Towards Unveiling Vulnerabilities of Large Reasoning Models in Machine Unlearning

Aobo Chen, Chenxu Zhao, Chenglin Miao, Mengdi Huai

TL;DR

This paper exposes a critical security vulnerability in Large Reasoning Models (LRMs) undergoing machine unlearning. The authors introduce a novel bi-level optimization framework that manipulates the unlearning process, forcing the model to output incorrect final answers supported by highly convincing, yet entirely deceptive, reasoning traces. This work highlights an urgent need for more secure unlearning pipelines as models increasingly rely on chain-of-thought reasoning.

What Problem Does This Paper Address?

Machine unlearning has emerged as a vital technique for enforcing the "right to be forgotten" (e.g., GDPR compliance), allowing model owners to erase specific sensitive or copyrighted training data without the massive computational expense of retraining from scratch. However, the unlearning process alters model weights and introduces a new attack surface.

While prior work has explored unlearning attacks on standard Large Language Models (LLMs)—typically trying to force simple label misclassifications—this paper is the first to target Large Reasoning Models (LRMs). LRMs generate explicit, multi-step reasoning traces (rationales) before arriving at a final conclusion.

An effective attack on an LRM must be stealthy: if an attacker only manipulates the final answer, a human reviewer will easily spot the mismatch against a correct reasoning trace. Therefore, the attack must jointly manipulate both the final answer and the reasoning chain so that the model "lies with convincing proof." Crafting such an attack introduces three major technical challenges:

  1. Non-differentiable logical constraints: Balancing answer incorrectness with reasoning coherence creates a discrete optimization landscape.
  2. Gradient dilution: LRMs generate long rationales, which dilutes gradient signals across many tokens and renders standard optimization ineffective.
  3. Discrete data selection: Exact unlearning requires selecting a discrete subset of training data to remove, which prevents end-to-end gradient-based optimization.

Methods

To overcome these hurdles, the authors propose a bi-level exact unlearning attack framework:

  • Differentiable Surrogate Objective: The authors replace the non-differentiable dual-constraint (wrong answer + coherent reasoning) with a surrogate loss ($\ell_1$) that maximizes the likelihood of pre-defined adversarial sequences containing the incorrect answer and a modified reasoning trace.
  • Trajectory Guidance Loss ($\ell_2$): To combat gradient dilution over long rationales, they introduce a loss that concentrates optimization signals onto a small set of highly influential anchor tokens (e.g., promoting words like "superior" or "positive" to steer a sentiment analysis model toward a false positive conclusion).
  • Relaxed Indicator Strategy & Gradient Matching: To handle the discrete nature of selecting which training samples to "forget," the authors relax the selection indicators to continuous weights. They then apply a gradient matching technique to align the gradient directions of the attack objective with the model updates, simplifying the complex bi-level optimization into a tractable integer programming problem.

The authors also extend their framework to an adversarial unlearning attack setting, where a fixed-length sequence of adversarial trigger tokens ($\delta$) is appended to the forget data using Greedy Coordinate Gradient (GCG) optimization.

Datasets

The framework was evaluated using several diverse datasets:

  • SST-2 & Twitter: Used for sentiment classification tasks.
  • Star-1: A dataset designed for the safe alignment of reasoning LLMs.
  • TOFU (Task of Fictitious Unlearning): A benchmark specifically designed to evaluate LLM unlearning performance.

Evaluations were performed across multiple models, including three LRMs (OpenReasoning-Nemotron-1.5B, Skywork-o1-Open-Llama-3.1-8B, and Skywork-OR1-7B) and two standard LLMs (Llama-2-7B and Mistral-7B-v0.1).

Key Insights and Ideas

  • Stealth Through Deception: The core insight is that models don't just fail randomly under unlearning attacks; they can be systematically trained to rationalize their errors. As shown in Figure 1, a successful attack doesn't break the chain of thought; it hijacks it to make an incorrect answer look perfectly logical to a human inspector.
Fig. 1: Visualization of our attack on LRM. The final answers are underlined, and the sentiment-related terms are bolded.
  • Anchor Tokens Matter More Than Full Traces: Trying to optimize every token in a long reasoning chain is mathematically inefficient. By focusing purely on high-impact trajectory-guiding tokens, attackers can effectively steer long-form model outputs.
  • High Black-Box Transferability: The vulnerabilities exposed by this attack are structural rather than model-specific. An attack optimized on a surrogate model transfers remarkably well to completely different target architectures, making it a highly practical threat in real-world scenarios.

Key Results

  • Attack Success Rate (ASR): In white-box settings, the proposed adversarial unlearning method achieved an ASR up to three times higher than random baselines at a 5% unlearning ratio.
  • Reasoning Quality Degradation: When configured to degrade reasoning quality while preserving final answer accuracy, the method successfully drove up the Performance Degradation Ratio (PDR) from 0.4 to 0.7 as the unlearning ratio increased, while keeping the Performance Preservation Ratio (PPR) above 1.0.
  • Black-Box Transferability: In black-box cross-model transfer experiments (using a surrogate dataset derived from Twitter data to attack an SST-2 target model), the attack achieved high max-normalized ASR scores, notably reaching 0.94 when transferring from an OpenReasoning-Nemotron source model to a Skywork-OR1 target model.

Limitations and Future Work

The authors focus primarily on demonstrating the existence and severity of this vulnerability. Consequently, the paper does not propose a concrete defense mechanism to mitigate these attacks. The authors note that their primary goal is to raise awareness, leaving the development of robust, attack-resistant LRM unlearning pipelines as an urgent open challenge for future research.

Summary updated

Abstract

Large language models (LLMs) possess strong semantic understanding, driving significant progress in data mining applications. This is further enhanced by large reasoning models (LRMs), which provide explicit multi-step reasoning traces. On the other hand, the growing need for the right to be forgotten has driven the development of machine unlearning techniques, which aim to eliminate the influence of specific data from trained models without full retraining. However, unlearning may also introduce new security vulnerabilities by exposing additional interaction surfaces. Although many studies have investigated unlearning attacks, there is no prior work on LRMs. To bridge the gap, we first in this paper propose LRM unlearning attack that forces incorrect final answers while generating convincing but misleading reasoning traces. This objective is challenging due to non-differentiable logical constraints, weak optimization effect over long rationales, and discrete forget set selection. To overcome these challenges, we introduce a bi-level exact unlearning attack that incorporates a differentiable objective function, influential token alignment, and a relaxed indicator strategy. To demonstrate the effectiveness and generalizability of our attack, we also design novel optimization frameworks and conduct comprehensive experiments in both white-box and black-box settings, aiming to raise awareness of the emerging threats to LRM unlearning pipelines.
BibTeX
@inproceedings{chen2026towards,
  author = {Aobo Chen and Chenxu Zhao and Chenglin Miao and Mengdi Huai},
  title = {{Towards Unveiling Vulnerabilities of Large Reasoning Models in Machine Unlearning}},
  booktitle = {PAKDD},
  year = {2026},
  doi = {10.1007/978-981-92-1468-6\_36},
}
Figure