China

LLMs are highly sensitive to prompt phrasing, yet standard benchmarks typically report performance using a single prompt, raising concerns about the reliability of such evaluations. In this work, we argue for a stochastic method of moments evaluation over the space of meaning-preserving prompt perturbations. We introduce a formal definition of *reliable evaluation* that accounts for prompt sensitivity, and suggest ReliableEval - a method for estimating the number of prompt resamplings needed to obtain meaningful results. Using our framework, we stochastically evaluate five frontier LLMs and find that even top-performing models like GPT-4o and Claude-3.7-Sonnet exhibit substantial prompt sensitivity. Our approach is model-, task-, and metric-agnostic, offering a recipe for meaningful and robust LLM evaluation.

EMNLP 2025

ReliableEval: A Recipe for Stochastic LLM Evaluation via Method of Moments

stochastic

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.

poster

In recent years, multi-agent frameworks powered by large language models (LLMs) have advanced rapidly. Despite this progress, there is still a notable absence of benchmark datasets specifically tailored to evaluate their performance. To bridge this gap, we introduce Auto-SLURP, a benchmark dataset aimed at evaluating LLM-based multi-agent frameworks in the context of smart personal assistants. Auto-SLURP extends the original SLURP dataset—initially developed for natural language understanding tasks—by relabeling the data and integrating simulated servers and external services. This enhancement enables a comprehensive end-to-end evaluation pipeline, covering language understanding, task execution, and response generation. Our experiments demonstrate that Auto-SLURP presents a significant challenge for current state-of-the-art frameworks, highlighting that truly reliable and intelligent multi-agent personal assistants remain a work in progress.

Auto-SLURP: A Benchmark Dataset for Evaluating Multi-Agent Frameworks in Smart Personal Assistant

In the context of fact-checking, claims are often repeated across various platforms and in different languages, which can benefit from a process that reduces this redundancy. While retrieving previously fact-checked claims has been investigated as a solution, the growing number of unverified claims and expanding size of fact-checked databases calls for alternative, more efficient solutions. A promising solution is to group claims that discuss the same underlying facts into clusters to improve claim retrieval and validation. However, research on claim clustering is hindered by the lack of suitable datasets. To bridge this gap, we introduce MultiClaimNet, a collection of three multilingual claim cluster datasets containing claims in 86 languages across diverse topics. Claim clusters are formed automatically from claim-matching pairs with limited manual intervention. We leverage two existing claim-matching datasets to form the smaller datasets within MultiClaimNet. To build the larger dataset, we propose and validate an approach involving retrieval of approximate nearest neighbors to form candidate claim pairs and an automated annotation of claim similarity using large language models. This larger dataset contains 85.3K fact-checked claims written in 78 languages. We further conduct extensive experiments using various clustering techniques and sentence embedding models to establish baseline performance. Our datasets and findings provide a strong foundation for scalable claim clustering, contributing to efficient fact-checking pipelines.

MultiClaimNet: A Massively Multilingual Dataset of Fact-Checked Claim Clusters

Large language models excel at many tasks but often incur high inference costs during deployment. To mitigate hallucination, many systems use a knowledge graph to enhance retrieval-augmented generation (KG-RAG). However, the large amount of retrieved knowledge contexts increase these inference costs further. A promising solution to balance performance and cost is LLM routing, which directs simple queries to smaller LLMs and complex ones to larger LLMs. However, no dedicated routing methods currently exist for RAG, and existing training-based routers face challenges scaling to this domain due to the need for extensive training data. We observe that the score distributions produced by the retrieval scorer strongly correlate with query difficulty. Based on this, we propose a novel, training-free routing framework—the first tailored to KG-RAG—that effectively balances performance and cost in a plug-and-play manner. Experiments show our method reduces calls to larger LLMs by up to 50% without sacrificing response quality, demonstrating its potential for efficient and scalable LLM deployment.

