What if an AI system could do more than write code? Baidu introduced ERNIE Evolve on November 9, 2026, describing a research engine that chooses directions, edits AI code, runs experiments, and keeps only changes that survive strict evaluation and review.
The system targets both AI research and inference optimization. Baidu reported better results than public comparisons on three NanoChat and NanoGPT metrics, along with 33.7% higher throughput in one deterministic ERNIE-5.5-Flash inference test. That combination points toward a new kind of engineering workflow, where the search for improvements becomes a structured loop instead of a sequence of isolated experiments.
An AI Research Loop Built Around Evidence
ERNIE Evolve starts with three inputs: a repository, an optimization objective, and evaluation criteria. It then modifies code and runs experiments inside isolated environments, using both failed and successful attempts to guide later iterations.
Its role reaches beyond code generation. The system selects lines of inquiry, tests hypotheses, compares candidate versions, and rejects changes that do not meet the required standard. The design separates the agents that create changes from the systems that decide whether those changes deserve to survive.
An outer-loop Director chooses a research direction and a starting version. It draws on an Experience Bank containing previous experiments and a Knowledge Base covering papers, open-source code, and domain knowledge. From there, parallel lanes hold trees of candidate versions, allowing multiple paths to develop at the same time.
Inside each lane, a Selector picks a candidate and a hypothesis, while an Executor implements the proposed change. This creates a clear chain from research idea to working code, with each step tied to a version that can be tested and reviewed.
Acceptance depends on both a fixed Evaluator and a separate Reviewer. The Evaluator sits outside the repository and remains invisible to the agents, excludes out-of-scope changes, and requires gains above a noise threshold. The Reviewer checks whether a result follows the selected research direction and discards exploits that target loopholes in evaluation.
That final safeguard changes what counts as progress. An off-direction candidate cannot become the best version, even if it posts a strong score through a tactic that misses the research goal.
Benchmarks Show Gains, With Limits Made Visible
Baidu’s NanoChat work adds two prediction branches alongside the neural network. One branch uses n-gram statistics from the training corpus, while the other retrieves matching documents and predicts from their continuations.
A learned gate adjusts the mixture based on each branch’s performance at earlier token positions. The result is a system that combines neural predictions with corpus statistics and document retrieval, then changes that balance as token generation proceeds.
The results repository publishes final code and measurement logs. Each log embeds the executed code or identifies it by its SHA-256, giving the experiments a trail that connects reported measurements to the code used to produce them.
The repository also lists seven SOL-ExecBench kernel artifacts with scores ranging from 0.6669 to 0.9920. Baidu’s comparison places ERNIE Evolve ahead of Recursive and Hyra-1.0 on that task, but the SOL-ExecBench row also qualifies the broader results: ERNIE Evolve did not lead every listed metric.
NVIDIA’s task definitions and workload data are not redistributed. That boundary matters because the published artifacts and logs support inspection without claiming that every underlying benchmark asset can be shared.
From Research Search to Faster Inference
Baidu also applied ERNIE Evolve to low-latency inference for the in-development ERNIE-5.5-Flash on Kunlunxin P800 hardware. In one deterministic test, the company reported 33.7% higher throughput.
This application brings the system’s research loop into a direct engineering setting. Instead of searching only for model or benchmark improvements, ERNIE Evolve can explore changes aimed at making inference faster under a defined objective and fixed evaluation process.
The larger idea is ambitious but concrete: provide a codebase, define what success means, and let a controlled system explore many candidate changes while separate checks guard against noise, scope violations, and evaluation loopholes. The system does not treat every successful run as a breakthrough; it retains improvements only when they pass both measurement and review.
With code, logs, candidate histories, and benchmark outcomes in the picture, ERNIE Evolve presents AI optimization as a repeatable search process. Its next test will be whether this approach can keep delivering useful gains as research directions become harder, inference targets become tighter, and the in-development ERNIE-5.5-Flash moves closer to real deployment.
Based on
- Baidu’s ERNIE Evolve Automates AI Research and Inference Optimization — unite.ai
- What matters most in the AI race — edition.cnn.com
- Unizen AI Publishes an Auditable ASI-Bench Result: Just Under 80 in Local Evaluation, With Every Score Receipted and Hash-Chained, for 20 Cents on a Laptop | Markets Insider — markets.businessinsider.com
- Chinese AI Models Fall Behind Anthropic, OpenAI in New AI Rankings – Business Insider — businessinsider.com
- Beyond LLMs? China forges new path for AI with focus on industrial systems | South China Morning Post — scmp.com
