New Tuning Method Stabilizes LLM Safety Across Traits

Large language models (LLMs) are expected to consistently exhibit safe behavior, refusing harmful requests while complying with safe ones. However, a significant issue, termed "trait-induced safety variation," occurs when the same user request elicits different safety decisions depending on the traits assigned to the LLM in its system prompt. This inconsistency undermines trust and reliability. To quantify this problem, new metrics like Trait-Induced Deviation and Trait-Induced Flip Rate were introduced. Analysis revealed that traits perturb the model's safety representations within a low-dimensional subspace. To counter this, researchers developed Trait-Invariant Safety Tuning (TIST), a self-distillation framework designed to align an LLM's trait-conditioned behavior with its no-trait baseline. A specific implementation of TIST, called Trait-Subspace Neutralization (TraSN), enforces invariance only within the identified trait subspace. Experiments demonstrate that TraSN effectively improves trait-invariant safety, strengthens the model's refusal of harmful requests, and preserves general capabilities. These findings highlight the critical role of traits in LLM safety and the importance of robust model behavior across varying contextual prompts.