Reproduce “Identifying Semantic Induction Heads to Understand In-Context Learning” on OpenResearch

About this paper

Jie Ren, Qipeng Guo, Hang Yan, Dongrui Liu, Quanshi Zhang, Xipeng Qiu, Dahua Lin

Researchers identified "semantic induction heads" in Large Language Models (LLMs), demonstrating that specific attention heads encode complex syntactic and semantic relationships like subject-predicate dependencies and knowledge graph relations. This work correlates the emergence of these specialized heads with the development of progressive levels of in-context learning, particularly for advanced pattern discovery, providing a mechanistic link beyond simple token copying.

Key insights

• Attention heads in LLMs can encode complex linguistic structures, including specific syntactic dependencies and various knowledge graph-derived semantic relationships, extending beyond basic token copying.
• In-Context Learning in LLMs develops progressively, starting with simple loss reduction, advancing to format compliance, and culminating in the more complex ability for pattern discovery.
• The formation and strengthening of semantic induction heads are strongly correlated with the emergence of advanced ICL capabilities, specifically the pattern discovery level, indicating a direct mechanistic link.

Summary by alphaXiv.