China

Despite the significant advancements of self-play fine-tuning (SPIN), which can transform a weak large language model (LLM) into a strong one through competitive interactions between models of varying capabilities, it still faces challenges in the Text-to-SQL task. SPIN does not generate new information, and the large number of correct SQL queries produced by the opponent model during self-play reduces the main model&#39;s ability to generate accurate SQL queries. To address this challenge, we propose a new self-play fine-tuning method tailored for the Text-to-SQL task, called SPFT-SQL. Prior to self-play, we introduce a verification-based iterative fine-tuning approach, which synthesizes high-quality fine-tuning data iteratively based on the database schema and validation feedback to enhance model performance, while building a model base with varying capabilities. During the self-play fine-tuning phase, we propose an error-driven loss method that incentivizes incorrect outputs from the opponent model, enabling the main model to distinguish between correct SQL and erroneous SQL generated by the opponent model, thereby improving its ability to generate correct SQL. Extensive experiments and in-depth analyses on six open-source LLMs and five widely used benchmarks demonstrate that our approach outperforms existing state-of-the-art (SOTA) methods.

EMNLP 2025

SPFT-SQL: Enhancing Large Language Model for Text-to-SQL Parsing by Self-Play Fine-Tuning

nlp application

question answering

generation

Despite the significant advancements of self-play fine-tuning (SPIN), which can transform a weak large language model (LLM) into a strong one through competitive interactions between models of varying capabilities, it still faces challenges in the Text-to-SQL task. SPIN does not generate new information, and the large number of correct SQL queries produced by the opponent model during self-play reduces the main model's ability to generate accurate SQL queries. To address this challenge, we propose a new self-play fine-tuning method tailored for the Text-to-SQL task, called SPFT-SQL. Prior to self-play, we introduce a verification-based iterative fine-tuning approach, which synthesizes high-quality fine-tuning data iteratively based on the database schema and validation feedback to enhance model performance, while building a model base with varying capabilities. During the self-play fine-tuning phase, we propose an error-driven loss method that incentivizes incorrect outputs from the opponent model, enabling the main model to distinguish between correct SQL and erroneous SQL generated by the opponent model, thereby improving its ability to generate correct SQL. Extensive experiments and in-depth analyses on six open-source LLMs and five widely used benchmarks demonstrate that our approach outperforms existing state-of-the-art (SOTA) methods.

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.

The wide adoption of Large Language Models (LLMs) has attracted significant attention from jailbreak attacks, where adversarial prompts crafted through optimization or manual design exploit LLMs to generate malicious contents. However, optimization-based attacks have limited efficiency and transferability, while existing manual designs are either easily detectable or demand intricate interactions with LLMs. In this paper, we first point out a novel perspective for jailbreak attacks: LLMs are more responsive to positive prompts. Based on this, we deploy Happy Ending Attack (HEA) to wrap up a malicious request in a scenario template involving a positive prompt formed mainly via a happy \ ending, it thus fools LLMs into jailbreaking either immediately or at a follow-up malicious request. This has made HEA both efficient and effective, as it requires only up to two turns to fully jailbreak LLMs. Extensive experiments show that our HEA can successfully jailbreak on state-of-the-art LLMs, including GPT-4o, Llama3-70b, Gemini-pro, and achieves 88.79\% attack success rate on average. We also provide quantitative explanations for the success of HEA.

Dagger Behind Smile: Fool LLMs with a Happy Ending Story

Large Language Models (LLMs) are increasingly aligned with human preferences through Reinforcement Learning from Human Feedback (RLHF). Among RLHF methods, Group Relative Policy Optimization (GRPO) has gained attention for its simplicity and strong performance, notably eliminating the need for a learned value function. However, GRPO implicitly assumes a balanced domain distribution and uniform semantic alignment across groups—assumptions that rarely hold in real-world datasets. When applied to multi-domain, imbalanced data, GRPO disproportionately optimizes for dominant domains, neglecting underrepresented ones and resulting in poor generalization and fairness. We propose Domain-Informed Self-Consistency Policy Optimization (DISCO), a principled extension to GRPO that addresses inter-group imbalance with two key innovations. Domain-aware reward scaling counteracts frequency bias by reweighting optimization based on domain prevalence. Difficulty-aware reward scaling leverages prompt-level self-consistency to identify and prioritize uncertain prompts that offer greater learning value. Together, these strategies promote more equitable and effective policy learning across domains. Extensive experiments across multiple LLMs and skewed training distributions show that DISCO improves generalization, outperforms existing GRPO variants by 5% on Qwen3 models, and sets new state-of-the-art results on multi-domain alignment benchmarks.

