← back to paper

Review history

arxiv: 2607.20908 · 2 revisions

Multi-turn RL with Structural and Performance Aware Rewards for CUDA Kernel Generation

1. 2026-08-15 CONDITIONAL MODERATE v1.4.0-alltime-deepseek-medium novelty 6.0
   100478 ms 15072 in 12422 out 2026-08-15T15:30:47.658597+00:00
2. 2026-08-01 CONDITIONAL LOW v1.3.0-alltime-deepseek novelty 6.0
   116072 ms 15030 in 13201 out 2026-08-01T09:01:31.394744+00:00