces are hard to obtain, synthetic trace generation is a promising alternative. This paper proposes a first-of-a-kind approach that relies on training a large language model (LLM) to generate synthetic workload traces, specifically microservice call graphs. To capture complex and arbitrary hierarchical structures and implicit constraints in such traces, we show how to fine-tune LLMs to generate recursively, making call graph generation a sequence of easier steps. To further enforce learning constraints in traces and generate uncommon situations, we argue for applying additional instruction tuning steps to align our model with the desired trace features. Our evaluation results show that we can generate diverse realistic traces under various conditions and outperform existing methods in accuracy and validity. We demonstrate that our synthetically generated traces can effectively replace real data to optimize important microservice management tasks. Additionally, our model adapts to downstream trace-related tasks, such as predicting key trace features and infilling missing data.

Large Language Models as Realistic Microservice Trace Generators

Personalization is essential for question answering systems that are user-centric. Despite its importance, personalization in answer generation has been relatively underexplored. This is mainly due to lack of resources for training and evaluating personalized question answering systems. We address this gap by introducing LaMP-QA---a benchmark designed for evaluating personalized long-form answer generation. The benchmark covers questions from three major categories: (1) Arts & Entertainment, (2) Lifestyle & Personal Development, and (3) Society & Culture, encompassing over 45 subcategories in total. To assess the quality and potential impact of the LaMP-QA benchmark for personalized question answering, we conduct comprehensive human and automatic evaluations, to compare multiple evaluation strategies for evaluating generated personalized responses and measure their alignment with human preferences. Furthermore, we benchmark a number of non-personalized and personalized approaches based on open-source and proprietary large language models (LLMs). Our results show that incorporating the personalized context provided leads to performance improvements of up to 39%. The benchmark is publicly released to support future research in this area.

LaMP-QA: A Benchmark for Personalized Long-form Question Answering

Multimodal question answering (QA) often requires identifying which video, audio, or sensor tokens are relevant to the question. Yet modality disagreements are common: off-camera speech, background noise, or motion outside the field of view often mislead fusion models that weight all streams equally. We present RAVEN, a unified QA architecture whose core is QuART, a query-conditioned cross-modal gating module that assigns scalar relevance scores to each token across modalities, enabling the model to amplify informative signals and suppress distractors before fusion. RAVEN is trained through a three-stage pipeline comprising unimodal pretraining, query-aligned fusion, and disagreement-oriented fine-tuning - each stage targeting a distinct challenge in multi-modal reasoning: representation quality, cross-modal relevance, and robustness to modality mismatch. To support training and evaluation, we release AVS-QA, a dataset of 300K synchronized Audio-Video-Sensor streams paired with automatically generated question-answer pairs. Experimental results on seven multi-modal QA benchmarks - including egocentric and exocentric tasks - show that RAVEN achieves up to 14.5% and 8.0% gains in accuracy compared to state-of-the-art multi-modal large language models, respectively. Incorporating sensor data provides an additional 16.4% boost, and the model remains robust under modality corruption, outperforming SOTA baselines by 50.23%. Our code and dataset are available at https://anonymous.4open.science/r/RAVEN/.

RAVEN: Query-Guided Representation Alignment for Question Answering over Audio, Video, Embedded Sensors, and Natural Language

In this work, we present sequence-driven structural causal models (SD-SCMs), a framework for specifying causal models with user-defined structure and language-model-defined mechanisms. We characterize how an SD-SCM enables sampling from observational, interventional, and counterfactual distributions according to the desired causal structure. We then leverage this procedure to propose a new type of benchmark for causal inference methods, generating individual-level counterfactual data to test treatment effect estimation. We create an example benchmark consisting of thousands of datasets, and test a suite of popular estimation methods for average, conditional average, and individual treatment effect estimation. We find under this benchmark that (1) causal methods outperform non-causal methods and that (2) even state-of-the-art methods struggle with individualized effect estimation, suggesting this benchmark captures some inherent difficulties in causal estimation. Apart from generating data, this same technique can underpin the auditing of language models for (un)desirable causal effects, such as misinformation or discrimination. We believe SD-SCMs can serve as a useful tool in any application that would benefit from sequential data with controllable causal structure.

Language Models as Causal Effect Generators

Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, but controlling their behavior reliably remains challenging, especially in open-ended generation settings. This paper introduces a novel supervised steering approach that operates in sparse, interpretable representation spaces. We employ sparse autoencoders (SAEs) to obtain sparse latent representations that aim to disentangle semantic attributes from model activations. Then we train linear classifiers to identify a small subspace of task-relevant dimensions in latent representations. Finally, we learn supervised steering vectors constrained to this subspace, optimized to align with target behaviors. Experiments across sentiment, truthfulness, and politics polarity steering tasks with multiple LLMs demonstrate that our supervised steering vectors achieve higher success rates with minimal degradation in generation quality compared to existing methods. Further analysis reveals that a notably small subspace is sufficient for effective steering, enabling more targeted and interpretable interventions.

