Marktechpost (@Marktechpost)T2·78 pts
- Two nested loops: a Director picks directions and parallel lanes evolve code trees, while a hidden Evaluator scores candidates and a Reviewer discards eval-gaming ones
- Hit 0.8763 BPB on NanoChat AutoResearch (best public 0.8878) and 38.9 s on NanoGPT Speedrun Track 1
- Pushed Kunlunxin P800 inference to 623.6% of baseline and deterministic GPU throughput up 33.7%, though SOL-ExecBench 0.819 trails the 0.827 leader