arXiv cs.AI· Hui Chen, Antoine Didisheim, Mo Pourmohammadi, Luciano Somoza, Hanqing Tian·· 11 小时前AI 评分40
用稀疏自编码器对 LLM 做“脑部扫描”：解读内部特征并定向调控经济偏好
A Financial Brain Scan of the LLM
AI 导读
研究用稀疏自编码器（SAE）对 LLM 做“脑部扫描”，将内部激活映射为可解释概念并实现对特定特征的定向调控。结果显示，提示模型采用特定偏好会激活许多非预期概念，而 SAE 能更平滑、更隔离地调控目标概念；模型内部特征贡献与其自生成解释存在明显脱节，后者不能可靠反映其推理过程。SAE 还为现有模型提供轻量干预手段，可生成可解释嵌入向量并减弱特定经济偏见。
正文 · 原文
Abstract:We use sparse autoencoders (SAEs) to "brain scan" large language models, mapping internal activations to interpretable concepts and enabling targeted "steering" of specific features. We show that prompting models to adopt specific preferences activates many unintended concepts, complicating causal inference. In contrast, SAEs permit smooth, more isolated steering of target concepts. Furthermore, our scans reveal a stark disconnect between the model's internal feature contributions and its self-generated explanations, showing that these explanations are unreliable guides to its reasoning. Finally, SAEs provide a lightweight intervention for existing models, producing interpretable embeddings and allowing researchers to attenuate specific economic biases.
| Comments: | 82 pages |
| Subjects: | General Finance (q-fin.GN); Artificial Intelligence (cs.AI); Computational Engineering, Finance, and Science (cs.CE); General Economics (econ.GN) |
| Cite as: | arXiv:2508.21285 [q-fin.GN] |
| (or arXiv:2508.21285v4 [q-fin.GN] for this version) | |
| https://doi.org/10.48550/arXiv.2508.21285 arXiv-issued DOI via DataCite |
Submission history
From: Hanqing Tian [view email]
[v1]
Fri, 29 Aug 2025 01:12:57 UTC (32,547 KB)
[v2]
Fri, 13 Feb 2026 23:28:19 UTC (32,547 KB)
[v3]
Tue, 3 Nov 2026 04:13:06 UTC (10,862 KB)
[v4]
Wed, 4 Nov 2026 04:06:59 UTC (10,862 KB)
来源：arXiv cs.AI · arxiv.org