China

Large Language Models (LLMs) require high quality preference datasets to align with human preferences. However, conventional methods for constructing such datasets face significant challenges: reliance on pre-collected instructions often leads to distribution mismatches with target models, while the need for sampling multiple stochastic responses introduces substantial computational overhead. In this work, we explore a paradigm shift by leveraging inherent regulation of LLMs&#39; representation space for efficient and tailored preference dataset construction, named Icon². Specifically, it first extracts layer-wise direction vectors to encode sophisticated human preferences and then uses these vectors to filter self-synthesized instructions based on their inherent consistency. During decoding, bidirectional inherent control is applied to steer token representations, enabling the precise generation of response pairs with clear alignment distinctions. Experimental results demonstrate significant improvements in both alignment and efficiency. Llama3-8B and Qwen2-7B achieve an average win rate improvement of 13.89\% on AlpacaEval 2.0 and 13.45\% on Arena-Hard, while reducing computational costs by up to 48.1\%.

EMNLP 2025

Icon²: Aligning Large Language Models Using Self-Synthetic Preference Data via Inherent Regulation

synthetic preference data

inherent regulation

llm alignment

Large Language Models (LLMs) require high quality preference datasets to align with human preferences. However, conventional methods for constructing such datasets face significant challenges: reliance on pre-collected instructions often leads to distribution mismatches with target models, while the need for sampling multiple stochastic responses introduces substantial computational overhead. In this work, we explore a paradigm shift by leveraging inherent regulation of LLMs' representation space for efficient and tailored preference dataset construction, named Icon². Specifically, it first extracts layer-wise direction vectors to encode sophisticated human preferences and then uses these vectors to filter self-synthesized instructions based on their inherent consistency. During decoding, bidirectional inherent control is applied to steer token representations, enabling the precise generation of response pairs with clear alignment distinctions. Experimental results demonstrate significant improvements in both alignment and efficiency. Llama3-8B and Qwen2-7B achieve an average win rate improvement of 13.89\% on AlpacaEval 2.0 and 13.45\% on Arena-Hard, while reducing computational costs by up to 48.1\%.

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.

With the different roles that AI is expected to play in human life, imbuing large language models (LLMs) with different personalities has attracted increasing research interest. While the ``personification'' enhances human experiences of interactivity and adaptability of LLMs, it gives rise to critical concerns about content safety, particularly regarding bias, sentiment, and toxicity of LLM generation. This study explores how assigning different personality traits to LLMs affects the toxicity and biases of their outputs. Leveraging the widely accepted HEXACO personality framework developed in social psychology, we design experimentally sound prompts to test three LLMs' performance on three toxic and bias benchmarks. The findings demonstrate the sensitivity of all three models to HEXACO personality traits and, more importantly, a consistent variation in the biases, negative sentiment, and toxicity of their output. In particular, adjusting the levels of several personality traits can effectively reduce bias and toxicity in model performance, similar to humans' correlations between personality traits and toxic behaviors. The findings highlight the additional need to examine content safety besides the efficiency of training or fine-tuning methods for LLM personification, they also suggest a potential for the adjustment of personalities to be a simple and low-cost method to conduct controlled text generation.

Exploring the Impact of Personality Traits on LLM Toxicity and Bias

Deploying large language models (LLMs) for long-context inference remains challenging due to their substantial memory and computational demands. While techniques such as Key-Value (KV) cache compression are designed to reduce memory usage, they often neglect the structured sparsity inherent in the relationship between hidden states and their corresponding KV cache. In this work, we explore the role of uncertainty as a potential indicator of sparsity within LLMs. We propose UNComp, an uncertainty-aware framework that leverages truncated matrix entropy to identify areas of low information content, thereby revealing sparsity patterns that can be used for adaptive compression. Unlike traditional methods that apply uniform compression, UNComp dynamically adjusts its approach to compression, guided by uncertainty measures that reflect the importance of various model components. Our analysis shows that sparsity patterns, when derived from uncertainty estimates, can be exploited to reveal special long-range dependencies, such as retrieval heads and retrieval layers. This perspective not only enhances our understanding of how compression can be optimized but also provides new insights into the inherent sparsity of LLMs during long-context inference. By focusing on uncertainty to analyze the sparsity pattern in detail, UNComp reduces the KV cache size to 4.74% of the original, achieves a 6% prefill speedup, and improves throughput by 6.4× — not only delivering strong lossless compression performance, but also validating the effectiveness of the underlying theoretical tool. Our codes are submitted with the paper.

