Towards Reasoning-Preserving Unlearning in Multimodal Large Language Models
Hongji Li, Junchi yao, Manjiang Yu, Priyanka Singh, Xue Li, Di Wang, Lijie Hu
TL;DR
This work tackles unlearning for reasoning-capable multimodal LLMs by introducing RMLLMU-Bench to evaluate not only forgetting but also leakage and retention of reasoning. It presents R-MUSE, a training-free inference-time framework that constructs span-based unlearning directions and protects general reasoning via a Reasoning Retain Subspace, augmented by Adaptive Calibration Steering grounded in optimal transport on the sphere. Empirical results show R-MUSE achieves stronger forgetting and lower reasoning leakage than baselines while maintaining downstream reasoning capabilities, and ablations confirm the necessity of the RRS and ACS components. The approach offers a practical, hyperparameter-light solution to privacy-aware unlearning in multimodal reasoning systems with broad implications for safe deployment. The work also provides a rigorous theoretical and visualization backing for alignment between geometric steering and reasoning preservation.
Abstract
Machine unlearning aims to erase requested data from trained models without full retraining. For Reasoning Multimodal Large Language Models (RMLLMs), this is uniquely challenging: intermediate chain-of-thought steps can still leak sensitive information even when final answers are forgotten, and overly aggressive interventions easily damage general reasoning ability. Yet no benchmark jointly evaluates how well unlearning methods suppress reasoning-level leakage while preserving reasoning competence. We address this gap with RMLLMU-Bench, the first benchmark for RMLLM unlearning that extends standard forgetting metrics with dedicated measures of reasoning leakage and reasoning retention. A systematic evaluation on RMLLMU-Bench reveals that existing unlearning methods for MLLMs and Large (Language) Reasoning Models (LRMs) either leave substantial leakage in the reasoning process or severely degrade reasoning performance. To address these gaps, we propose R-MUSE (Reasoning-preserving MLLM Unlearning via Subspace guidance and Adaptive Steering), a training-free and inference-time intervention framework that steers internal representations to forget both answers and reasoning traces while explicitly preserving general reasoning. Experiments on RMLLMU-Bench demonstrate that R-MUSE achieves a substantially better balance between effective forgetting and reasoning retention.