ERNIE Evolve is an automated research system that Baidu announced on the ERNIE official blog on November 9, 2026 — and the first step in Baidu's move toward recursive self-improvement. Give it a code repository, objectives, and evaluation criteria, and it can change code and run experiments on its own in an isolated environment, with independent evaluation deciding which improvements stay. Across three public AI R&D tasks, it beat the best publicly verifiable results on three metrics; applied to the in-development ERNIE-5.5-Flash, it pushed low-latency inference on the Kunlunxin P800 to 623.6% of baseline speed.
Five roles, each with its own job
The system works in two nested loops: the outer loop sets research directions, the inner loop develops and tests improvements within each direction, with several research lanes running in parallel, each maintaining its own tree of candidate code versions. Five roles split the work: the Director picks the next research direction and starting version, consulting an Experience Bank of past candidates and results plus a Knowledge Base of papers and open-source code; the Selector chooses candidates within a lane and sets the next hypothesis to test; the Executor changes code in an isolated environment; the Evaluator measures performance and checks correctness on real workloads with a fixed procedure, sitting outside the repository being optimized and invisible to any agent; the Reviewer checks whether gains come from gaming the evaluation, discarding candidates that do, and only improvements above the noise threshold are accepted.
The public-task scorecard reads in three lines:
- NanoChat AutoResearch: with a five-minute single-GPU budget, validation BPB came in at 0.8763, better than the previous best of 0.8878.
- NanoGPT Speedrun: time to target loss was 38.9 seconds and steps to target loss 2,590, both beating the previous bests of 39.9 seconds and 2,690 steps.
- SOL-ExecBench: across 235 real-model kernel optimization problems, the mean score was 0.819, slightly below the leaderboard best of 0.827.
All three tasks used the same research loop, with no tuning for any single metric, and the optimization results have been open-sourced on GitHub.
The internal rollout matters more: automated research has entered the ERNIE team's daily engineering for the first time, with changes required to pass the same load tests and correctness checks as engineer submissions. Beyond the Kunlunxin speedup, deterministic inference throughput rose 33.7% after about three days of research, decoding speed gained another 14.9%, and accuracy stayed the same. Baidu also drew a safety line: recursive self-improvement must run under human control, interruptibly, within defined objectives, permissions, and resource limits, and under ongoing safety evaluation and supervision. ERNIE Evolve has not, however, announced any public access, pricing, or API — ordinary users cannot use it yet, and it is aimed at teams working on model training, inference optimization, and GPU kernel optimization.
The next step is to extend the approach to training-system optimization first, then let model capability metrics guide automated hypothesis generation and experiment design. When ERNIE 5.0 appeared on Qianfan, it was positioned as a natively fully multimodal foundation model; ERNIE Evolve looks like the 'research engine' added to that product line.
