Execution-Feedback on 兜忆轩 | Cedric's Compiler Notes

Execution-Feedback on 兜忆轩 | Cedric's Compiler Noteshttps://douyixuan.github.io/tags/execution-feedback/Recent content in Execution-Feedback on 兜忆轩 | Cedric's Compiler Notes兜忆轩 | Cedric's Compiler Noteshttps://douyixuan.github.io/images/papermod-cover.pnghttps://douyixuan.github.io/images/papermod-cover.pngHugo -- 0.146.0en-usThu, 08 Oct 2026 12:11:00 +0800023 · 从两百轮里挑最快的，算稳定胜过编译器吗？https://douyixuan.github.io/posts/ai-silicon-papers/2026-10-08/Thu, 08 Oct 2026 12:11:00 +0800https://douyixuan.github.io/posts/ai-silicon-papers/2026-10-08/CUDA Agent 在执行反馈环境中训练 CUDA 优化策略，并在 KernelBench 上显著超过 torch.compile；但评测允许每题最多 200 轮并选取轨迹中的最快正确解，策略能力与搜索预算仍需分开理解。