arXiv cs.AI· Sungguk Cha, Mintae Kim, Youngsub Han, Byoung-Ki Jeon, Sangyeob Lee·· 11 小时前AI 评分26
一致性不等于理解：面向神经场权重读取器的 Orbit 审计
Consistency Is Not Understanding: Orbit Audits for Neural Field Weight Readers
AI 导读
研究用隐藏神经元置换与正 ReLU 缩放等保函数变换审计神经网络权重读取器，发现统计读取器在联合变换下准确率大幅下降，而归一化集合读取器预测保持不变。实验覆盖 480 个独立拟合形状场与 1,000 个 MNIST 场，四个小型冻结语言模型用下一 token 打分取得原始准确率，另有 1,920 个评分会话的推理实验显示相同原始提示在部分样本上生成答案不一致。
正文 · 原文
Abstract:Recognizing visual content from neural-network weights requires separating functional information from arbitrary parameter coordinates. We audit this separation with hidden-neuron permutations and positive ReLU rescalings that preserve the represented function. The study contains 480 independently fitted shape fields and 1,000 MNIST fields, with transformed copies grouped by origin. On shape recognition, a statistics reader falls from \ShapeRawStats\% to \ShapeJointStats\% accuracy under joint transformations; a normalized set reader retains \ShapeNormSet\% with unchanged predictions. Four small frozen language models obtain \LMRange\% raw accuracy with next-token scoring. A separate four-configuration reasoning experiment with 1,920 scored sessions obtains \FrontRawRange\% raw accuracy with generated answers, while fresh identical raw prompts disagree on \FrontRawRepeatRange\% of origins. Canonicalization makes equivalent prompts identical, but this neither establishes recognition nor guarantees identical generated responses. Joint correctness/agreement metrics, semantic-cue controls, fresh repeats, and exact-rational rounding bounds separate these failure modes. The result is an inspectable evaluation protocol for visual weight readers, with claims restricted to the tested fields, prompts, and execution configurations.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.05798 [cs.AI] |
| (or arXiv:2611.05798v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2611.05798 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Sungguk Cha [view email]
[v1]
Wed, 4 Nov 2026 01:12:33 UTC (146 KB)
来源：arXiv cs.AI · arxiv.org