LLM Neuroanatomy: How I Topped the AI Leaderboard Without Changing a Single Weight
Summary
This article is a deep dive into LLM neuroanatomy, describing how duplicating middle Transformer layers (i, j configurations) can boost performance without changing weights. It introduces the idea of functional circuits within layers, presents a brain-scanner style methodology, and shares leaderboard results (RYS-XLarge) and insights on how early, middle, and late layers contribute to encoding, reasoning, and decoding.