arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2604.02113cs.CL
收藏

用于大语言模型转向推理的可靠控制点选择

Reliable Control-Point Selection for Steering Reasoning in Large Language Models

University of Notre Dame(圣母大学)

Haomin Zhuang, Hojun Yoo, Xiaonan Luo, Kehan Guo, Xiangliang Zhang

AI总结:

本文提出稳定过滤方法,通过概率模型识别稳定的行为边界,提升转向向量的稳定性与准确性,在MATH-500数据集上取得0.784准确率。

摘要:

转向向量提供一种无需训练的机制,用于控制大语言模型的推理行为,但构建有效的向量需要识别模型隐藏状态中的真实行为信号。对于可通过提示切换的行为,这很简单。然而,许多推理行为——如自我反思——自发出现且难以通过提示级控制。当前方法通过链式思维轨迹中的关键词匹配检测这些行为,隐含假设每个检测到的边界都编码真实行为信号。我们证明这一假设普遍错误:在541个关键词检测边界中,93.3%行为不稳定,无法在相同前缀下重复检测行为。我们开发了一个概率模型,将内在推理行为形式化为具有上下文依赖触发概率的随机事件,并显示不稳定边界稀释了转向信号。基于此分析,我们提出稳定过滤,仅保留模型一致重复目标行为的边界。结合内容子空间投影去除残余问题特定噪声,我们的方法在MATH-500数据集上达到0.784准确率(比最强基线高5.0)。所获得的转向向量在相同架构家族的模型间转移,无需重新提取,提升了Nemotron-Research-Reasoning-1.5B(+5.0)和DeepScaleR-1.5B-Preview(+6.0)。代码可在https://github.com/zhmzm/stability-steering获取。

英文摘要:

Steering vectors offer a training-free mechanism for controlling reasoning behaviors in large language models, but constructing effective vectors requires identifying genuine behavioral signals in the model's hidden states. For behaviors that can be toggled via prompts, this is straightforward. However, many reasoning behaviors -- such as self-reflection -- emerge spontaneously and resist prompt-level control. Current methods detect these behaviors through keyword matching in chain-of-thought traces, implicitly assuming that every detected boundary encodes a genuine behavioral signal. We show that this assumption is overwhelmingly wrong: across 541 keyword-detected boundaries, 93.3\% are behaviorally unstable, failing to reproduce the detected behavior under re-generation from the same prefix. We develop a probabilistic model that formalizes intrinsic reasoning behaviors as stochastic events with context-dependent trigger probabilities, and show that unstable boundaries dilute the steering signal. Guided by this analysis, we propose stability filtering, which retains only boundaries where the model consistently reproduces the target behavior. Combined with a content-subspace projection that removes residual question-specific noise, our method achieves 0.784 accuracy on MATH-500 (+5.0 over the strongest baseline). The resulting steering vectors transfer across models in the same architecture family without re-extraction, improving Nemotron-Research-Reasoning-1.5B (+5.0) and DeepScaleR-1.5B-Preview (+6.0). Code is available at https://github.com/zhmzm/stability-steering.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