Overview
ERNIE Evolve is an automated research system from Baidu's ERNIE team, released in November 2026. Hand it a repository, a goal and an evaluation standard, and it picks research directions, edits code and runs experiments on its own, keeping only the changes that survive a hidden evaluator and an independent review.
The structure is a double loop: the outer loop chooses a direction, the inner loop iterates on experiments. Roles are split into Director for direction, Selector for picking proposals, Executor for editing code and running experiments, Evaluator for scoring against the hidden evaluator, and Reviewer for independent checks. The weight sits on the last two. The evaluator is not exposed, and review is separated from execution, so the system cannot certify its own work. That is the step automated research systems most often gloss over, and the only part of this design worth calling out on its own.
Results concentrate on two public automated-research benchmarks, with state of the art on both NanoChat AutoResearch and NanoGPT Speedrun. The engineering gains are measurable: deterministic GPU inference throughput up 33.7% and single-request decoding 14.9% faster, with those optimizations already shipped into ERNIE-5.5-Flash inference.
Baidu frames this as a step toward recursive self-improvement, with plans to extend to training systems, model capability and evaluation methods themselves. The framing is large; what lands today is smaller and concrete. Having the system hunt for optimizations beats waiting for someone to profile them one at a time.
Key Features
- Autonomous research from a goal: Given a repository, goal and evaluation standard, it chooses directions, edits code and runs experiments without a human specifying each next step.
- Double loop: An outer loop selects direction while an inner loop iterates on experiments, which keeps direction and execution from capturing each other.
- Five role split: Director sets direction, Selector picks proposals, Executor edits and runs, Evaluator scores, Reviewer checks independently, a finer division than a single agent grading itself.
- Hidden evaluator and independent review: The evaluator stays hidden from the execution side, and a change must clear both hidden scoring and independent review to survive.
- State of the art on two benchmarks: Best results on both NanoChat AutoResearch and NanoGPT Speedrun.
- Measurable engineering gains: Deterministic GPU inference throughput up 33.7% and single-request decoding 14.9% faster, optimizations already applied to ERNIE-5.5-Flash inference.
Use Cases
- Performance work on inference and training code, letting the system find bottlenecks and validate fixes.
- Architecture search that requires large numbers of repeated experiments.
- Internal benchmarks with the evaluation criteria expressed in code, handed to the system to improve against.
- Offloading the engineering side of research so humans review results instead of running every trial.
Pros
- State of the art on two public automated-research benchmarks, with reproducible comparisons.
- A hidden evaluator plus independent review narrows the room for self-grading in automated research.
- The optimizations shipped into ERNIE-5.5-Flash, so this is not only a paper result.
- Throughput up 33.7% and decoding 14.9% faster translate directly into cost.
Pricing
Baidu has not published pricing or a public access path. It is a research system for now, and its output already feeds ERNIE-5.5-Flash inference optimization. Whether it becomes an API or platform follows later announcements.
Summary
ERNIE Evolve is Baidu's automated research system: give it a repository, goal and evaluation standard, and it picks directions, edits code and runs experiments, keeping only changes that clear a hidden evaluator and independent review. It sets state of the art on NanoChat AutoResearch and NanoGPT Speedrun, and its inference work lifted deterministic GPU throughput 33.7% and single-request decoding 14.9%, already shipped into ERNIE-5.5-Flash. Baidu positions it as a step toward recursive self-improvement.
Version History
- ERNIE Evolve launches, setting two speedrun records (2026-11-09): Baidu's ERNIE team released ERNIE Evolve, an automated research system that organizes direction selection and experiment iteration in a double loop, with Director, Selector, Executor, Evaluator and Reviewer roles, keeping only changes that clear a hidden evaluator and independent review. It reaches state of the art on NanoChat AutoResearch and NanoGPT Speedrun, lifts deterministic GPU inference throughput 33.7% and single-request decoding 14.9%, with those optimizations deployed to ERNIE-5.5-Flash inference. Baidu describes it as a step toward recursive self-improvement, with plans to extend to training systems, model capability and evaluation methods.