SkewRoute: Training-Free LLM Routing for Knowledge Graph Retrieval-Augmented Generation via Score Skewness of Retrieved Context

LLMs are used predominantly in synchronous communication, where a human user and a model communicate in alternating turns. In contrast, many real-world settings are inherently asynchronous. For example, in group chats, online team meetings, or social games, there is no inherent notion of turns; therefore, the decision of when to speak forms a crucial part of the participant's decision making. In this work, we develop an adaptive asynchronous LLM-agent which, in addition to determining what to say, also decides when to say it. To evaluate our agent, we collect a unique dataset of online Mafia games, including both human participants, as well as our asynchronous agent. Overall, our agent performs on par with human players, both in game performance, as well as in its ability to blend in with the other human players. Our analysis shows that the agent's behavior in deciding when to speak closely mirrors human patterns, although differences emerge in message content. We release all our data and code to support and encourage further research for more realistic asynchronous communication between LLM agents. This work paves the way for integration of LLMs into realistic human group settings, from assistance in team discussions to educational and professional environments where complex social dynamics must be navigated.

Time to Talk: LLM Agents for Asynchronous Group Communication in Mafia Games

As texts generated by Large Language Models (LLMs) are ever more common and often indistinguishable from human-written content, research on automatic text detection has attracted growing attention. Many recent detectors report near-perfect accuracy, often boasting AUROC scores above 99\%. However, these claims typically assume fixed generation settings, leaving open the question of how robust such systems are to changes in decoding strategies. In this work, we systematically examine how sampling-based decoding impacts detectability, with a focus on how subtle variations in a model’s (sub)word-level distribution affect detection performance. We find that even minor adjustments to decoding parameters - such as temperature, top-p, or nucleus sampling - can severely impair detector accuracy, with AUROC dropping from near-perfect levels to 1\% in some settings. Our findings expose critical blind spots in current detection methods and emphasize the need for more comprehensive evaluation protocols. To facilitate future research, we release a large-scale dataset encompassing 37 decoding configurations, along with our code and evaluation framework.

How Sampling Affects the Detectability of Machine-written texts: A Comprehensive Study

AI-based code generation is increasingly prevalent, with GitHub Copilot estimated to generate 46% of GitHub code. Accurately evaluating the alignment of generated code with developer intent remains a critical challenge. Traditional evaluation methods, like unit tests, are often unscalable and costly. Syntactic similarity metrics (e.g., BLEU, ROUGE) fail to capture code functionality, and metrics like CodeBERTScore require reference code that is not always available. Recognizing the gap in reference-free evaluation, with few alternatives like ICE-Score, this paper introduces MATCH, a novel reference-free metric. MATCH employs Contrastive Learning to generate meaningful embeddings for code and natural language task descriptions, enabling similarity scoring that reflects how well generated code implements the task. We demonstrate that MATCH achieves stronger correlations with functional correctness and human preference than existing metrics across several programming languages.

MATCH: Task-Driven Code Evaluation through Contrastive Learning

Existing low-resource Knowledge Graph Question Answering (KGQA) methods rely heavily on Large Language Models (LLMs) for semantic parsing of natural language question to its corresponding logical form (LF) such as SPARQL, S-Expression, etc. However, LLMs becomes bottleneck for practical applications due to: (i) its high computational resource requirements; (2) limited knowledge of LLM about different LFs; (3) unavailability of low-resource annotated data for new KGs and settings. This motivates us to design a KGQA framework that can operate in a zero-shot setting without the need for additional resources. In this paper, we propose (NS-KGQA): a zero-shot neuro-symbolic approach based on neural KG embeddings that have demonstrated their ability to effectively model KG structure without the need of additional data. We extract a link-prediction based symbolic question subgraph. We then propose a Symbolic Resolver that uses Dual KG Embeddings combined with a symbolic approach to resolve the symbolic question subgraph. Our extensive experiments on Complex KGQA benchmarks such as KQA Pro demonstrate the effectiveness of our approach. NS-KGQA outperforms all other LLM-based zero-shot baselines by 26% (avg).

