Neel Nanda et al. reverse-engineer grokking, showing it emerges in three continuous phases
burny_tech · x · 2026-11-01
In this mechanistic interpretability paper, Neel Nanda and colleagues fully reverse-engineer the algorithm a small transformer learns on modular addition: it uses discrete Fourier transforms and trig identities to turn addition into rotation on a circle, verified via activation/weight analysis and Fourier-space ablations. They define continuous progress measures that split training into memorization, circuit formation, and cleanup phases — showing grokking is not a sudden shift but the gradual amplification of structured mechanisms followed by removal of memorizing components.
More from Models
- DeepSeek-V4.1-Flash uncensored GGUF quant hits Hugging Face trending — AtomicChat · 2026-11-01
- The cheapest path to learn RL: David Silver lectures, DQN, PPO, GRPO — all on a laptop — cneuralnetwork · 2026-11-01
- Reddit bench: EPYC + multi-3090 rig runs Qwen Q8 at 94 tok/s with near-linear scaling to 8 streams — sleight42 · 2026-11-01
- User shares ChatGPT's own confession: answers differ by person and reflect its programmers' choices — JMD_923 · 2026-11-01
- GPT's forced 'zoomer' persona annoys users, custom instructions don't fix it — Aizkmusic · 2026-11-01
- From-scratch CUDA engine runs full GLM-5.3 on 3-6 DGX Sparks; steering vectors flip output language — burny_tech · 2026-11-01