News

Baidu details ERNIE Evolve automated research system

Baidu says ERNIE Evolve uses nested generate, evaluate and review loops to automate code experiments, moving from public optimization tasks into inference work for the in-development ERNIE 5.5 Flash.

D
Nov 20, 2026 · 3 min read

Baidu has published ERNIE Evolve, an automated research system that proposes code changes, runs experiments and keeps only candidates that pass evaluation and review. The company says it tested the approach on three public optimization tasks before bringing it into day-to-day inference work for the in-development ERNIE 5.5 Flash.

Baidu has not provided independent validation of the internal performance figures or confirmed that the resulting changes entered production. The company says proposed changes for ERNIE 5.5 Flash had to pass the same load tests and correctness checks as engineer-submitted work. They also had to stay within specified modules and preserve accuracy, latency and, where required, deterministic output.

The system runs two nested loops. In the outer loop, a director chooses research directions and assigns parallel lanes. Within each lane, the inner loop selects an earlier candidate, states a hypothesis, edits code in an isolated environment, then sends the result to a fixed evaluator and a separate reviewer.

According to Baidu, the evaluator sits outside the repository being optimized and is hidden from the agents. It measures performance and correctness on the actual workload, rejects out-of-scope changes, and accepts improvements only when they clear a noise threshold. The reviewer checks whether a candidate follows its assigned direction and probes for attempts to exploit evaluation loopholes. Only candidates that pass both checks are retained. Their code changes, results and failure reasons then enter an Experience Bank used in later rounds.

Baidu first applied that process to NanoChat AutoResearch, NanoGPT Speedrun and NVIDIA’s SOL-ExecBench, covering training-algorithm search, training acceleration and GPU-kernel optimization. It reported a NanoChat validation result of 0.8763 bits per byte, compared with 0.8878 in the public result it selected. The company also reported 38.9 seconds and 2,590 steps for two NanoGPT tracks, against selected public comparators of 39.9 seconds and 2,690 steps. On SOL-ExecBench, Baidu reported an average score of 0.819, below the 0.827 leader in its comparison table but above two other entries it selected. The public workloads were not rerun as part of the supplied research, so the comparisons remain Baidu’s reported results against particular snapshots rather than independently reproduced rankings.

For its internal work, Baidu said four automated research rounds raised ERNIE 5.5 Flash generation speed on Kunlunxin P800 hardware to 623.6% of its baseline while passing the company’s correctness checks. It separately reported that about three days of work increased deterministic GPU inference throughput by 33.7%, reduced average time per output token by 26% and cut time to first token by 12%. On an already tuned single-request GPU service, Baidu said the system explored seven directions, retained improvements from three and increased decoding speed by another 14.9% without changing accuracy from the baseline. The opened evidence included no independent benchmark or production telemetry for those internal results.

Baidu describes inference infrastructure as ERNIE Evolve’s first real-world application and says it plans to extend the loop to training-system optimization and evaluation design. The company says any expansion should remain interruptible and operate with bounded permissions, safety evaluation, human control and oversight.

More news