A Zero-Shot Neuro-Symbolic Approach for Complex Knowledge Graph Question Answering

Large language models (LLMs) have large potential for molecular optimization, as they can gather external chemistry tools and enable collaborative interactions to iteratively refine molecular candidates. However, this potential remains underexplored, particularly in the context of structured reasoning, interpretability, and comprehensive tool-grounded molecular optimization. To address this gap, we introduce MT-MOL, a multi-agent framework for molecular optimization that leverages tool-guided reasoning and role-specialized LLM agents. Our system incorporates comprehensive RDKit tools, categorized into five distinct domains: structural descriptors, electronic and topological features, fragment-based functional groups, molecular representations, and miscellaneous chemical properties. Each category is managed by an expert analyst agent, responsible for extracting task-relevant tools and enabling interpretable, chemically grounded feedback. MT-MOL produces molecules with tool-aligned and stepwise reasoning through the interaction between the analyst agents, a molecule-generating scientist, a reasoning-output verifier, and a reviewer agent. As a result, we show that our framework shows the state-of-the-art performance of the PMO-1K benchmark on 17 out of 23 tasks.

MT-Mol: Multi Agent System with Tool-based Reasoning for Molecular Optimization

Hallucination detection remains a fundamental challenge for the safe and reliable deployment of large language models (LLMs), especially in applications requiring factual accuracy. Existing hallucination benchmarks often operate at the sequence level and are limited to English, lacking the fine-grained, multilingual supervision needed for comprehensive evaluation. In this work, we introduce PsiloQA, a large-scale, multilingual dataset annotated with span-level hallucinations across 14 languages. PsiloQA is constructed through an automated three-stage pipeline: generating question–answer pairs from Wikipedia using GPT-4o, eliciting potentially hallucinated answers from diverse LLMs in a no-context setting, and automatically annotating hallucinated spans using GPT-4o by comparing against golden answers and retrieved context. We evaluate a wide range of hallucination detection methods-including uncertainty quantification, LLM-based tagging, and fine-tuned encoder models-and show that encoder-based models achieve the strongest performance across languages. Furthermore, PsiloQA demonstrates effective cross-lingual generalization and supports robust knowledge transfer to other benchmarks, all while being significantly more cost-efficient than human-annotated datasets. Our dataset and results advance the development of scalable, fine-grained hallucination detection in multilingual settings.

When Models Lie, We Learn: Multilingual Span-Level Hallucination Detection with PsiloQA

Temporal question-answer (QA) is an established method to assess temporal reasoning in large language models (LLMs). Expected answers are often numeric (e.g., dates or durations), yet the model responses are evaluated like regular text with exact match (EM), unable to distinguish small from large errors. In this investigative work, we frame temporal QA as a numerical estimation task to assess the shortcomings of EM. We introduce TempAnswerQA, a benchmark distilled from Test of Time and TempTabQA, where all questions require a numerical temporal answer, allowing us to evaluate models beyond EM. We used the forecasting metrics symmetric mean absolute percentage error (sMAPE) and mean absolute scaled error (MASE). With sMAPE, we found that error size and EM are decoupled. Models with low EM still had low sMAPE (both ~20%), and some models had high sMAPE despite high EM. Scaling errors by the deviance of the ground truth data with MASE reshuffles model rankings compared to EM, revealing gaps in models' understanding of temporal domain knowledge, especially when trained with synthetic data. Lastly, the models' most frequent error was to deviate by only ± 1 from the ground truth. sMAPE and MASE, unlike EM, adequately weight these errors. Our findings underscore the need for specialised metrics for temporal QA tasks.

Downloads

Next from EMNLP 2025

Auto-SLURP: A Benchmark Dataset for Evaluating Multi-Agent Frameworks in Smart Personal Assistant

Similar lecture

Flaw or Artifact? Rethinking Prompt Sensitivity in Evaluating LLMs