Search papers, labs, and topics across Lattice.
This paper investigates the phenomenon of trait-induced safety variation in aligned large language models (LLMs), where the same request can lead to different safety decisions based on the assigned traits in the system prompt. To address this issue, the authors introduce Trait-Invariant Safety Tuning (TIST) and Trait-Subspace Neutralization (TraSN), which align the model's behavior across different traits while maintaining its general capabilities. Experimental results demonstrate that TraSN effectively enhances trait-invariant safety and improves responses to harmful requests, underscoring the significance of traits in LLM safety dynamics.
Trait-induced safety variation can lead to inconsistent safety decisions in LLMs, but a new tuning method stabilizes their behavior across different traits.
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.
