Baidu's ERNIE Evolve tops public bests on NanoGPT Speedrun, brings RSI to real inference optimization
xiaohu · x · 2026-11-11
At Baidu World 2026, Baidu published the tech report, blog, and downloads for ERNIE Evolve, an automated research system. It organizes research in two loops: an outer Director picks research directions from experience/knowledge bases; inner Selector/Executor agents iterate code changes across parallel lanes, gated by an independent Evaluator (outside the repo, invisible to agents) and a Reviewer that rejects off-direction work and benchmark gaming.
On public tasks it beats the best publicly verified results on NanoChat AutoResearch (0.8763 BPB) and both NanoGPT Speedrun tracks (38.9s; 2590 steps), and scores 0.819 on SOL-ExecBench vs the public best of 0.827.
Baidu then applied it to ERNIE-5.5-Flash inference optimization, with production-validated gains: 6.24x decode speed on Kunxin P800 and +33.7% GPU throughput. The company frames this as the first RSI participation in ERNIE inference optimization and calls automated research systems "the new infrastructure."
More from Infra
- CoreWeave: 2027 GPU renewals could add ~$4B annualized revenue, analyst skeptical — zephyr_z9 · 2026-11-11
- Mark Says George Hotz Was Prescient About Nvidia's Situation — mark_k · 2026-11-11
- Amazon Prime 360 uses AI to rebuild NBA plays in 3D for live, any-angle replays — import_jmr · 2026-11-11
- Altman says OpenAI aims to replace Nvidia in the medium to long term — zephyr_z9 · 2026-11-11
- Chinese University Bought Nvidia H200 Chips Despite US Blacklist — pstAsiatech · 2026-11-11
- Baidu says RSI auto-optimized ERNIE inference, 6.24x decode speedup on Kunxin P800 — xiaohu · 2026-11-11