ACTR: Aligning Chain of Thought with Answers to Improve Multilingual Safety
Start here
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 the neuron-selective consistency method.
Generated by AI from reports · Updated 4 days ago
Progress timeline
- arXiv：越狱、提示注入、投毒与防御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.
- arXiv：越狱、提示注入、投毒与防御ACTR: Aligning Reasoning with Answers to Improve Multilingual Safety of Reasoning Large Language Models
The researchers propose the ACTR framework, which improves multilingual safety alignment by strengthening the safety reasoning that reasoning large language models already have. The method first uses the think gap score (TGS) to compare the normalized contribution of reasoning traces to attention outputs across languages, and uses reasoning trace 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 trace and the answer on safety categories, updating only parameters related to selected neurons, without requiring human-annotated answers or preference data.