What it is
Searches neural architectures in open program space by having an LLM rewrite full PyTorch blocks and evolution decide which few get trained. Discover new model architectures by evolving executable architecture programs instead of picking from a fixed search space, with LLM-generated rewrites, surrogate screening, and a browsable 171-model zoo.
Key results
- Screens thousands of LLM proposals per iteration; trains only <=16
- Ran 10 iterations and trained 151 new architectures in ~132 V100 GPU-days
- Includes a 171-architecture zoo with code, rationale, diagrams, and metrics
Shows that LLM-guided evolution can search open architecture program space and discover competitive native spiking sequence models.