arXiv cs.AI· Weihan Li, Xinlei Chen, Yuhan Song, Xiaofeng Lin, Tianshi Zheng·· 3 天前AI 评分22
LLM 中可操控的效价方向究竟表征了什么？相同结果、不同读出
Same Outcome, Different Readout: What Does a Steerable Valence Direction in LLMs Represent?
AI 导读
一项研究以迷宫任务中的“好-坏”结果方向为对象，用受控干预区分实际结果与信息历史，检验该方向的构念效度。跨多个 LLM checkpoint，基于一种显式结果编码拟合的方向能良好迁移到另一种编码，但同一结果经由“预告”与“未预告”两种历史达成时迁移显著下降，事件后读出仍强烈依赖早先的预告。
正文 · 原文
Abstract:Claims about what an internal direction in an LLM represents need evidential constraints beyond an observer's prior beliefs about the system. Decodability and successful activation steering do not, by themselves, establish which construct the direction tracks. This gap is especially consequential for welfare-relevant interpretations, where a proposed functional state must be distinguished from correlated features of the extraction contrast. We treat the question as one of construct validity and study a good-bad outcome direction in a maze task, using controlled interventions that separate the realised outcome from the informational history through which it became known. Across multiple LLM checkpoints, directions fitted on one explicit outcome encoding transfer well to another, indicating that the readout is not tied to surface form. When the same realised outcome is reached through announced and unannounced histories, however, transfer degrades substantially: even after both histories receive the same explicit outcome, the post-event readout remains strongly conditioned on the earlier announcement. In the base model, steering along the direction changes actions, yet removing it leaves the natural cue effect almost intact. In a matched maze-RL run, the post-RL direction becomes substantially more predictive of reference-MDP remaining return and the policy becomes more dependent on it at the tested sites, while the history dependence persists. These dissociations support a functional, value-related interpretation of the direction, but not its identification with a history-invariant scalar valence state.
| Comments: | 19 pages, 13 figures, 10 tables |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2609.22850 [cs.LG] |
| (or arXiv:2609.22850v4 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2609.22850 arXiv-issued DOI via DataCite |
Submission history
From: Weihan Li [view email]
[v1]
Sat, 19 Sep 2026 07:46:19 UTC (2,849 KB)
[v2]
Tue, 22 Sep 2026 18:34:11 UTC (581 KB)
[v3]
Sun, 27 Sep 2026 12:57:17 UTC (1,301 KB)
[v4]
Tue, 13 Oct 2026 10:38:38 UTC (1,358 KB)
来源：arXiv cs.AI · arxiv.org