Large reasoning models learn to forget sensitive details naturally
GUARD: Natural Forgetting in Large Reasoning Models via Guided Answer-Reasoning Distillation
Artificial Intelligence
Summary
Large AI models that think through answers step-by-step can accidentally reveal private or unsafe information during their reasoning process. The authors propose a method called GUARD that teaches these models to forget such sensitive details in a natural way, by guiding the reasoning to safely avoid disclosing protected facts and instead provide a clear refusal when needed. They also created a way to measure how well the model’s new responses behave after forgetting sensitive content. Their tests show GUARD effectively reduces unsafe disclosures while keeping the model’s reasoning ability intact.
What this means in practice
- •For ai safety engineers: Improve AI systems by enabling natural forgetting of sensitive information in multi-step reasoning to prevent accidental disclosures.
- •For privacy compliance teams: Enforce data removal requests by retraining language models to refuse answering with protected information naturally during reasoning.
Authors
Zeyu Yan, Guanghao Zhou, Minghui Qiu, Ming Gao, Cen Chen
Abstract
Recent advances in large reasoning models (LRMs) have made machine unlearning more challenging, as protected facts or unsafe rationales may surface in intermediate chain-of-thought (CoT) traces before the final answer is produced. Existing unlearning objectives typically suppress the target content or redirect internal representations, but they never specify how the post-forgetting trajectory should continue, which can lead to hallucinated substitutes, malformed boundaries, or repetitive outputs. We argue that LRM unlearning should instead learn a natural forgetting trajectory: a coherent non-disclosing CoT followed by a stable refusal-style answer that replace the original disclosure. To this end, we propose Guided Answer-Reasoning Distillation (GUARD), which converts model-generated unsafe disclosures into safe-exit trajectories, aligns a frozen LRM via guidance tokens, and distills the guided behavior into model parameters.To address the lack of metrics for replacement quality beyond leakage, we further introduce Natural Forgetting Reasoning Score (NFRS), which captures structural stability, fluency, and unsupported substitutes in forgotten outputs. Extensive experiments on R-TOFU and a STAR-1-derived harmful-intent setting show that GUARD substantially reduces unsafe and privacy disclosures across two widely adopted distilled LRMs while preserving reasoning utility. Codes are available at https://github.com/zeyu-Yan/GUARD