arXiv cs.AI· Seyed Arshan Dalili, Ajay Narayanan Sridhar, Vijaykrishnan Narayanan, Mehrdad Mahdavi·· 13 小时前AI 评分22
Cobras：基于黎曼激活引导的条件最优桥接方法
Conditional Optimal Bridge for Riemannian Activation Steering
AI 导读
研究者提出 Cobras，将大语言模型的激活引导建模为残差流超球面上的薛定谔桥，从良定义优化问题中推导出对数密度比引导目标，并构建查询自适应引导场，把对数密度比引导作为极限情形纳入。该方法还能给出支撑分数以控制 OOD 干预。在四个模型上，Cobras 在有用性、真实性和去毒等任务中几乎全部模型—任务设置下优于既有引导方法，同时基本保持 OOD 性能，代码已开源。
正文 · 原文
Abstract:Activation steering offers a lightweight alternative to fine-tuning for controlling large language models at inference time. Many existing steering methods implicitly optimize a log-density-ratio objective between desired and undesired activation distributions without a principled derivation. Methods with fixed steering directions can also degrade out-of-distribution (OOD) performance. We introduce Conditional Optimal Bridge for Riemannian Activation Steering (\textsc{Cobras}), which casts steering as a Schrödinger Bridge on the residual-stream hypersphere. This formulation yields a principled derivation of the log-density-ratio steering objective from a well-posed optimization problem. We construct a query-adaptive steering field that recovers log-density-ratio steering as a limiting case. The fitted transport also provides a support score for controlling OOD interventions. Across four models, \textsc{Cobras} outperforms prior steering methods in nearly all model--task settings spanning helpfulness, truthfulness, and detoxification, while largely preserving OOD performance. The code is available \href{this https URL}{here}.
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2607.10517 [cs.LG] |
| (or arXiv:2607.10517v2 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2607.10517 arXiv-issued DOI via DataCite |
Submission history
From: Seyed Arshan Dalili [view email]
[v1]
Sun, 12 Jul 2026 00:36:46 UTC (1,589 KB)
[v2]
Wed, 21 Oct 2026 11:42:23 UTC (2,130 KB)
来源：arXiv cs.AI · arxiv.org