Forking: NanoGPT autoresearch uncovers sudden overfitting under data replay
ceciletamura · x · 2026-10-20
A new arXiv paper studies forking, a generalization failure discovered during NanoGPT autoresearch: under data replay, models with an over-encoding n-gram memory show a sharp train/validation loss split at epoch boundaries. The team reproduces it in controlled NanoGPT settings and in a DeepSeek-style model with Engram. Mechanistically, repeated updates sharpen seen continuations while suppressing unseen ones; weakly interacting context-specific subspaces from the n-gram module amplify the effect, with low-frequency contexts contributing most of the gap. Forking also appears in short-budget, heavily repeated SFT and RL-like regimes. The authors frame it alongside grokking and double descent—and warn it's an unpleasant by-product of tricks proposed by autoresearch agents, whose many seemingly useful results deserve scrutiny. Author Shaoyang Guo (Peking University) presents the work in a talk on Oct 20, 7pm PDT.
More from Research
- Stanford and Databricks use AI agents to reproduce medical findings and test analytical robustness — ChengleiSi · 2026-10-23
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23