ACTR: Aligning Reasoning Chains with Answers to Improve Multilingual Safety of Reasoning Large Models

The researchers propose the ACTR framework, which improves multilingual safety alignment by strengthening the safety reasoning that reasoning large models already have. The framework first uses the think gap score (TGS) to compare the normalized contribution of reasoning chains to attention outputs across languages, and uses reasoning chain replacement to measure cross-language safety gaps; then, using jailbreak query corpora, it evaluates neuron importance through the response representation changes caused by neuron masking, compares the sets of high-importance neurons when reasoning is enabled versus disabled, and identifies safety think neurons that support safety reasoning; finally, it proposes neuron-selective consistency optimization (NSCO), which uses a frozen judge model to reward consistency between the reasoning chain and the answer on safety categories, updating only the parameters of selected neurons, without requiring human-annotated answers or preference data.