China

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.

EMNLP 2025

R-TOFU: Unlearning in Large Reasoning Models

large reasoning models

machine unlearning

large language models

## 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

Large Language Models (LLMs) are often quantized to lower precision to reduce the memory cost and latency in inference. However, quantization often degrades model performance, thus fine-tuning is required for various downstream tasks. Traditional fine-tuning methods such as stochastic gradient descent and Adam optimization require backpropagation, which is error-prone in the low-precision settings. To overcome these limitations, we propose the Quantized Zeroth-Order (QuZO) framework, specifically designed for fine-tuning LLMs through low-precision (e.g., 4- or 8-bit) forward passes. Our method avoids the low-precision straight-through estimator, which requires backward computation, and instead utilizes optimized stochastic rounding to mitigate increased bias. QuZO simplifies the training process, while achieving results comparable to first-order methods in rm FP8 and superior accuracy in rm INT8 and rm INT4 training. Experiments demonstrate that QuZO achieves competitive performance on classification, multi-choice, and generation tasks under low-bit training, including zero-shot reasoning tasks. Notably, QuZO incurs minimal overhead and reduces memory consumption by 2.94 times–5.47 times compared to quantized first-order methods during LLaMA-7B fine-tuning.

QuZO: Quantized Zeroth-Order Fine-Tuning for Large Language Models

Grouped-Query Attention (GQA) is a widely adopted strategy for reducing the computational cost of attention layers in large language models (LLMs). However, current GQA configurations are often suboptimal because they overlook how context length influences inference cost. Since inference cost grows with context length, the most cost-efficient GQA configuration should also vary accordingly. In this work, we analyze the relationship among context length, model size, GQA configuration, and model loss, and introduce two innovations: (1) we decouple the total head size from the hidden size, enabling more flexible control over attention FLOPs; and (2) we jointly optimize the model size and the GQA configuration to arrive at a better allocation of inference resources between attention layers and other components. Our analysis reveals that commonly used GQA configurations are highly suboptimal for long-context scenarios. More importantly, we propose a recipe for deriving cost-optimal GQA configurations. Our results show that for long-context scenarios, one should use fewer attention heads while scaling up model size. Configurations selected by our recipe can reduce both memory usage and FLOPs by more than 50% compared to Llama-3's GQA, with no degradation in model capabilities. Our findings offer valuable insights for designing efficient long-context LLMs.

Cost-Optimal Grouped-Query Attention for Long-Context Modeling

Machine unlearning aims to selectively remove targeted knowledge from Large Language Models (LLMs), ensuring they forget specified content while retaining essential information. Existing unlearning metrics assess whether a model correctly answers retain queries and rejects forget queries, but they fail to capture real-world scenarios where forget queries rarely appear in isolation. In fact, forget and retain queries often coexist within the same prompt, making mixed-query evaluation crucial. We introduce SEPS, an evaluation framework that explicitly measures a model’s ability to both forget and retain information within a single prompt. Through extensive experiments across three benchmarks, we identify two key failure modes in existing unlearning methods: (1) untargeted unlearning indiscriminately erases both forget and retain content once a forget query appears, and (2) targeted unlearning overfits to single-query scenarios, leading to catastrophic failures when handling multiple queries. To address these issues, we propose Mixed Prompt (MP) unlearning, a strategy that integrates both forget and retain queries into a unified training objective. Our approach significantly improves unlearning effectiveness, demonstrating robustness even in complex settings with up to eight mixed forget and retain queries in a single prompt.

SEPS: A Separability Measure for Robust Unlearning in LLMs

Sequence labeling remains a significant challenge in low-resource, domain-specific scenarios, particularly for character-dense languages. Existing methods primarily focus on enhancing model comprehension and improving data diversity to boost performance. However, these approaches still struggle with inadequate model applicability and semantic distribution biases in domain-specific contexts. To overcome these limitations, we propose a novel framework that combines an LLM-based knowledge enhancement workflow with a span-based Knowledge Fusion for Rich and Efficient Extraction (KnowFREE) model. Our workflow employs explanation prompts to generate precise contextual interpretations of target entities, effectively mitigating semantic biases and enriching the model's contextual understanding. The KnowFREE model further integrates extension label features, enabling efficient nested entity extraction without relying on external knowledge during inference. Experiments on multiple domain-specific sequence labeling datasets demonstrate that our approach achieves state-of-the-art performance, effectively addressing the challenges posed by low-resource settings.

Improving Low-Resource Sequence Labeling with Knowledge Fusion and Contextual Label Explanations

Despite the impressive performance of large language models (LLMs) in general domains, they often underperform in specialized domains. Existing approaches typically rely on data synthesis methods and yield promising results by using unlabeled data to capture domain-specific features. However, these methods either incur high computational costs or suffer from performance limitations, while also demonstrating insufficient generalization across different tasks. To address these challenges, we propose textbfAQuilt, a framework for constructing instruction-tuning data for any specialized domains from corresponding unlabeled data, including textbfAnswer, textbfQuestion, textbfUnlabeled data, textbfInspection, textbfLogic, and textbfTask type. By incorporating logic and inspection, we encourage reasoning processes and self-inspection to enhance model performance. Moreover, customizable task instructions enable high-quality data generation for any task. As a result, we construct a dataset of 703K examples to train a powerful data synthesis model. Experiments show that AQuilt is comparable to DeepSeek-V3 while utilizing just 17% of the production cost. Further analysis demonstrates that our generated data exhibits higher relevance to downstream tasks. Codes are available at https://anonymous.4open.science/r/AQuilt-4C3B.

