arXiv cs.CL· Ihor Vitenko·· 11 小时前AI 评分31
VERS：通过编辑激活解释来引导推理模型
VERS: Steering Reasoning Models by Editing Their Activation Explanations
AI 导读
VERS 将推理轨迹中"存疑点"的解释编辑转化为一个固定方向，并在段落断点处叠加该方向以引导模型推理。在 GSM8K 完整测试集上，生成 token 减少 45.8%，准确率仅下降 0.82 个百分点；MATH-500 上节省 31.3% token，准确率下降 3.05 个百分点。
正文 · 原文
Abstract:Reasoning language models write a chain of thought before they answer, and many checks in it do not change the final answer. Natural language autoencoders (NLAs) explain a model's activations in text and rebuild activations from text, and editing an explanation can steer a model at a single token. We study whether such edits can also steer how a model reasons. We build VERS (Verbalize, Edit, Reconstruct, Steer): it turns explanation edits at the doubt points of one reasoning trace into one fixed direction, and adds this direction at paragraph breaks. On the full GSM8K test, VERS cuts generated tokens by 45.8% at an accuracy change of -0.82 percentage points (pp); on MATH-500 it saves 31.3% at -3.05 pp. At half strength, its trade-off between tokens and accuracy lies between those of the published methods SEAL and DEER. With fitted affine maps and a strength tuned for each model, the direction transfers to larger Qwen3 models, up to 235B-A22B, and saves 30-35% of tokens. On competition problems, however, VERS loses 3-7 pp at every model size, and a fixed direction did not separate necessary checks from unneeded ones.
| Comments: | 36 pages. Code and checkpoints: this https URL |
| Subjects: | Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2611.03040 [cs.CL] |
| (or arXiv:2611.03040v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2611.03040 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Ihor Vitenko [view email]
[v1]
Mon, 2 Nov 2026 12:39:13 UTC (285 KB)
来源：arXiv cs.CL · arxiv.org