Amazon's CompoundingKernel Hits #1 on SOL-ExecBench with Self-Improving GPU Kernel Agents
AmazonScience · hf · 2026-11-20
Amazon presents CompoundingKernel, a self-improving agent system that writes GPU kernels by search. As of 2026-11-11 its kernels rank #1 on the public L1, L2 and Quantization boards of SOL-ExecBench (NVIDIA B200), holding #1 on L1 since 2026-10-27.
- Search from zero: agents get only each problem's PyTorch definition—no reference kernels or official scores. All 235 problems have site-verified correct kernels; on the 209 public-board problems they are 4.40x faster than the reference baseline.
- Architecture: a per-problem archive outside the model stores every correct kernel and its lineage; an LLM controller reads it, picks a parent or new design, and gives each parallel child agent one direction; correctness is decided by checks outside the model.
- At equal wall time it beat a coded UCB-style tree on 35 of 38 problems with 2.4x as many correct attempts; over a thousand agents ran concurrently, and one warm-restart fix raised evaluation rate 3.6x.
- Key finding: the rate of reliable measurements, not agent count, limits how far such search scales.
Code and kernels are open-sourced.
More from coding & agent
- Cognee open-sources agent memory platform: ontology from Postgres, 91% on LoCoMo — nikola_mr64990 · 2026-11-20
- AI writes all code to make a real robotic arm play chess, open-sourced — FinanceYF5 · 2026-11-20
- Indie dev builds a personal bot with Fable 5.2 + T3 Code, linking X and Meta data via a shared knowledge base — yihui_indie · 2026-11-20
- Clairvoyance 0.91 ships Agentic Web browser for Claude and Codex — draginol · 2026-11-20
- GPT-5.1 Checkbox Hallucinations: A Document Extraction War Story Stuck at 80% Accuracy — ComfortableThought55 · 2026-11-20
- emulate ships local OpenAI & Anthropic API emulator for deterministic zero-cost AI testing — cramforce · 2026-11-20