Kernel Autoresearch for Open-Ended Model Discovery
Kernaut uses coding agents plus validity contracts to search for new kernel programs without locking the search to a fixed grammar.
The paper says LLM-generated kernels can pass random numerical checks and still fail at other scales or dimensions. Kernaut accepts only kernels that satisfy construction contracts, then keeps diverse high performers in an archive and pushes agents toward novel behavior. In the reported tests, discovered kernels generalized to unseen black-box optimization and enzyme-kinetic tasks, beating several tuned or meta-learned baselines. The authors also report that a human-refined discovered kernel further cut held-out predictive error and optimization regret.
ArXiv · AI/CL/LG's note
The paper says LLM-generated kernels can pass random numerical checks and still fail at other scales or dimensions. Kernaut accepts only kernels that satisfy construction contracts, then keeps diverse high performers in an archive and pushes agents toward novel behavior. In the reported tests, discovered kernels generalized to unseen black-box optimization and enzyme-kinetic tasks, beating several tuned or meta-learned baselines. The authors also report that a human-refined discovered kernel further cut held-out predictive error and optimization regret.
ArXiv · AI/CL/LG's note
score 5