Baidu's ERNIE Evolve automates AI research, boosts ERNIE 5.5 Flash inference speed 6x
BaiduResearch · x · 2026-11-10
Baidu unveils ERNIE Evolve, an automated research system aimed at recursive self-improvement
Baidu detailed ERNIE Evolve, a system that autonomously proposes research directions, runs experiments in isolated environments, validates results, and accumulates reusable experience — a step toward recursive self-improvement (RSI) in real AI development.
Architecture
- Dual-loop design: an outer Director decides what to research; an inner Selector + Executor iterate within parallel research lanes, each maintaining a candidate tree
- Independent gating: the Evaluator lives outside the repo being optimized and is invisible to agents; improvements must exceed noise thresholds, and a Reviewer rejects off-direction changes or evaluation exploit
- Memory: an Experience Bank and Knowledge Base feed findings into future research rounds
Results
- Outperformed baselines on NanoChat AutoResearch, NanoGPT Speedrun, and SOL-ExecBench, beating the best publicly verified results on the first two
- Applied to ERNIE 5.5 Flash inference optimization with production-verified gains: 623.6% of baseline generation speed on Kunlunxin P800, 133.7% throughput for deterministic inference while preserving bitwise determinism, plus improved MoE scheduling and GPU decoding
Baidu frames such AI-native research infrastructure as the new infrastructure, arguing RSI at scale will accelerate intelligence gains unprecedentedly.
Related event: Baidu Releases Open-Source ERNIE Evolve Automated Research System(6 posts)→
More from Infra
- simdjson disables 256-bit LASX on Loongson 3A6000: 128-bit LSX is 6.7% faster — lemire · 2026-11-10
- Vera launches gyms, benchmarks and agents for cloud infrastructure operations — tianyin_xu · 2026-11-10
- NVIDIA paper: masking ~10 'intrusion heads' recovers 57% of distractor accuracy loss — rohanpaul_ai · 2026-11-10
- Memory pricing itself out of demand: PC BOM share hits 35-38%, shipments plunge — tengyanAI · 2026-11-10
- Two engineers burned $500K in API usage in one month — $6M annualized — bingxu_ · 2026-11-10
- Flide: exact state forking speeds decision-model agents up to 9.5x — NationalUniversityofSingapore · 2026-11-10