免微调的 Task Vector 方法 TFTVs 实现 LLM 行为控制
Training-Free Task Vectors for LLM Behavioral Control
AI 导读
NeurIPS 2026 论文提出免微调任务向量方法 TFTVs，仅用前向传播统计量将激活引导向量映射为秩一权重空间编辑，无需微调即可计算任务向量。该方法支持通过加法学习、减法遗忘及多编辑组合，在 LLM 行为控制任务上能持续放大、抑制和组合目标行为，同时保持通用知识与解题能力。对比其他编辑与引导基线，TFTVs 实现更强的特质控制且效用保持相当或更优，代码已在项目网站开源。
正文 · 原文
Abstract:Task vectors enable post-training model editing by identifying semantically meaningful directions in weight space, typically computed as the difference between a fine-tuned model and its pretrained initialization. However, this reliance on fine-tuning makes discovering such directions costly and limits the practicality of post-training model editing. To address this limitation, we introduce Training-Free Task Vectors (TFTVs), a novel method to compute task-vector-like directions without requiring fine-tuning. Our method maps activation steering vectors to rank-one weight-space edits using only forward-pass statistics, while satisfying arithmetic properties that directly support learning via addition, forgetting via subtraction, and the composition of multiple edits. Empirically, we evaluate TFTVs on large language model behavioral control tasks and show that they consistently amplify, suppress, and compose target behaviors while preserving general knowledge and problem-solving skills. We also validate our method against other editing and steering baselines, experimentally demonstrating that TFTVs achieve stronger trait control with better or competitive utility preservation. We hope our work opens new directions for the community in post-training model editing and broader training-free model control. Code is available on the project website: this http URL.
| Comments: | Conference on Neural Information Processing Systems (NeurIPS 2026) |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2609.09054 [cs.LG] |
| (or arXiv:2609.09054v3 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2609.09054 arXiv-issued DOI via DataCite |
Submission history
From: Gabriel Perin [view email]
[v1]
Tue, 8 Sep 2026 17:10:28 UTC (683 KB)
[v2]
Wed, 14 Oct 2026 17:32:51 UTC (659 KB)
[v3]
Thu, 22 Oct 2026 17:33:28 UTC (659 KB)
来源：arXiv cs.AI · arxiv.org