arXiv cs.CL· Chung-En Sun, Ge Yan, Zimo Wang, Tsui-Wei Weng·· 11 小时前AI 评分32
Steer2Edit：从激活引导到组件级权重编辑的无训练框架
Steer2Edit: From Activation Steering to Component-Level Editing
AI 导读
Steer2Edit 提出一种无需训练的框架，将推理时的引导向量转化为诊断信号，用于对注意力头和 MLP 神经元做组件级 rank-1 权重编辑，而非在生成时全局注入固定方向。在安全对齐、幻觉缓解和推理效率任务上，匹配下游性能时安全提升最高 17.2%、真实性提升 9.8%、推理长度平均缩短 12.2%。该方法保持标准前向传播并兼容并行推理优化，代码已开源，论文被 NeurIPS 2026 接收。
正文 · 原文
Abstract:Steering methods influence Large Language Model behavior by identifying semantic directions in hidden representations, but are typically realized through inference-time activation interventions that apply a fixed, global modification to the model's internal states. While effective, such interventions often induce unfavorable attribute-utility trade-offs under strong control, as they ignore the fact that many behaviors are governed by a small and heterogeneous subset of model components. We propose Steer2Edit, a theoretically grounded, training-free framework that transforms steering vectors from inference-time control signals into diagnostic signals for component-level rank-1 weight editing. Instead of uniformly injecting a steering direction during generation, Steer2Edit selectively redistributes behavioral influence across individual attention heads and MLP neurons, yielding interpretable edits that preserve the standard forward pass and remain compatible with optimized parallel inference. Across safety alignment, hallucination mitigation, and reasoning efficiency, Steer2Edit consistently achieves more favorable attribute-utility trade-offs: at matched downstream performance, it improves safety by up to 17.2%, increases truthfulness by 9.8%, and reduces reasoning length by 12.2% on average. Overall, Steer2Edit provides a principled bridge between representation steering and weight editing by translating steering signals into interpretable, training-free parameter updates. Our code is available at this https URL
| Comments: | Accepted to NeurIPS 2026 |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2602.09870 [cs.CL] |
| (or arXiv:2602.09870v3 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2602.09870 arXiv-issued DOI via DataCite |
Submission history
From: Chung En Sun [view email]
[v1]
Tue, 10 Feb 2026 15:15:15 UTC (1,543 KB)
[v2]
Tue, 3 Mar 2026 02:47:10 UTC (1,543 KB)
[v3]
Wed, 4 Nov 2026 03:41:39 UTC (1,596 KB)
来源：arXiv cs.CL · arxiv.org