Baidu says a team of AI agents it built to improve its own models made its next ERNIE model generate text up to 6.2 times faster on its own Kunlunxin P800 chips.
The system, called ERNIE Evolve, sets research goals, edits code, runs experiments and keeps only changes that pass both an independent evaluator and a separate “reviewer” agent. Baidu’s ERNIE team described it on Monday in its first public account of Recursive Self-Improvement (RSI), in a post republished by the Chinese tech outlet QbitAI.
The low-latency work on the Kunlunxin chip took about 168 hours. In a second task lasting three days, the agents raised throughput for deterministic inference by 33.7% and cut the time per token by 26%. A third task lifted top decoding speed by 14.9% with no loss of accuracy. Baidu says all the changes passed stress testing and are being used for ERNIE 5.5 Flash, which has not yet been released. Baidu open-sourced the model’s base version on Nov. 5.
Baidu plans to extend the system to training and model research “within defined goals, permissions and resource boundaries,” with human oversight. The figures are Baidu’s own, measured against its own baselines, and have not been independently verified.
Source: QbitAI