UNComp: Can Matrix Entropy Uncover Sparsity? — A Compressor Design from an Uncertainty-Aware Perspective

Large language models (LLMs) are commonly trained on multi-domain datasets, where domain sampling strategies significantly impact model performance due to varying domain importance across downstream tasks. Existing approaches for optimizing domain-level sampling strategies struggle with maintaining intra-domain consistency and accurately measuring domain impact. In this paper, we present Domain Impact-aware Data Sampling (DIDS). To ensure intra-domain consistency, a gradient clustering algorithm is proposed to group training data based on their learning effects, where a proxy language model and dimensionality reduction are employed to reduce computational overhead. To accurately measure domain impact, we develop a Fisher Information Matrix (FIM) guided metric that quantifies how domain-specific parameter updates affect the model's output distributions on downstream tasks, with theoretical guarantees. Furthermore, to determine optimal sampling ratios, DIDS combines both the FIM-guided domain impact assessment and loss learning trajectories that indicate domain-specific potential, while accounting for diminishing marginal returns. Extensive experiments demonstrate that DIDS achieves 3.4% higher average performance while maintaining comparable training efficiency. The code is available at https://anonymous.4open.science/r/DIDS.

DIDS: Domain Impact-aware Data Sampling for Large Language Model Training

Federated finetuning of Large Language Models (LLMs) using Low-Rank Adaptation (LoRA) offers computational efficiency and preserves data privacy. However, applying LoRA in federated settings faces significant challenges: standard approaches struggle with data heterogeneity, and existing personalization techniques fail to precisely adapt shared global knowledge to individual client needs. To address these issues, we propose pFedGPT, a framework that leverages Hierarchical Bayesian Optimization (HBO) for fine-grained, personalized LoRA aggregation. pFedGPT intelligently partitions LoRA parameters based on model structure and client information, then employs HBO to hierarchically search for optimal, module-specific weights. This enables a nuanced integration of the downloaded global LoRA state with each client's local model, precisely capturing client-specific requirements. To manage the optimization cost inherent in HBO, pFedGPT incorporates efficient multi-fidelity evaluations and a curriculum learning strategy. Extensive experiments demonstrate that pFedGPT achieves state-of-the-art (SOTA) performance on personalized FL benchmarks, showcasing robustness and scalability while introducing only minimal (approx. 4%) additional optimization overhead. Our results also underscore the limitations of traditional FL methods for LoRA-based LLM personalization, highlighting the need for tailored approaches like pFedGPT.

pFedGPT: Hierarchically Optimizing LoRA Aggregation Weights for Personalized Federated GPT Models

Quantization is widely adopted to accelerate inference and reduce memory consumption in large language models (LLMs). While activation-weight joint quantization enables efficient low-precision decoding, it suffers substantial performance degradation on multi-step reasoning tasks. We propose QSPEC, a novel quantization paradigm that decouples efficiency from quality by integrating two complementary schemes via speculative decoding: low-precision joint quantization for fast drafting and high-precision weight-only quantization for accurate verification. QSPEC reuses both weights and KV cache across stages, enabling near-zero-cost switching without retraining or auxiliary models. Compared to high-precision baselines, QSPEC achieves up to 1.64x speedup without quality degradation, and outperforms state-of-the-art speculative decoding methods by up to 1.55x in batched settings. Furthermore, QSPEC supports plug-and-play deployment and generalizes well across model scales, quantization methods, and workloads. These properties make QSPEC a practical and scalable solution for high-fidelity quantized LLM serving under memory-constrained scenarios.

QSpec: Speculative Decoding with Complementary Quantization Schemes

Prominent Large Language Model (LLM) services from providers like OpenAI and Google excel at general tasks but often underperform on domain-specific applications. Current customization services for these LLMs typically require users to upload data for fine-tuning, posing significant privacy risks. While differentially private (DP) data synthesis presents a potential alternative, its application commonly results in low effectiveness due to the introduction of excessive noise on data for DP. To overcome this, we introduce Llamdex, a novel framework that facilitates LLM customization as a service, where the client uploads pre-trained domain-specific models rather than data. This client-uploaded model, optionally protected by DP with much lower noise, is inserted into the base LLM via connection modules. Significantly, these connecting modules are trained without requiring sensitive domain data, enabling clients to customize LLM services while preserving data privacy. Experiments demonstrate that Llamdex improves domain-specific accuracy by up to 26% over state-of-the-art private data synthesis methods under identical privacy constraints and, by obviating the need for users to provide domain context within queries, maintains inference efficiency comparable to the original LLM service.

