Bilevel Autoresearch, by independent researchers Yaonan Qu and Meng Lu (arXiv), treats Karpathy’s autoresearch loop as itself a research target: the inner loop proposes and tests changes to a GPT pretraining run, while the outer loop diagnoses the inner loop’s search behavior and writes new mechanisms drawn from bandits, combinatorial optimization and design of experiments. On Karpathy’s GPT pretraining benchmark the outer loop reports a 5x improvement over the inner loop alone (-0.045 vs -0.009 val_bpb) with no stronger meta-level model. It enters the corpus in the March 31 development as the point where the recursion stops needing a smarter supervisor, and is the immediate predecessor of Apple’s self-distillation result.
AISystem
Bilevel Autoresearch
A two-level autoresearch framework in which an outer loop reads the inner hyperparameter-search loop's code and traces and injects new Python search mechanisms at runtime, both loops running on the same model.
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to March 31, 2026 (theinnermostloop.substack.com)
Development timeline
Newest reporting dates first. Reporting dates may differ from event dates.
Knowledge graphopen in browser →
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.