Kernaut: coding agents discover ML kernels automatically as 22-58% of LLM-generated ones fail at scale
Cohere · youtube · 2026-11-14
PhD candidate Richard Cornelius Suwandi (CUHK-Shenzhen) presented Kernel Autoresearch (Kernaut) at Cohere Labs Open Science Community.
The problem: kernels encode inductive biases of many ML models, but automating their design involves a trade-off. Fixed grammars guarantee validity but limit the search space; free-form LLM programs remove that limit but are brittle — in tests, 22–58% of LLM-generated kernels that pass simple numerical checks fail at different scales or dimensions.
Method:
- Coding agents write kernels as programs, with construction contracts ensuring every accepted kernel is valid
- A quality-diversity archive retains strong kernels with different behaviors
- Novelty screening pushes agents toward genuinely new ideas
Results: discovered kernels transfer to unseen tasks (black-box optimization, enzyme kinetics) without further LLM calls, and remain interpretable programs humans can refine.
Suwandi co-developed OpenEvolve, an open-source evolutionary coding agent inspired by DeepMind's AlphaEvolve, and co-founded AIDDA. Paper: arxiv.org/abs/2610.10394.
More from coding & agent
- Val Town kills *.val.run links on Jan 10, 2027; adds MCP tools to set up redirects — threepointone · 2026-11-14
- Transcript Testing: Apache DataFusion maintainer says AI-era tests should be written for humans to read — mgill25 · 2026-11-14
- Scope launches Competitors to track how AI agents evaluate and bypass your product — RaphaelDabadie · 2026-11-14
- Open-source IMAP MCP lets Claude read and send email from any mailbox, with OAuth and hard safety rails — Ordinary_Ad_404 · 2026-11-14
- Claude Code 2.1.325 is about to be released — ClaudeCodeLog · 2026-11-14
- Grok Bot joins Google Meet, Zoom and Teams as a working meeting teammate — XFreeze · 2026-11-14