A research loop writes the strategies for its own outer loop

The recursion stops needing a smarter supervisor. Bilevel Autoresearch (Qu and Lu, arXiv) nests two loops: the inner one is the hyperparameter-search loop Karpathy popularized as autoresearch, and the outer one reads the inner loop’s code and traces, identifies bottlenecks and injects new Python search mechanisms at runtime; 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), instantiating mechanisms from bandits, combinatorial optimization and design of experiments that no human specified. Both loops use the same model, which is the point: the gain comes from the architecture rather than a stronger supervisor. Meta’s AIRA2 (arXiv) attacks the other side of the problem, replacing synchronous single-GPU execution with an asynchronous multi-GPU worker pool, adding a hidden consistent evaluation protocol to close the generalization gap and using ReAct agents that debug interactively, and reaches an 81.5% mean percentile rank on MLE-bench-30 at 24 hours against a 72.7% baseline. The pair extend Meta’s hyperagents of the previous week and Karpathy’s 650-experiment autoresearch run, and are overtaken within days by Apple’s self-distillation result, which removes the search loop entirely.