解码却不可控：审计冻结 LLM 中表征与行为在几何约束上的落差
Decodable but Not Actionable: Auditing the Representation-Behavior Gap in Frozen LLMs for Geometric Constraints
AI 导读
一项 arXiv 研究提出基于参数化 CAD 的受控审计框架，区分 LLM 在几何约束上的编码、表达与激活控制三类失败。对 8 个预训练模型和随机初始化对照的测试显示，局部关系可线性解码但零样本预测远低于监督探针性能，激活转向也无法提供可靠定向控制。仅用 400 个标注关系对做低秩微调即可显著提升三个代表性模型的生成表现，说明表征—行为落差可通过适配缩小。
正文 · 原文
Abstract:Large language models (LLMs) perform well on structured reasoning tasks, but what they encode and whether encoded information guides their behavior remain unclear. We introduce a controlled auditing framework that uses parametric CAD to distinguish failures of encoding, expression, and activation-level control. Applied to eight pretrained models and matched randomly initialized controls, our framework reveals that pretraining contributes substantially more to linear decoding of local relations than to global constraint status, which is already highly decodable from random representations. Yet strong local decodability does not translate into the models' zero-shot predictions, which remain far below supervised probe performance. Restoring clean internal activations can repair probe predictions, whereas simple activation steering does not provide reliable targeted control. With only 400 labeled relation pairs, low-rank fine-tuning substantially improves generation on three representative models, showing that the representation--behavior gap can be reduced through adaptation. Together, these findings show why decodability alone is insufficient evidence that encoded information will be expressed or controllable in model behavior.
| Comments: | Revised manuscript with expanded experiments, analyses, and discussion |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2608.17843 [cs.CL] |
| (or arXiv:2608.17843v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2608.17843 arXiv-issued DOI via DataCite |
Submission history
From: Man Liang [view email]
[v1]
Tue, 18 Aug 2026 14:40:28 UTC (1,567 KB)
[v2]
Tue, 27 Oct 2026 23:04:55 UTC (1,024 KB)
来源：arXiv cs.AI · arxiv.org