Baidu launches ERNIE Evolve, a self-improving AI research agent

Baidu's ERNIE research team published ERNIE Evolve on November 9, 2026, an autonomous research system that runs nested outer and inner loops to propose, implement, evaluate and review code-level improvements to Baidu's own models and infrastructure inside isolated environments.

What ERNIE Evolve changes

• An outer loop picks research directions while an inner loop develops and tests improvements, letting the system perform recursive self-improvement (RSI) on model code.
• Five roles divide the work: a Director that selects direction using an Experience Bank and Knowledge Base, a Selector that forms hypotheses, an Executor that edits code in isolated environments, an Evaluator that measures correctness on real workloads, and a Reviewer that checks against evaluation gaming.
• Applied to the in-development ERNIE-5.5-Flash model, it produced a 33.7% throughput increase and 26% lower time-per-token on deterministic GPU inference.
• On Kunlunxin P800 chips, it delivered a 623.6% speed improvement for low-latency inference, plus a 14.9% decoding-speed improvement.
• On public benchmarks it reached a 0.8763 validation BPB on NanoChat AutoResearch, hit target loss in 38.9 seconds on NanoGPT Speedrun Track 1, used 2,590 steps on NanoGPT Speedrun Track 3, and scored a mean SOL of 0.819 on SOL-ExecBench.

What it means for people who build with AI

This is a concrete, benchmarked example of an AI lab letting an agent loop directly on its own training and inference pipeline rather than just generating application code. The gains on Kunlunxin P800 chips, which are not Nvidia hardware, matter to engineers tracking agentic research tooling and inference optimization.

Open-sourced results and next steps

Baidu says the optimized results on public benchmarks have been open-sourced on GitHub, while the real-world inference improvements are already integrated into ERNIE-5.5-Flash production systems. The team's stated plans include extending the system to training-system optimization and eventually letting it help design its own benchmarks and evaluation suites, while keeping human oversight as RSI capability expands.

Sources

• ERNIE Evolve (Baidu, November 9, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.