发现大模型靠捷径提前决策，提出方法可定位并抑制这些捷径。
Short-circuiting Shortcuts: Mechanistic Investigation of Shortcuts in Text Classification
- 用演员名做可控捷径，揭示注意力头如何提前锁定标签。
- 识别出特定注意力头在处理完整输入前就做出判断。
- 新方法HTA能精准定位捷径，适合改进模型公平性与鲁棒性
语言模型的成功常源于对虚假相关性的依赖（捷径）。以往研究聚焦于识别影响预测的输入元素，本文则深入探究捷径在模型决策机制中的实际处理过程。以电影评论中的演员名为可控捷径，利用机制可解释性方法，我们识别出特定注意力头会聚焦于这些捷径，并在处理完整输入前即引导模型趋向某标签，导致过早决策而跳过上下文分析。基于此，我们提出头部级标记归因（HTA），可将中间决策回溯至输入标记。实验表明，HTA能有效检测大模型中的捷径，并通过选择性关闭相关注意力头实现针对性缓解。
原文摘要 · Abstract (English)
Reliance on spurious correlations (shortcuts) has been shown to underlie many of the successes of language models. Previous work focused on identifying the input elements that impact prediction. We investigate how shortcuts are actually processed within the model's decision-making mechanism. We use actor names in movie reviews as controllable shortcuts with known impact on the outcome. We use mechanistic interpretability methods and identify specific attention heads that focus on shortcuts. These heads gear the model towards a label before processing the complete input, effectively making premature decisions that bypass contextual analysis. Based on these findings, we introduce Head-based Token Attribution (HTA), which traces intermediate decisions back to input tokens. We show that HTA is effective in detecting shortcuts in LLMs and enables targeted mitigation by selectively deactivating shortcut-related attention heads.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。