---
id: 20260409-T0-20
title: "潜在泛化错觉：双向性与反转诅咒"
title_en: "Bilingual Supervision Mitigates Reversal Curse in AI Models"
url: https://ai.daily.yangsir.net/daily/20260409-T0-20
issue_date: 2026-04-09
publish_date: 2026-04-08T04:00:00.000Z
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2604.04943
---

# 潜在泛化错觉：双向性与反转诅咒

该研究探讨了自回归语言模型在反向事实检索中的失败现象，即“反转诅咒”。例如，模型在训练时学习“A > B”的事实，但在测试时无法正确处理“B < A”的表述。最新研究表明，采用双向监督（如双向语言模型）的目标函数可以缓解这一问题。通过实验验证，双向模型在反向任务上的准确率显著提升，相比传统单向模型提高了约30%。这一发现对提升模型的逻辑推理能力和泛化性能具有重要意义，为未来语言模型的设计提供了新方向。

## English Version

**Bilingual Supervision Mitigates Reversal Curse in AI Models**

A study investigates a failure in autoregressive language models known as the 'reversal curse,' where a model trained on facts like 'A > B' fails to process their reverse, 'B < A.' The research demonstrates that using bidirectional supervision in the objective function, such as with bidirectional language models, can effectively address this issue. Experimental validation shows that bidirectional models achieve a significant performance boost, improving accuracy on reverse tasks by approximately 30% compared to traditional unidirectional models. This finding is crucial for enhancing logical reasoning and generalization capabilities, offering new directions for the future design of language models.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2604.04943)

**详情页**：https://ai.daily.yangsir.net/daily/20260409-T0-20

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*