Baidu discloses ERNIE Evolve, an RSI research system already used to optimize ERNIE 5.5 Flash inference

On November 9, Baidu ERNIE disclosed its first progress on recursive self-improvement (RSI): ERNIE Evolve, an automated research system that has been applied to inference optimization for the upcoming ERNIE 5.5 Flash. The details come from a Baidu technical blog post, republished with authorization by Chinese-language outlet QbitAI (量子位), which carried it as content supplied by Baidu ERNIE.

What ERNIE Evolve does

According to Baidu, ERNIE Evolve works from a goal and constraints. It proposes research directions, runs experiments, verifies results and accumulates experience. It can keep refining existing paths or explore new directions based on observed results.

Baidu claims the system reached the best publicly verified results on three metrics across two tasks, NanoChat AutoResearch and NanoGPT Speedrun.

Architecture

The system uses a two-layer loop:

• Outer loop: combines past experiment results and existing knowledge to propose new research directions and run several lines in parallel.
• Inner loop: for a given direction, repeatedly modifies code, evaluates the effect and records what was learned.

Several agents divide the work. The Director selects research directions. The Selector sets each round's objective and code starting point. The Executor modifies code and produces new versions. After each round, an independent Evaluator program tests code performance and experiment results, and a Reviewer checks whether the research has drifted from the goal and whether the evaluation has loopholes. Only improvements that pass both evaluation and review are kept. Code changes, results and lessons from every round are logged for later reference.

Results on ERNIE 5.5 Flash inference

Baidu reports three production-oriented optimization tasks:

Task	Effort	Result
Low-latency optimization on Kunlunxin P800 (昆仑芯 P800)	About 168 hours, 4 rounds of direction exploration	Best version reached 623.6% of baseline generation speed
Deterministic inference throughput	3 days, 3 directions	Throughput at 133.7% of baseline; average per-token generation time down 26%; first-token latency down 12%, with determinism strictly preserved
Extreme decoding speed	7 directions explored, 3 produced gains	Cumulative inference performance at 114.9% of baseline, accuracy matching baseline

On the P800 task, the system first adopted native fused operators and enabled speculative decoding, then optimized host scheduling, synchronization and communication paths. In the decoding task, the gains came from optimizing MoE and mHC operators, matrix multiplication and scheduling strategy, and the critical path of the draft model.

Baidu states all changes passed the ERNIE team's stress and correctness testing and completed production-grade validation.

Context and next steps

RSI means having AI take part directly in its own development to form a self-improvement loop. The Baidu post says OpenAI, Anthropic, Tencent and Zhipu are among the companies exploring RSI in model development.

Baidu's blog says inference infrastructure is the first place RSI has landed in real development. The same loop is to extend gradually to training system optimization, model capability research and evaluation system iteration. Baidu says this expansion will happen within defined goals, permissions and resource boundaries, under continuing safety assessment and human oversight.

On November 5, Baidu open-sourced its latest base model, ERNIE 5.5 Flash Base. Baidu claims it reaches state-of-the-art results among models of its size on benchmarks covering knowledge, reasoning, math, code, multilingual and long-context tasks. A person close to Baidu said a new ERNIE model may be released soon. Pricing, context window and parameter count for ERNIE 5.5 Flash have not yet been disclosed.

What this means

This is the first time Baidu has described RSI work in concrete terms, and the first application is narrow: inference speed on a model about to ship, not model capability. The reported gains were measured against baselines on the ERNIE team's own production stack and passed its stress and correctness checks, including accuracy parity in the decoding task and strict determinism in the throughput task.

The Evaluator and Reviewer roles are the notable design choice. Keeping only changes that pass an independent evaluation and a check for goal drift and evaluation loopholes is how Baidu addresses the risk of an automated system gaming its own metrics. Baidu's stated plan to extend the loop to training systems and capability research, with safety assessment and human oversight, means the scope of what ERNIE Evolve touches is expected to widen. The ERNIE 5.5 Flash release will show how these inference gains appear in the shipped model.