潜在泛化的幻觉：双向性与反转诅咒
The Illusion of Latent Generalization: Bi-directionality and the Reversal Curse
AI总结：
本文研究了自回归语言模型在反向顺序检索中的失败现象，比较了双向监督目标与传统掩码语言模型在四个基准测试中的表现，发现反向准确性依赖于明确将源实体作为预测目标的训练信号，而非单一方向无关的表示。
摘要：
反转诅咒描述了自回归语言模型在反向顺序检索中的失败（例如，训练于"A > B"但无法处理"B < A"）。近期研究表明，具有双向监督的目标（如双向注意力或基于掩码的解码器-only模型重构）可以缓解此问题。我们扩展了此评估，包括传统掩码语言模型（MLM）目标，并在四个反转基准上与解码器-only掩码训练进行比较，随后提供了一项最小机制研究，探讨这些目标如何成功。我们发现，反向准确性需要训练信号明确使源实体成为预测目标，并发现很少有证据表明成功对应于单一方向无关的事实表示。相反，表示距离和线性探针表明，正向和反向方向被存储为不同的条目，MLM与解码器-only掩码训练具有不同的索引几何结构。我们的结果警告，目标层面的“修复”可以改善反向行为，而无需诱导人们可能从统一概念中期望的潜在泛化类型。
英文摘要：
The reversal curse describes a failure of autoregressive language models to retrieve a fact in reverse order (e.g., training on ``$A > B$'' but failing on ``$B < A$''). Recent work shows that objectives with bidirectional supervision (e.g., bidirectional attention or masking-based reconstruction for decoder-only models) can mitigate the reversal curse. We extend this evaluation to include a vanilla masked language modeling (MLM) objective and compare it to decoder-only masking-based training across four reversal benchmarks and then provide a minimal mechanistic study of \emph{how} these objectives succeed. We show that reversal accuracy requires training signal that explicitly makes the source entity a prediction target, and we find little evidence that success corresponds to a single direction-agnostic representation of a fact. Instead, representation distances and linear probes are consistent with storing forward and reverse directions as distinct entries, with different indexing geometry for MLM versus decoder-only masking-based training. Our results caution that objective-level ``fixes'' can improve reversal behavior without necessarily inducing the kind of latent generalization one might expect from a unified concept.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
