Add new paper: Understanding Knowledge Hijack Mechanism in In-context Learning through Associative Memory

[wyzh0912]

Title: Understanding Knowledge Hijack Mechanism in In-context Learning through Associative Memory
Head: Induction Head
Published: arxiv
Summary:

• Innovation: Investigated how the model prioritizes in-context knowledge versus global knowledge when generating outputs, using both theoretical analysis and experimental evaluation.
• Tasks: Analyzed two key aspects of induction heads: the impact of positional encoding on the oversight, and the contributions of in-context information and pretrained knowledge during inference.
• Significant Result: Induction head learned by a transformer using RPE can avoid oversight of in-context knowledge within the framework of associative memory