← 研踪 PaperTrailarXiv:2611.17620 · v1
Safety Side Effects of Preventative Persona Steering: An Empirical Audit
首次提交 2026-11-13 08:45
版本更新 2026-11-13 08:45
雷达观测 2026-11-17 04:58
摘要
High-level behaviors such as sycophancy and hallucination correspond to directions in a model's activation space (Chen et al., 2025). These directions, known as persona vectors, can suppress undesirable traits either at inference time, subtracting the vector during generation, or through preventative steering, adding the vector during fine-tuning, so the optimizer never needs to encode the trait in weights, then removing the vector at evaluation. Chen et al. (2025) recommend preventative steering because it preserves Massive Multitask Language Understanding (MMLU) accuracy, but validations of this recommendation have measured capability alone. We primarily evaluate results on preventative steering of sycophancy and hallucination vectors on three models, measuring three safety benchmarks alongside trait expression, coherence, and MMLU. Our findings show that preventative steering still changes refusal behavior, but by an amount that depends on the trait and does not establish that preventative persona steering is safety-neutral, and steering methods should be evaluated on several safety benchmarks.
作者备注:5 pages, 1 figure (6 charts), 1 table, accepted at NeurIPS 2026 Bialign Workshop
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-13 08:45 · 最新观测版本
Safety Side Effects of Preventative Persona Steering: An Empirical Audit
High-level behaviors such as sycophancy and hallucination correspond to directions in a model's activation space (Chen et al., 2025). These directions, known as persona vectors, can suppress undesirable traits either at inference time, subtracting the vector during generation, or through preventative steering, adding the vector during fine-tuning, so the optimizer never needs to encode the trait in weights, then removing the vector at evaluation. Chen et al. (2025) recommend preventative steering because it preserves Massive Multitask Language Understanding (MMLU) accuracy, but validations of this recommendation have measured capability alone. We primarily evaluate results on preventative steering of sycophancy and hallucination vectors on three models, measuring three safety benchmarks alongside trait expression, coherence, and MMLU. Our findings show that preventative steering still changes refusal behavior, but by an amount that depends on the trait and does not establish that preventative persona steering is safety-neutral, and steering methods should be evaluated on several safety benchmarks.
查看此版本解读 →查看此版本原文 ↗