Baidu Unveils ERNIE Evolve to Auto-Evolve Training and Inference Code

Baidu has unveiled ERNIE Evolve, a recursive self-improvement (RSI) system that autonomously proposes research directions, modifies code, and validates results through a multi-agent dual-loop architecture. The system has already been applied to optimize inference for Wenxin 5.5 Flash, achieving a 623.6% speedup in generation speed on Kunlun P800 chips under low-latency scenarios, while increasing deterministic inference throughput to 133% of baseline without compromising determinism.

ERNIE Evolve employs a hierarchical structure with Director, Selector, Executor, Evaluator, and Reviewer agents, forming a "director + parallel evolution tree" nested loop. Only improvements that pass independent evaluation and review are retained, preventing reward hacking. On public benchmarks, it achieved 0.8763 BPB on NanoChat AutoResearch, surpassing the previous best of 0.8878, and completed NanoGPT Speedrun Track 1 in 38.9 seconds. However, the system itself is not open-sourced; only the final code and logs from benchmark tests have been released.

The launch marks a significant step for Baidu in automated AI research, demonstrating a closed-loop from research to production by applying self-evolution directly to its own chips and models. Wenxin 5.5 Flash Base was open-sourced on November 5, and a new model may be released soon, potentially intensifying competition in AI development efficiency and sparking discussions on transparency and reproducibility of automated research systems.