Making Your LLMs More Objective: Stabilizing LLM Safety Behavior Across Traits with Trait-Invariant Safety Tuning

2026-08-12

AI Summary

This work identifies and quantifies a previously unexamined phenomenon—trait-induced safety variation—where large language models exhibit inconsistent safety judgments for identical requests under different system prompts that induce distinct personality traits, leading to unstable safety behaviors. To address this issue, the authors propose Trait-Subspace Neutralization (TraSN), a method that neutralizes the perturbations caused by personality traits on safety representations within a low-dimensional subspace. Built upon a self-distillation framework coupled with representation analysis, TraSN effectively aligns safety behaviors across diverse trait settings. The approach significantly enhances the model’s ability to refuse harmful requests while preserving its general-purpose capabilities, thereby improving overall safety stability without compromising utility.

Abstract

Aligned large language models (LLMs) are expected to exhibit safety behavior based on the content of the user request: they should refuse unsafe requests and comply with safe ones. However, we show that the same request can elicit substantially different safety decisions under different traits assigned in the system prompt, a failure mode we call trait-induced safety variation. To measure this failure, we introduce refusal-based metrics: Trait-Induced Deviation measures dataset-level deviation from the no-trait baseline, while Trait-Induced Flip Rate measures whether the same request receives different safety decisions across traits. We then provide a representation-level analysis of the mechanism behind trait-induced safety shifts and find that traits perturb the model's safety representations within a low-dimensional subspace. To achieve trait-invariant safety, where safety behavior remains stable across traits, we introduce Trait-Invariant Safety Tuning (TIST), a simple yet effective self-distillation framework that aligns an LLM's trait-conditioned behavior with its no-trait behavior. Guided by our analysis, we further propose Trait-Subspace Neutralization (TraSN), an instantiation of TIST, which enforces invariance only within the identified trait subspace. Experiments show that TraSN improves trait-invariant safety and strengthens harmful-request safety while preserving general capability. Our results highlight traits as an important factor in LLM safety and robust model behavior.

Problem

• trait-induced safety variation
• safety behavior
• large language models
• system prompt traits
• refusal consistency

Innovation

• trait-invariant safety
• safety alignment
• self-distillation
• representation subspace
• large language models