跳转至

推理捷径与价值对称性:对称性所允许的、架构所实现的与优化所选择的

文章背景与核心概要

本文从价值对称性(value symmetries)和自同构群(automorphism groups)的视角(遵循 Takemura、Inoue 和 Nishino 引入的框架)深入研究了“推理捷径”(reasoning shortcuts)——即通过非预期概念得出正确预测的基于规则的解决方案。文章指出了现有框架在异构基准测试中的局限性,并引入了按分量划分的层次结构,有效解决了诸如填充等朴素嵌入方法产生的严重误报病态问题。

在理论与计算方面,作者分析了十一个规则家族,推导了六个关于传递性条件及其失效的定理,并精确分类了布尔传递性。此外,本文证明了电路给定规则的坐标对称性惰性(coordinate symmetry-inertness)是 coNP-complete 的,自同构的存在性是 coNP-hard 的,并探讨了其在多项式层次结构中的复杂度。在实证方面,弱监督模型在理论层面上完美隔离了所有观察到的捷径,而端到端训练则验证了传递性预测,展现出极高的准确性与理论契合度。


摘要摘要 (Abstract Summary)

本论文通过价值对称性和自同构群的视角(遵循 Takemura、Inoue 和 Nishino 引入的框架),研究了推理捷径(reasoning shortcuts)——即通过非预期概念得出正确预测的基于规则的解决方案。

该研究的主要发现和贡献包括: * 对现有框架的批判: 证明了当前定义(例如跨所有位置共享的单一值置换)在异构基准测试中失效,而诸如填充(padding)之类的朴素嵌入会产生严重的假病态(例如,在 CLE4EVR 上有 90.91% 的未解释解对,而本文引入的分量级层次结构下该比例为 0%)。 * 理论基础与复杂性: 分析了十一个规则家族,确立了未解释对的比例从 0% 到 99.9999% 不等,并追踪了可证明的结构。作者提供了关于传递性条件及其失效的六个定理,精确地对布尔传递性进行了分类:当且仅当解集是一个仿射陪集(affine coset)时,自同构才能解释一切。 * 计算复杂性结果: 证明了电路给定规则的坐标对称性惰性(coordinate symmetry-inertness)是 coNP-complete 的。此外,自同构的存在性是 coNP-hard 的(通过随机归约),位于 \(\Sigma_2^p\) 中,在布尔情况下不是 \(\Sigma_2^p\)-complete 的(除非多项式层次结构坍塌),并且在单调电路上是 coNP-complete 的。 * 实证与基于模型的见解: 弱监督模型在框架标记的理论级别上隔离了所有 94 个观察到的捷径,而在 48 个传递性或 12 个类型模糊的级别上没有放置任何捷径。当在规则域和异构域上进行端到端训练时,模型产生了 20,223 个标签保持错误且零个不同轨道异常——这与传递性预测一致——而填充方法则误报了其中 78–88% 的错误。

This paper investigates reasoning shortcuts—rule-based solutions that arrive at correct predictions via unintended concepts—through the lens of value symmetries and automorphism groups (following the framework introduced by Takemura, Inoue, and Nishino).

Key findings and contributions of the work include: * Critique of Existing Frameworks: Demonstrates that current definitions (e.g., a single value permutation shared across all positions) fail across heterogeneous benchmarks, while naïve embeddings like padding produce severe false pathologies (e.g., 90.91% unexplained solution pairs on CLE4EVR compared to 0% under the componentwise hierarchy introduced here). * Theoretical Foundations & Complexity: Analyzes eleven rule families, establishing that unexplained-pair rates span from 0% to 99.9999% and track provable structure. The author provides six theorems for transitivity conditions and their failures, classifying Boolean transitivity exactly: automorphisms explain everything if and only if the solution set is an affine coset. * Computational Complexity Results: Proves that coordinate symmetry-inertness for circuit-given rules is coNP-complete. Furthermore, automorphism existence is coNP-hard (via randomized reductions), lies in \(\Sigma_2^p\), is not \(\Sigma_2^p\)-complete in the Boolean case (unless the polynomial hierarchy collapses), and is coNP-complete on monotone circuits. * Empirical & Model-Based Insights: Weakly supervised models isolate all 94 observed shortcuts at the theoretical level flagged by the framework, while placing none at the 48 transitive or 12 typed-ambiguous levels. When trained end-to-end on rule and heterogeneous domains, models yield 20,223 label-preserving errors with zero different-orbit exceptions—aligning with transitivity predictions—whereas padded methods misreport 78–88% of them.


附加元数据 (Additional Metadata)

  • 评论: 62 页,2 个图表,8 个表格。审稿中。
  • 提交历史:
  • [v1] 2026年8月11日 星期二 03:10:20 UTC
  • [v2] 2026年8月21日 星期五 18:39:52 UTC (当前版本)
  • Comments: 62 pages, 2 figures, 8 tables. Under review.
  • Submission History:
  • [v1] Tue, 11 Aug 2026 03:10:20 UTC
  • [v2] Fri, 21 Aug 2026 18:39:52 UTC (current version)