DISCO Balances the Scales: Adaptive Domain- and Difficulty-Aware Reinforcement Learning on Imbalanced Data

Large Language Models (LLMs) have transformed natural language processing, yet their internal mechanisms remain largely opaque. Recently, mechanistic interpretability has attracted significant attention from the research community as a means to understand the inner workings of LLMs. Among various mechanistic interpretability approaches, Sparse Autoencoders (SAEs) have emerged as a promising method due to their ability to disentangle the complex, superimposed features within LLMs into more interpretable components. This paper presents a comprehensive survey of SAEs for interpreting and understanding the internal workings of LLMs. Our major contributions include: (1) exploring the technical framework of SAEs, covering basic architecture, design improvements, and effective training strategies; (2) examining different approaches to explaining SAE features, categorized into input-based and output-based explanation methods; (3) discussing evaluation methods for assessing SAE performance, covering both structural and functional metrics; and (4) investigating real-world applications of SAEs in understanding and manipulating LLM behaviors.

A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models

With the development of Large Language Models (LLMs), numerous efforts have revealed their vulnerabilities to jailbreak attacks. Although these studies have driven the progress in LLMs' safety alignment, it remains unclear whether LLMs have internalized authentic knowledge to deal with real-world crimes, or are merely forced to simulate toxic language patterns. This ambiguity raises concerns that jailbreak success is often attributable to a hallucination loop between jailbroken LLM and judger LLM. By decoupling the use of jailbreak techniques, we construct knowledge-intensive Q&A to investigate the misuse threats of LLMs in terms of dangerous knowledge possession, harmful task planning utility, and harmfulness judgment robustness. Experiments reveal a mismatch between jailbreak success rates and harmful knowledge possession in LLMs, and existing LLM-as-a-judge frameworks tend to anchor harmfulness judgments on toxic language patterns. Our study reveals a gap between existing LLM safety assessments and real-world threat potential.

Confusion is the Final Barrier: Rethinking Jailbreak Evaluation and Investigating the Real Misuse Threat of LLMs

This paper proposes Complete Textual Concept Bottleneck Model (CT-CBM), a novel TCBM generator building concept labels in a fully unsupervised manner using a small language model, eliminating both the need for predefined human labeled concepts and LLM annotations. CT-CBM iteratively targets and adds important and identifiable concepts in the bottleneck layer to create a complete concept basis. CT-CBM achieves striking results against competitors in terms of concept basis completeness and concept detection accuracy, offering a promising solution to reliably enhance interpretability of NLP classifiers.

Towards Achieving Concept Completeness for Textual Concept Bottleneck Models

Large language models (LLMs) are trained using massive datasets. However, these datasets often contain undesirable content, e.g., harmful texts, personal information, and copyrighted material. To address this, \emph{machine unlearning} aims to remove information from trained models. Recent work has shown that soft token attacks (\sta) can successfully extract unlearned information from LLMs. In this work, we show that \sta{s} can be an inadequate tool for auditing unlearning. Using common unlearning benchmarks (\textit{Who Is Harry Potter?} and \textit{TOFU}), we demonstrate that, in a \emph{strong auditor} setting, such attacks can elicit any information from the LLM, regardless of (1) the deployed unlearning algorithm, and (2) whether the queried content was originally present in the training corpus. Also, we show that \sta with just a few soft tokens (1-10) can elicit random strings over 400-characters long. Thus showing that \sta{s} must be used carefully to effectively audit unlearning.