Model-based Large Language Model Customization as Service

Reinforcement learning from human feedback (RLHF) offers a promising approach to aligning large language models (LLMs) with human preferences. Typically, a reward model is trained or supplied to act as a proxy for humans in evaluating generated responses during the reinforcement training phase. However, current reward models operate as sequence-to-one models, allocating a single, sparse, and delayed reward to an entire output sequence. This approach may overlook the significant contributions of individual tokens toward the desired outcome. To this end, we propose a more fine-grained, token-level guidance approach for RL training. Specifically, we introduce RED, a novel REward reDistribition method that evaluates and assigns specific credit to each token using an off-the-shelf reward model. Utilizing these fine-grained rewards enhances the model's understanding of language nuances, leading to more precise performance improvements. Notably, our method does not require modifying the reward model or introducing additional training steps, thereby incurring minimal computational costs. Experimental results across diverse datasets and tasks demonstrate the superiority of our approach.

RED: Unleashing Token-Level Rewards from Holistic Feedback via Reward Redistribution

Large Language Models (LLMs) have demonstrated unprecedented capabilities across various natural language processing tasks, yet they remain vulnerable to jailbreaking attacks designed to deliberately induce harmful content generation. Despite numerous defensive efforts, we lack understanding of whether models' safety behaviors change when assuming different roles. In this paper, we reveal a critical safety inconsistency: LLMs can more effectively identify harmful requests as discriminators than defend against them as generators. To address this gap, we propose SDGO (Self-Discrimination-Guided Optimization), a reinforcement learning framework that leverages the model's own discrimination capabilities as a reward signal to enhance generation safety through iterative self-improvement without additional annotated data or external models. Extensive experiments across various LLMs and jailbreaking attacks demonstrate that SDGO significantly improves model safety compared to both prompt-based and training-based baselines, effectively narrowing the discrimination-generation safety gap while maintaining utility on general benchmarks. Our approach enables mutual benefit between LLMs' discrimination and generation capabilities, resulting in robust performance against out-of-distribution (OOD) jailbreaking attacks. Additionally, we find that SDGO can be further enhanced by fine-tuning with a small amount of harmful-labeled discrimination samples, indicating that SDGO effectively transforms discrimination into part of the model's generation, achieving tight coupling between these two aspects.

SDGO: Self-Discrimination-Guided Optimization for Consistent Safety in Large Language Models

Instruction tuning is vital for aligning large language models (LLMs) with human intent, but current methods typically rely on costly human-annotated seed data or powerful external teacher models. While instruction back-translation techniques reduce this dependency, they remain fundamentally tethered to an initial seed set, which limits full automation, introduces biases, and can lead to inefficient use of unlabeled corpora. In this paper, we propose Cycle-Instruct, a novel framework that achieves fully seed-free instruction tuning. Inspired by cycle consistency, Cycle-Instruct employs a dual self-training loop where two models—an answer generator and a question generator—are bootstrapped solely from raw, unlabeled text. These models mutually supervise each other by reconstructing original text segments from their counterpart's generated pseudo-labels, effectively learning from the intrinsic structure of the data without any human-provided seeds. We demonstrate Cycle-Instruct's efficacy across four diverse data tracks, including general instruction-following, domain-specific tasks, dialogue logs, and plain text. Our extensive experiments show that Cycle-Instruct not only outperforms seed-driven back-translation baselines but also achieves performance comparable to strongly supervised methods.

CYCLE-INSTRUCT: Fully Seed-Free Instruction Tuning via Dual Self-Training and Cycle Consistency

Large Reasoning Models (LRMs) embed private or copyrighted information not only in their final answers but also throughout multi-step chain-of-thought (CoT) traces, making reliable unlearning far more demanding than in standard LLMs. We introduce Reasoning-TOFU (R-TOFU), the first benchmark tailored to this setting. R-TOFU augments existing unlearning tasks with realistic CoT annotations and provides step-wise metrics that expose residual knowledge invisible to answer-level checks. Using R-TOFU, we carry out a comprehensive comparison of gradient-based and preference-optimization baselines and show that conventional answer-only objectives leave substantial forget traces in reasoning. We further propose Reasoned IDK, a preference-optimization variant that preserves coherent yet inconclusive reasoning, achieving a stronger balance between forgetting efficacy and model utility than earlier refusal styles. Finally, we identify a failure mode: decoding variants such as ZeroThink and LessThink can still reveal forgotten content despite seemingly successful unlearning, emphasizing the need to evaluate models under diverse decoding settings. Together, the benchmark, analysis, and new baseline establish a systematic foundation for studying and improving unlearning in LRMs while preserving their reasoning capabilities.

Downloads

Next from EMNLP 2025

Exploring the Impact of Personality Traits on LLM Toxicity and Bias