跳转至

推理捷径与价值对称性：对称性允许什么、架构实现什么以及优化选择什么

文章背景与核心概要

本文深入探讨了“推理捷径”（Reasoning Shortcuts）——即神经符号系统通过非本意、非逻辑的概念得出正确预测的现象。作者对 Takemura、Inoue 和 Nishino 提出的近期框架进行了批判，证明其关于价值重新标记（value relabeling）的核心假设在标准基准测试中无法成立。通过对 11 个规则家族的严格重新评估，该论文引入了一个新的对称性层次结构，并提供了形式化证明（包括“自由槽引理”），以解释某些捷径为何会出现。研究最后区分了对称性在理论上允许的内容与弱监督模型中优化过程实际选择的内容。

This paper investigates "reasoning shortcuts"—instances where neurosymbolic systems arrive at correct predictions via unintended, non-logical concepts. The author critiques a recent framework by Takemura, Inoue, and Nishino, demonstrating that its core assumptions regarding value relabeling fail to hold across standard benchmarks. Through a rigorous re-evaluation of eleven rule families, the paper introduces a new hierarchy of symmetry and provides formal proofs—including the "Free Slot Lemma"—to explain why certain shortcuts emerge. The research concludes by distinguishing between what symmetry theoretically permits and what optimization processes actually select in weakly supervised models.

摘要

Summary

本文探讨了“推理捷径”——神经符号系统通过非预期、非逻辑概念得出正确预测的实例。作者对 Takemura、Inoue 和 Nishino 最近提出的框架进行了批判，证明其关于价值重新标记的核心假设在各项标准基准测试中均无法成立。通过对 11 个规则家族的严格重新评估，本文引入了一个新的对称性层次结构，并提供了形式化证明（包括“自由槽引理”），以解释某些捷径产生的原因。研究最后区分了对称性理论上允许的内容与弱监督模型中优化过程实际选择的内容。

This paper investigates "reasoning shortcuts"—instances where neurosymbolic systems arrive at correct predictions via unintended, non-logical concepts. The author critiques a recent framework by Takemura, Inoue, and Nishino, demonstrating that its core assumptions regarding value relabeling fail to hold across standard benchmarks. Through a rigorous re-evaluation of eleven rule families, the paper introduces a new hierarchy of symmetry and provides formal proofs—including the "Free Slot Lemma"—to explain why certain shortcuts emerge. The research concludes by distinguishing between what symmetry theoretically permits and what optimization processes actually select in weakly supervised models.

核心发现

Key Findings

1. 现有框架的批判

1. Critique of Existing Frameworks

作者证明了在每个位置应用共享置换的流行方法在经验上存在缺陷。当应用于异构基准测试时，该方法会产生“自信的错误病理现象”（confident false pathology），在 CLE4EVR 数据集上报告的未解释解率高达 90.91%，而作者提出的层次结构下该比例为 0%。

The author demonstrates that the prevailing approach of applying a shared permutation at every position is empirically flawed. When applied to heterogeneous benchmarks, the methodology produces "confident false pathology," with reported unexplained solution rates as high as 90.91% on the CLE4EVR dataset, compared to 0% under the author's proposed hierarchy.

2. 理论贡献

2. Theoretical Contributions

  • 对称性与传递性： 本文确立了六个定理，为对称性的传递性提供了充分条件。
  • 计算复杂性：
    • 判定一个坐标的对称惰性（symmetry-inertness）是 coNP-complete 的。
    • 非平凡自同构的存在性是 coNP-hard 的，并且位于 \(\Sigma_2^p\) 中.
    • 对于布尔情况，作者给出了一个完整的分类：自同构解释了解集当且仅当该集是一个仿射陪集（affine coset）。
  • 自由槽引理： 纯粹基于语法开发的形式化工具，用于证明系统（如 Kandinsky 基准测试）中的病理现象。
  • Symmetry and Transitivity: The paper establishes six theorems providing sufficient conditions for the transitivity of symmetry.
  • Computational Complexity:
    • Deciding the symmetry-inertness of a coordinate is coNP-complete.
    • The existence of nontrivial automorphisms is coNP-hard and resides in \(\Sigma_2^p\).
    • For Boolean cases, the author provides a complete classification: automorphisms explain the solution set if and only if the set is an affine coset.
  • The Free Slot Lemma: A formal tool developed to certify pathology in systems (such as the Kandinsky benchmarks) based purely on syntax.

3. 经验观察

3. Empirical Observations

通过根据 15 个预先指定的预测来衡量 11 个规则家族，本研究发现未解释对（unexplained-pair）的比率差异巨大（从 0% 到 99.9999%）。该研究突出了理论对称性与实际模型行为之间的明显分歧： * 弱监督模型一致倾向于分量理论（componentwise theory）所标出的捷径。 * 模型避开了理论证实具有传递性的领域的捷径，这表明优化过程在对称可能性的空间中主动选择了特定的路径。

By measuring eleven rule families against fifteen pre-specified predictions, the study found that unexplained-pair rates vary drastically (0% to 99.9999%). The research highlights a clear divergence between theoretical symmetry and practical model behavior: * Weakly supervised models consistently gravitate toward shortcuts flagged by the componentwise theory. * The models avoid shortcuts in areas where the theory certifies transitivity, suggesting that optimization processes actively select specific paths within the space of symmetric possibilities.

元数据与访问路径

Metadata & Access