Baidu launches ERNIE Evolve, an autonomous research system for training and inference

On November 9, 2026, Baidu Wenxin disclosed for the first time the latest progress in recursive self-improvement (RSI), unveiling ERNIE Evolve, an automated research system that participated in the development of Wenxin 5.5 Flash. The disclosure frames RSI as an approach in which AI models take part directly in their own development process, iteratively improving their capabilities in a self-improvement loop. According to the report, OpenAI, Anthropic, Tencent, Zhipu and several other companies are also exploring ways to bring RSI into model development.

ERNIE Evolve organizes research as a dual-loop system. The outer loop draws on historical experiment results and existing knowledge to propose new research directions, while the inner loop operates under given objectives and constraints, autonomously running experiments, verifying results and accumulating experience. This design allows the system both to keep optimizing established paths and to explore new directions based on observed outcomes. Baidu Wenxin stated that on the three evaluation metrics of two tasks, NanoChat AutoResearch and NanoGPT Speedrun, the system achieved the best publicly verified SOTA results to date.

The system is already in use ahead of the Wenxin 5.5 Flash release: it is applied to inference optimization for the upcoming model, including performance tuning for low-latency scenarios on the Kunlunxin P800, deterministic inference throughput optimization, and extreme decoding speed optimization. Wenxin 5.5 Flash is expected to launch in the near future.