Search papers, labs, and topics across Lattice.
This paper introduces LEMUR, a novel inference-time unlearning framework designed to address privacy vulnerabilities in reinforcement-learning (RL) trained multimodal large reasoning models (MLRMs). The authors identify that even after unlearning sensitive facts, these models can inadvertently leak information through their reasoning traces, particularly due to distinctive entropy signatures left by RL-induced exploration. LEMUR effectively mitigates this leakage by redirecting the reasoning trajectory using entropy dynamics to sanitize sensitive content while maintaining the model's overall utility and fluency.
RL-trained multimodal models can leak sensitive information through reasoning traces, but LEMUR offers a training-free solution that effectively sanitizes this leakage without sacrificing output quality.
Reinforcement-learning (RL) post-training equips multimodal large reasoning models (MLRMs) with exploratory chains of thought (CoT), substantially improving visual reasoning. However, we find that this capability introduces a distinct privacy vulnerability: even when a sensitive fact is successfully unlearned from the final answer, the model may still reproduce it in its reasoning trace. This leakage is substantially more pronounced in natively RL-trained MLRMs than in their non -reasoning base models, revealing a privacy risk that existing unlearning methods are not designed to address. We show that RL-induced exploration leaves sensitive content with a distinctive token-level entropy signature that is largely absent from base models. Based on this observation, we propose LEMUR, a fully training-free, inference-time unlearning framework for natively RL-trained multimodal models. LEMUR uses entropy dynamics as a control signal to identify when sensitive reasoning begins and when sanitization should stop. During this interval, it redirects the reasoning trajectory through entropy-modulated visual-anchor latent injection, replacing committed tokens with sanitized, probability-weighted embeddings re-grounded in the input image. Across diverse MLRMs, LEMUR consistently outperforms existing unlearning met hods in suppressing both reasoning-trace and answer leakage, while better preserving non-sensitive utility and output fluency. These results demonstrate that RL-induced entropy dynamics provide a distinctive signal for privacy leakage and that exploiting this signal enables effective training-free unlearning for reasoning-capable multimodal models.
