KernelGYM & Dr.Kernel

RL for GPU kernel generation done right (ICML 2026): KernelGYM is a distributed RL environment for Triton kernel generation with robust multi-turn feedback, and the paper introduces TRLOO, an unbiased turn-level advantage estimator for multi-turn RL. The resulting Dr.Kernel-14B (RL-tuned from Qwen3-14B-Base; an 8B sibling from Qwen3-8B) beats Claude-4.5-Sonnet and GPT-5 on KernelBench Level-2 speedup metrics.

Sole HKUST work from Junxian He's hkust-nlp — the systems-flavored continuation of the lab's SimpleRL zero-RL line into verifiable, performance-rewarded coding environments.