LEMUR Enhances Privacy in Multimodal LLMs by Unlearning Sensitive Reasoning.

Reinforcement learning (RL) post-training significantly improves the visual reasoning capabilities of multimodal large reasoning models (MLRMs) by enabling them to generate exploratory chains of thought (CoT). However, this enhanced reasoning introduces a new privacy vulnerability: even if a sensitive fact is removed from the final answer, it can still appear in the model's internal reasoning process. This leakage is particularly pronounced in RL-trained MLRMs compared to their base models. The researchers observed that RL-induced exploration leaves a distinct token-level entropy signature associated with sensitive content. Leveraging this insight, they developed LEMUR (Latent Entropy-aware Multimodal Unlearning via Visual-anchored Reasoning Redirection). LEMUR is a training-free, inference-time framework that uses these entropy dynamics as a control signal. When sensitive reasoning is detected, LEMUR redirects the model's reasoning trajectory. It achieves this by using entropy-modulated visual-anchor latent injection, replacing problematic tokens with sanitized, probability-weighted embeddings re-grounded in the input image. Experiments show LEMUR effectively suppresses both reasoning-trace and answer leakage across various MLRMs, outperforming existing methods while maintaining non-sensitive utility and output fluency.