SAGE: Experience-Learning Agent Solves 84/88 NPU Kernel Synthesis Tasks on Ascend 910C
jiqizhixin · x · 2026-11-11
A preprint from Fudan, USTC and Huawei, SAGE studies experience learning for coding agents doing Ascend NPU kernel synthesis, with correctness checks and real-device timing.
- Bottleneck: A large archive of past attempts is not automatically useful memory — irrelevant advice wastes context, and a successful final kernel doesn't reveal which earlier suggestion helped.
- Mechanism: SAGE tracks which retrieved experiences the agent adopts and assigns credit by development stage; reuse verified across operators is distilled into compact resident rules, while other records stay retrievable outside the active context.
- Results: With GLM-5.2 on Ascend 910C and 30 rounds per operator, SAGE solves 84 of 88 NPUKernelBench operators.
More from coding & agent
- How should UIs attribute actions taken by agents through a signed-in human account? — max__drake · 2026-11-11
- Fable 5.2 audits an Opus 5.5-built codebase and finds issues Opus missed — CtrlAltDwayne · 2026-11-11
- Google Antigravity ships Swarm Mode: parallel agents coordinating via shared message board — testingcatalog · 2026-11-11
- Comet Builds Docker-Based Incident Testbed to Evaluate Remediation Agents with Opik — craigsdennis · 2026-11-11
- Claude agent drives a real Leica microscope via MCP, autofocus and all — jwt0625 · 2026-11-11
- Yacine's agent talk: context moves outside the model, harnesses self-edit — yacinelearning · 2026-11-11