一致性不等于理解：面向神经场权重读取器的轨道审计
Consistency Is Not Understanding: Orbit Audits for Neural Field Weight Readers
Sungguk Cha · Mintae Kim · Youngsub Han · Byoung-Ki Jeon · Sangyeob Lee
中文摘要
从神经网络权重中识别视觉内容，需要将功能信息与任意参数坐标分离。通过保持所表示函数不变的隐藏神经元置换与正 ReLU 缩放来审计这种分离。研究包含 480 个独立拟合的形状场和 1,000 个 MNIST 场，变换副本按来源分组。在形状识别上，统计读取器在联合变换下准确率从 \ShapeRawStats\% 降至 \ShapeJointStats\%；归一化集合读取器保持 \ShapeNormSet\% 且预测不变。四个小型冻结语言模型用下一词元评分获得 \LMRange\% 原始准确率。另一项四配置推理实验对 1,920 个评分会话用生成答案获得 \FrontRawRange\% 原始准确率，而相同原始提示的新鲜重复在 \FrontRawRepeatRange\% 的来源上不一致。规范化使等价提示相同，但这既不确立识别，也不保证生成响应相同。联合正确性/一致性指标、语义线索控制、新鲜重复与精确有理数舍入边界区分了这些失败模式。结果是面向视觉权重读取器的可检查评估协议，结论限于所测场、提示与执行配置。
关键要点
- 01用保持函数不变的隐藏神经元置换与正 ReLU 缩放审计权重读取器，区分功能信息与参数坐标。
- 02形状识别中统计读取器在联合变换下准确率从 \ShapeRawStats\% 降至 \ShapeJointStats\%，归一化集合读取器保持 \ShapeNormSet\%。
- 03四个小型冻结语言模型下一词元评分原始准确率为 \LMRange\%。
- 04推理实验中生成答案原始准确率 \FrontRawRange\%，相同提示新鲜重复在 \FrontRawRepeatRange\% 来源上不一致。
- 05规范化使等价提示相同，但不保证识别或相同生成响应，结论限于所测场、提示与配置。
解读
尚无解读。
原始英文摘要
arXiv:2611.05798v1 Announce Type: new 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.