Baidu ERNIE Evolve: an AI research system that rewrites its own code and beat NanoChat records

The loop: two nested searches, one hidden judge#

ERNIE Evolve starts with three human-supplied inputs: a code repository, an optimization objective and evaluation criteria. From there it runs on its own. Two nested loops do the work. An outer-loop Director picks the next research direction and a starting code version, reading an Experience Bank of past experiments and a Knowledge Base of papers, open-source code and domain knowledge. Several research lanes run in parallel, each holding a tree of candidate code versions. Inside a lane, a Selector picks a candidate and a hypothesis, and an Executor implements the change in an isolated environment.

The part Baidu stresses most is the gate. A fixed Evaluator sits outside the repository, invisible to the agents, and a change only counts if its gain clears a noise threshold. A separate Reviewer checks whether the change follows the intended research direction or exploits an evaluation loophole. Off-direction candidates cannot become the best version. Failed experiments are recorded in the Experience Bank, so later rounds do not repeat them. Baidu has not said which language models power the agents.

It is the same shape as other automated-research systems — Google’s AlphaEvolve runs a similar evolve-and-test loop on code — but ERNIE Evolve is the first to publish a full ledger of wins and losses against public benchmarks and then against a production model. Google is pairing AlphaEvolve with its HPC push, and the two systems are now the clearest public comparison points for machine-led R&D.

The scoreboard#

Baidu tested the loop on three public benchmark suites, then turned it on real engineering: inference for ERNIE-5.5-Flash, a model still in development whose open-weights base checkpoint went public on November 5. The public numbers:

Three wins, one loss — and the wins are narrow. On NanoChat the system added two prediction branches beside the neural network: one using n-gram statistics from the training corpus, the other retrieving matching documents and predicting from their continuations, mixed by a learned gate. On Speedrun Track 3 it decomposed weights into direction and magnitude and added EMA-Nesterov extrapolation. Baidu published the final code and measurement logs in a results repository, with executed code embedded in each log or identified by its SHA-256 hash.

The production numbers are bigger and harder to check. On Baidu’s Kunlunxin P800 accelerator, four research rounds took ERNIE-5.5-Flash generation speed to 623.6% of baseline over about 168 hours, with the model and deployment setup unchanged. Round one enabled fused manifold-constrained hyper-connections and speculative decoding; round two cut host-side scheduling overhead; a fourth-round dispatch fix engaged DeepEP’s low-latency path. In a deterministic-inference task for reinforcement-learning training — where generated tokens and log probabilities must be bitwise identical regardless of batch composition — throughput rose 33.7% in about three days, time per output token fell 26% and time to first token fell 12%. The largest single gain came from replacing deterministic sampling, which had consumed about 14% of GPU time per step, with a Triton kernel that processes vocabulary chunks in parallel. Single-request decoding on GPUs rose 14.9% with accuracy unchanged.

The catch#

Every number above is Baidu’s. The public benchmarks are reproducible in principle — the code and logs are on GitHub — but the production gains run on Kunlunxin P800 hardware and an unreleased model, so nobody outside Baidu can re-run them. The evaluator that grades the work is also Baidu’s, which is exactly the conflict the hidden-evaluator design is supposed to solve: the system cannot see the test, but the company that built the test also built the system and wrote the report.

The margins cut both ways. A 1.3% edge on NanoChat and a 2.5% edge on Speedrun Track 1 are real but small, and on SOL-ExecBench the system finished behind the leader. Some useful changes were smaller than measurement noise, Baidu says, and had to prove their value in end-to-end load tests rather than isolated kernel measurements. And the system itself stays closed: no orchestrator, no prompts, no Experience Bank — only the artifacts it produced.

Baidu’s own framing is cautious. The company says recursive self-improvement must remain human-controlled and interruptible, within explicit boundaries for goals, permissions and resources, with continuous safety evaluation and human oversight. Its next steps are to apply the loop to training-system optimization, then to use model-capability metrics to guide automated hypothesis generation, experiment design and validation — and, longer term, for AI to identify capability gaps and generate its own evaluation tasks. That last step is where the oversight question stops being theoretical. For now, the ledger is the news: an AI research loop with published numbers, a named failure, and a 6.2× speedup it can point to.

Sources: Baidu ERNIE Research, “ERNIE Evolve: Toward Automated Research for Real-World AI R&D” (November 9, 2026); Unite.AI, “Baidu’s ERNIE Evolve Automates AI Research and Inference Optimization” (November 9, 2026), here; MarkTechPost, “Baidu’s ERNIE Evolve Automated Research System Beats NanoChat and NanoGPT Records” (November 9, 2026), here.