SAE-SSV: Supervised Steering in Sparse Representation Spaces for Reliable Control of Language Models

Large Language Models (LLMs) have achieved remarkable success in question answering (QA) tasks, but often encode biases that hinder fairness and reliability. Existing bias mitigation techniques are constrained to predefined bias categories, limiting their ability to address novel or emergent biases. To address this, we introduce Open-DeBias, a comprehensive framework for open-set bias mitigation in QA. We propose a large-scale open-set QA dataset with 473, 602 instances spanning 31 bias categories and 9, 594 subgroups to enable the detection and mitigation of both known and unseen biases. Furthermore, we propose a novel debiasing framework using adapters, a parameter efficient fine-tuning approach to mitigate existing biases while generalizing to the newer ones. Compared to the state-of-the-art BMBI method, Open-DeBias improves QA accuracy on the BBQ dataset by 48.3% and 5.2% for ambiguous and disambiguous cases, respectively. Additionally, with adapters fine-tuned on only 4.27% of total training data, our method obtains 84% accuracy on Korean BBQ, demonstrating strong language-agnostic generalization. These results demonstrate a new direction for scal027 able, adaptive bias mitigation in LLM-based QA without compromising task performance.

Open-DeBias: Toward Mitigating Open-Set Bias in Language Models

This system paper presents the DeMeVa team's approaches to the third edition of the Learning with Disagreements shared task (LeWiDi 2025; Leonardelli et al., 2025). We explore two directions: in-context learning (ICL) with large language models, where we compare example sampling strategies; and label distribution learning (LDL) methods with RoBERTa (Liu et al., 2019b), where we evaluate several fine-tuning methods. Our contributions are twofold: (1) we show that ICL can effectively predict annotator-specific annotations (perspectivist annotations), and that aggregating these predictions into soft labels yields competitive performance; and (2) we argue that LDL methods are promising for soft label predictions and merit further exploration by the perspectivist community.

DeMeVa at LeWiDi-2025: Modeling Perspectives with In-Context Learning and Label Distribution Learning

Test-time scaling is a family of techniques to improve LLM outputs at inference time by performing extra computation.
To the best of our knowledge, test-time scaling has been limited to domains with verifiably correct answers, like mathematics and coding.
We transfer test-time scaling to the LeWiDi-2025 tasks to evaluate annotation disagreements.
We experiment with three test-time scaling methods: two benchmark algorithms (Model Averaging and Majority Voting), and a Best-of-N sampling method.
The two benchmark methods improve LLM performance consistently on the LeWiDi tasks, but the Best-of-N method does not.
Our experiments suggest that the Best-of-N method does not currently transfer from mathematics to 
LeWiDi tasks, and we analyze potential reasons for this gap.

BoN Appetit Team at LeWiDi-2025: Best-of-N Test-time Scaling Can Not Stomach Annotation Disagreements (Yet)

The Learning With Disagreements (LeWiDi) 2025 shared task aims to model annotator disagreement through soft label distribution prediction and perspectivist evaluation, which focuses on modeling individual annotators. We adapt DisCo (Distribution from Context), a neural architecture that jointly models item-level
and annotator-level label distributions, and present detailed analysis and improvements. In this paper, we extend DisCo by introducing annotator metadata
embeddings, enhancing input representations, and multi-objective training
losses to capture disagreement patterns better. Through extensive experiments,
we demonstrate substantial improvements in both soft and perspectivist evaluation
metrics across three datasets. We also conduct in-depth calibration and error
analyses that reveal when and why disagreement-aware modeling improves.
Our findings show that disagreement can be better captured by conditioning on
annotator demographics and by optimizing directly for distributional metrics, yielding consistent improvements across datasets.

LPI-RIT at LeWiDi-2025: Improving Distributional Predictions via Metadata and Loss Reweighting with DisCo

The VariErrNLI task requires detecting the degree to which each Natural Language Inference (NLI) label is acceptable to a group of annotators. 
This paper presents an approach to VariErrNLI which incorporates measures of uncertainty, namely Semantic Entropy (SE), to model human label variation. 
Our method is based on the assumption that if two labels are plausible alternatives, then their explanations must be non-contradictory.
We measure SE over Large Language Model (LLM)-generated explanations for a given NLI label, which represents the model uncertainty over the semantic space of possible explanations for that label. 
The system employs SE scores combined with an encoding of the inputs and generated explanations, and reaches a 0.31 Manhattan distance score on the test set, ranking joint first in the soft evaluation of VariErrNLI.

Premium content

Downloads

Next from EMNLP 2025

Large Language Models as Realistic Microservice Trace Generators

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES