AI-generated conceptual image: a circular track and an inspection gate symbolize experimental iteration and independent acceptance
A model that can write code is still some distance from being able to conduct research continuously: it must also decide what to change next, discern whether an improvement is real, and remember which attempts have already failed. ERNIE Evolve, publicly unveiled by Baidu's ERNIE team on November 9, puts these steps into a single automated research loop. The news here is the disclosure of methods and results; it is a separate event from the previously released ERNIE 5.5 Flash Base weights.[B1]
According to the official description, the outer loop selects research directions, while the inner loop modifies and tests candidate code. The evaluation program is placed outside the repository being optimized and is not made available to the agent; a separate review stage determines whether modifications cross boundaries or exploit scoring loopholes. Here, “independent” means evaluation is separated from experiment execution; it does not mean that an external institution has independently verified all results.[B1]
The public results repository gives some figures a traceable anchor: the validation-set BPB on the NanoChat task is 0.8763; NanoGPT Speedrun Track 1 records 38.92 seconds over 18 runs, and Track 3 records 2590 steps over 8 runs. The repository also provides the corresponding code and logs, and states that the code used for runs is embedded in logs or identified by SHA-256. Reviewers still need to match hardware, data, and software conditions; having logs does not mean reproduction has been completed.[B2]
Baidu's blog comparison table shows that the first two task categories exceed the best public results they were compared with, but the SOL-ExecBench average score of 0.819 remains below the table's leading value of 0.827. Progress on a few tasks cannot be generalized as overall leadership. Baidu also says that the system, used for the still-under-development ERNIE-5.5-Flash, increased throughput by 33.7% under a specific deterministic GPU inference workload. Follow-up media reports also make clear that these gains come from Baidu's own disclosures; this is not evidence that a new model has already been launched externally.[B1][B3]
From an engineering perspective, what deserves continued attention is where acceptance authority sits. If the agent modifying code can also modify the scoring method, a impressive number may simply mean the test was changed. Separating the evaluator from the candidate implementation at least preserves a clear boundary for subsequent review; actual reliability also depends on whether tests cover real work and whether new errors go undetected.
The inspiration such processes offer enterprise R&D need not start with a grand term like “self-evolution.” One can begin with a clearly bounded optimization task, fix correctness conditions, resource budget, and human stop points, and then log each round of experiments. Only when the execution costs saved exceed the search, evaluation, and review costs does automated research have stable economic significance.
As of this verification, what the official materials provide is a research description and a partial results repository; on that basis, it cannot be claimed that the entire ERNIE Evolve system has been open-sourced. The blog's direction of extending the research loop to training and capability optimization remains a future plan.[B1]
Sources:
[B1] Baidu Research Blog | 2026-11-09
[B2] ERNIE Evolve Results official repository
[B3] The Next Web | 2026-11-09
