Baidu’s ERNIE Evolve automates AI research and inference tuning

By Nathan Brooks

Baidu reported on November 9, 2026, that ERNIE Evolve delivered 33.7% higher throughput in one deterministic test of the in-development ERNIE-5.5-Flash. The system also beat public comparison results on three NanoChat and NanoGPT metrics, according to the company. The announcement covered active inference-engineering work on ERNIE-5.5-Flash.

Baidu describes ERNIE Evolve as an automated research system whose role extends beyond writing code. It selects research directions, edits AI code, runs experiments, and retains only improvements that pass separate evaluation and review. Baidu said failures and successes guide later iterations.

How ERNIE Evolve tests its own work

Given a repository, an optimization objective and evaluation criteria, ERNIE Evolve modifies code and runs experiments in isolated environments, Baidu said.

An outer-loop Director chooses a direction and starting version from an Experience Bank of previous experiments and a Knowledge Base of papers, open-source code and domain knowledge. Parallel lanes hold trees of candidate versions. Within each lane, a Selector chooses a candidate and hypothesis, while an Executor implements the change.

Acceptance rests with a fixed Evaluator and a separate Reviewer, Baidu said. The Evaluator sits outside the repository, remains invisible to agents, excludes out-of-scope changes and requires gains above a noise threshold.

The Reviewer checks research-direction compliance and discards evaluation-loophole exploits. Off-direction candidates cannot become the best version.

Public benchmarks and released code

Against the public comparison results used in its evaluation, Baidu reported:

• NanoChat AutoResearch: In five-minute training on one B200, ERNIE Evolve achieved 0.8763 validation bits per byte, where lower is better. The public comparison averaged 0.887791 across 10 seeds.
• NanoGPT Speedrun Track 1: On eight H100 GPUs, it reached 3.28 FineWeb validation loss in 38.92 seconds over 18 runs, versus 39.9 seconds for the public comparison.
• NanoGPT Speedrun Track 3: With fixed data, batch size and architecture, it took 2,590 steps over eight runs to reach the 3.28 FineWeb validation loss, versus 2,690 steps over eight runs publicly.
• SOL-ExecBench: Across 235 kernel problems, it scored 0.819 on the mean speed-of-light measure against an analytical B200 roofline, where higher is better. The leading public score was 0.827.

Baidu qualified that result. ERNIE Evolve did not lead every listed metric, although the company placed it ahead of Recursive and Hyra-1.0 on SOL-ExecBench.

The winning NanoChat version added two prediction branches alongside the neural network, Baidu said. One used n-gram statistics from the training corpus; the other retrieved matching documents and predicted from their continuations. A learned gate adjusted the mixture using each branch’s performance at earlier token positions.

Baidu’s results repository publishes final code and measurement logs. Executed code appears in each log or is identified by its SHA-256. The repository lists seven SOL-ExecBench kernel artifacts scoring from 0.6669 to 0.9920. It does not redistribute NVIDIA’s task definitions or workload data.

Inference optimization

Baidu applied ERNIE Evolve to low-latency inference for ERNIE-5.5-Flash on Kunlunxin P800, deterministic inference throughput on GPUs and single-request decoding speed. Most baselines had already been optimized by engineers, and each automated change had to pass the same load tests and correctness checks as engineer-submitted work.

On Kunlunxin P800, four research rounds raised generation speed to 623.6% of baseline—more than six times the starting rate—while passing correctness checks, Baidu reported. The model and deployment setup remained unchanged.

Early rounds enabled fused manifold-constrained hyper-connections and speculative decoding, then reduced host-side scheduling and synchronization overhead. A fourth-round dispatch fix enabled DeepEP’s low-latency path.

For reinforcement-learning training, Baidu’s deterministic task required generated tokens and log probabilities to be bitwise identical regardless of batch composition and to match full-sequence recomputation. SGLang’s deterministic mode met those requirements by fixing reduction order, but the approach limited GPU utilization.

After roughly three days of work, Baidu reported a 26% cut in average time per output token and a 12% reduction in time to first token, with determinism preserved. The largest individual gain came from replacing deterministic sampling, which consumed about 14% of GPU time per step, with a Triton kernel that processes vocabulary chunks in parallel.

For single-request GPU decoding, Baidu reported a 14.9% speed increase over an already engineer-tuned service, with unchanged accuracy. Three of seven research directions produced gains.

The winning changes used native FP16 instructions for FP8 matrix multiplications with fewer than eight input rows and replaced sequential expert scans with a two-level mixture-of-experts scheduler, Baidu said.

The tests also exposed tighter module and determinism constraints, Baidu said. Measurement noise was comparable to individual gains, and the environments spanned inference engines, kernel libraries, hardware plugins and service startup scripts. Some useful changes fell below the noise floor, so candidates had to prove value in end-to-end load tests rather than isolated kernel measurements.

What comes next

Baidu plans to apply the loop to training-system optimization. It then wants model-capability metrics to guide automated hypothesis generation, experiment design and validation. The longer-term goal is for AI to identify capability gaps and generate and refine evaluation tasks.

The company says recursive self-improvement must remain human-controlled and interruptible, within explicit boundaries for goals, permissions and resources. Continuous safety evaluation and human oversight remain central as the system moves beyond fixed engineering metrics.

What does this change for AI model builders?

For practitioners, ERNIE Evolve offers a concrete operating model: isolate experiments, retain an experience record, and require evaluation and review before accepting code. Baidu’s released logs make the process inspectable, while its SOL-ExecBench score shows where the automated system still trails the public leader.

If Baidu extends that loop into training and benchmark design, its value will depend less on generating ideas than on keeping evaluations independent enough to catch loopholes before they shape the next model.