Soft Token Attacks Cannot Reliably Audit Unlearning in Large Language Models

Oversensitivity—where language models defensively reject benign prompts—not only disrupts user interactions but also obscures the boundaries between harmful and harmless content. Existing benchmarks rely on static datasets that degrade over time as models evolve, leading to data contamination and diminished evaluative power. To address this, we develop a framework that dynamically generates model-specific challenging datasets, capturing emerging defensive patterns and aligning with each model’s unique behavior. Building on this approach, we construct OverBench, a benchmark that aggregates these datasets across diverse LLM families, encompassing 450,000 samples from 26 models. OverBench provides a dynamic and evolving perspective on oversensitivity, allowing for continuous monitoring of defensive triggers as models advance, highlighting vulnerabilities that static datasets overlook.

Dynamic Evaluation for Oversensitivity in LLMs

Currently, large language models (\textbf{L}LMs) based \textbf{O}pen domain \textbf{N}atural language plannin\textbf{G} (LONG) has considerable room for improvement. E.g., nonreusable plans with incomplete intermediate states and missing steps hinder real-world applications. To remedy these flaws, this paper establishes a dataset with a baseline for LONG. The dataset, GOLD, provides the largest dataset for textual procedures with corresponding reusable formal planning domain definitions to date. The baseline, DIGGER, leverages entity-attribute-level action models, which reveal relevant implicit physical properties (aka attributes) of salient entities in actions. DIGGER first extracts action models and builds typed entity lists from textual procedures. Then, it builds goal states for new tasks and instantiates grounded actions using domain prediction. At last, plans are generalized and translated into textual procedures by LLM. Reference-based metrics, LLM-as-Judge, and human evaluation are employed to evaluate LONG comprehensively. Experiments on GOLD validate that DIGGER is stronger and more generalizable than recently proposed approaches and LLMs. I.e., DIGGER is the best in seen domains and applicable to unseen domains without adaptation. Specifically, the best BLEU-1 score increased from 0.385 to 0.408 on seen domains and boosted to 0.310 on unseen domains.

LLM-based Open Domain Planning by Leveraging Entity-Attribute-Level Domain Models

Cross-prompt trait scoring task aims to learn generalizable scoring capabilities from source- prompt data, enabling automatic scoring across multiple dimensions on unseen essays. Existing research on cross-prompt trait essay scoring primarily focuses on improving model generalization by obtaining prompt-invariant representations. In this paper, we approach the research problem from a different perspective on invariance learning and propose a scoring-invariant learning objective. This objective encourages the model to focus on intrinsic information within the essay that reflects its quality during training, thereby learning generic scoring features. To further enhance the model's ability to score across multiple dimensions, we introduce a trait feature extraction network based on routing gates into the scoring architecture and propose a trait consistency scoring objective to encourage the model to balance the diversity of trait-specific features with scoring consistency across traits when learning trait-specific essay features. Extensive experiments demonstrate the effectiveness of our approach, showing advantages in multi-trait scoring performance and achieving significant improvements with low-resource prompts.

Improving Prompt Generalization for Cross-prompt Essay Trait Scoring from the Scoring-invariance Perspective

Large language models (LLMs) have demonstrated remarkable capabilities in natural language processing tasks. However, they remain vulnerable to semantic inconsistency, where minor, semantically equivalent variations in input formatting result in divergent predictions. Our comprehensive evaluation reveals that this brittleness persists even in state-of-the-art models such as GPT-4o, posing a serious challenge to their reliability. Through a mechanistic analysis, we attribute this phenomenon to deep representational failures, whereby semantic-equivalent input changes induce instability in the model’s internal representations. We further examine standard mitigation strategies and uncover their fundamental limitations. In particular, even direct fine-tuning on format variations frequently fails to yield format-invariant semantic representations, highlighting the difficulty of the problem. By explaining the failure of existing methods through our representational diagnosis, we underscore the need for representation-aware strategies to achieve robust and reliable LLM behavior.

Downloads

Next from EMNLP 2025

Dagger Behind Smile: Fool LLMs with a Happy Ending Story