Research framework
Autonomous Intelligent Research for agentic neural architecture search.
AIR is Rkive's framework for agentic neural architecture search (NAS). Three agents — researcher, supervisor, analyst — operate with persistent memory, intelligent budget allocation, and structured experiment logs. Each cycle inherits prior learning through MAIC-RL, a memory-augmented in-context reinforcement learning protocol. AIR extends Karpathy's autoresearch into structured multi-agent search with taxonomic legibility.
AIR works across experiments, steering cycles, and broader sweeps, compounding learning from failures rather than running blind trial-and-error. The supervisor reallocates effort as results arrive. The analyst synthesizes patterns across runs. Custom activation functions and loss formulations are invented at runtime and AST-validated before injection.
AIR powers an open record submission at 1.1396 val_bpb in the OpenAI Parameter Golf competition, ahead of the merged leaderboard at 1.1428 (March 2026). Across 70+ experiments and five sweep cycles, the framework produced novel techniques including gradient-guided adaptive quantization and runtime function invention, alongside reusable search traces for meta-model training.
Experiment, cycle, and sweep
Budget is reallocated across experiments, cycles, and sweeps as results come in — not treated as isolated runs. The supervisor prunes stale approaches and concentrates effort where signal is strongest.
Architecture, technique, and hyperparameter
Architecture, activation function, training recipe, and hyperparameter choices are explored together. In constrained optimization, gains usually come from interactions between variables, not from tuning single knobs.
Budget allocation and runtime function writing
Budget is adjusted under noisy reward signals. The supervisor can define new activation functions and loss formulations at runtime, reshaping the search space as promising patterns emerge.
AIR emits structured experiment records, steering decisions, and analyst outputs that compound across model families rather than being discarded after each run.
Search becomes cumulative: the framework retains what worked, what failed, and why.