AQuilt: Weaving Logic and Self-Inspection into Low-Cost, High-Relevance Data Synthesis for Specialist LLMs

In-Context derived Vector (ICV) methods extract task-relevant representations from large language models (LLMs) and reinject them during inference, achieving comparable performance to few-shot In-Context Learning (ICL) without repeated demonstration processing. However, existing ICV methods remain sensitive to ICL-specific factors, often use coarse or semantically fragmented representations as the source of the vector, and rely on heuristic-based injection positions, limiting their applicability. To address these issues, we propose Dynamic Vector (DyVec), which incorporates an Exhaustive Query Rotation (EQR) strategy to extract robust semantically aggregated latent representations by mitigating variance introduced by ICL. It then applies Dynamic Latent Segmentation and Injection to adaptively partition representations based on task complexity and leverages REINFORCE-based optimization to learn optimal injection positions for each segment. Experiments results show that DyVec outperforms few-shot ICL, LoRA, and prior ICV baselines. Further analysis highlights the effectiveness of dynamically segmenting and injecting semantically aggregated latent representations. DyVec provides a lightweight and data-efficient solution for inference-time task adaptation.

Beyond Demonstrations: Dynamic Vector Construction from Latent Representations

In this work, we propose FoRA-UA, a novel method that, using only 1--5% of the standard LoRA's parameters, achieves state-of-the-art performance across a wide range of tasks. Specifically, we explore scenarios with extremely limited parameter budgets and derive two key insights: (1) fix-sized sparse frequency representations approximate small matrices more accurately; and (2) with a fixed number of trainable parameters, introducing a smaller intermediate representation to approximate larger matrices results in lower construction error. These findings form the foundation of our FoRA-UA method. By inserting a small intermediate parameter set, we achieve greater model compression without sacrificing performance. We evaluate FoRA-UA across diverse tasks, including natural language understanding (NLU), natural language generation (NLG), instruction tuning, and image classification, demonstrating strong generalisation and robustness under extreme compression.

Tiny Budgets, Big Gains: Parameter Placement Strategy in Parameter Super-Efficient Fine-Tuning

While large language models (LLMs) demonstrate remarkable capabilities across a wide range of tasks, they remain vulnerable to generating outputs that are potentially harmful. Red teaming, which involves crafting adversarial inputs to expose vulnerabilities, is a widely adopted approach for evaluating the robustness of these models. Prior studies have indicated that LLMs are susceptible to vulnerabilities exposed through multi-turn interactions as opposed to single-turn scenarios. Nevertheless, existing methods for multi-turn attacks mainly utilize a predefined dialogue pattern, limiting their effectiveness in realistic situations. Effective attacks require adaptive dialogue strategies that respond dynamically to the initial user prompt and the evolving context of the conversation. To address these limitations, we propose DAMON, a novel multi-turn jailbreak attack method. DAMON leverages Monte Carlo Tree Search (MCTS) to systematically explore multi-turn conversational spaces, efficiently identifying sub-instruction sequences that induce harmful responses. We evaluate DAMON's efficacy across five LLMs and three datasets. Our experimental results show that DAMON can effectively induce undesired behaviors.

DAMON: A Dialogue-Aware MCTS Framework for Jailbreaking Large Language Models

As demonstrated by the proprietary Large Language Models (LLMs) such as GPT and Claude series, LLMs have the potential to achieve remarkable proficiency across a wide range of domains, including law, medicine, finance, science, code, etc., all within a single model. These capabilities are further augmented during the Supervised Fine-Tuning (SFT) phase. Despite their potential, existing work mainly focuses on domain-specific enhancements during fine-tuning, the challenge of which lies in catastrophic forgetting of knowledge across other domains. In this study, we introduce **VersaTune**, a novel data composition framework designed for enhancing LLMs' overall multi-domain capabilities during training. We begin with detecting the distribution of domain-specific knowledge within the base model, followed by the training data composition that aligns with the model's existing knowledge distribution. During the subsequent training process, domain weights are dynamically adjusted based on their learnable potential and forgetting degree. Experimental results indicate that VersaTune is effective in multi-domain fostering, with an improvement of 29.77% in the overall multi-ability performances compared to uniform domain weights. Furthermore, we find that Qwen-2.5-32B + VersaTune even surpasses frontier models, including GPT-4o, Claude3.5-Sonnet and DeepSeek-V3 by 0.86%, 4.76% and 4.60%. Additionally, in scenarios where flexible expansion of a specific domain is required, VersaTune reduces the performance degradation in other domains by 38.77%, while preserving the training efficacy of the target domain.

VersaTune: An Efficient Data Composition Framework for Training Multi-Capability LLMs

Recent works have revealed the great potential of speculative decoding in accelerating the autoregressive generation process of large language models. The success of these methods relies on the alignment between draft candidates and the sampled outputs of the target model. Existing methods mainly achieve draft-target alignment with training-based methods, e.g., EAGLE, Medusa, involving considerable training costs. In this paper, we present a training-free alignment-augmented speculative decoding algorithm. We propose alignment sampling, which leverages output distribution obtained in the prefilling phase to provide more aligned draft candidates. To further benefit from high-quality but non-aligned draft candidates, we also introduce a simple yet effective flexible verification strategy. Through an adaptive probability threshold, our approach can improve generation accuracy while further improving inference efficiency. Experiments on 8 datasets (including question answering, summarization and code completion tasks) show that our approach increases the average generation score by 3.3 points for the LLaMA3 model. Our method achieves a mean acceptance length up to 2.39 and speed up generation by 2.23×.

Downloads

Next from EMNLP 2025

QuZO: Quantized Zeroth-Order Fine-Tuning for Large Language Models

Similar lecture

Reasoning Model Unlearning: Forgetting Traces, Not Just Answers, While Preserving Reasoning Skills