What was announced
On November 9, 2026, Baidu introduced ERNIE Evolve, an automated research system. Given a repository, an objective and evaluation criteria, it picks research directions, edits code and runs experiments. Only changes that pass a hidden evaluator and a separate reviewer are kept.
Baidu framed it as a practical step toward recursive self-improvement (RSI), where AI systems improve the software that builds and serves AI.
How it works
The system runs two nested loops: the outer loop decides what to investigate, the inner loop develops and tests improvements within each direction. Work is split across four roles:
- Director: picks the next research direction and a starting code version, reading an experience bank of past results and a knowledge base of papers and code.
- Selector: chooses a candidate from a version tree and names the next hypothesis.
- Executor: implements the plan in an isolated environment.
- Evaluator and reviewer: the evaluator sits outside the repository and is invisible to agents; the reviewer checks whether a change follows the intended direction or exploits an evaluation loophole.
Failed experiments are recorded so later rounds can avoid repeating them.
Results
Baidu tested the system on NanoChat AutoResearch, NanoGPT Speedrun and SOL-ExecBench. According to the company:
- NanoChat reached 0.8763 validation BPB.
- NanoGPT Speedrun Track 1 hit the target loss in 38.9 seconds, versus the 39.9-second record.
- Track 3 reached the target in 2,590 steps, versus 2,690.
- SOL-ExecBench scored 0.819 mean SOL, short of the 0.827 leaderboard best.
On the production side, inference for ERNIE-5.5-Flash was optimized. Baidu reports generation speed on Kunlunxin P800 reached 623.6% of baseline after four research rounds, with +33.7% deterministic throughput and +14.9% decode speed.
Caveats
All figures are self-reported. The system has not been released and Baidu has not disclosed which LLMs power its agents. Margins on public benchmarks are small in places. The comparison table also lists Tencent Hunyuan's Recursive and Hyra-1.0 and rekursiv.ai's (auto)²-research, but their Speedrun times were set against an older record, so they are not like-for-like.
