Baidu Lets AI Modify Its Own Code and Run Experiments, First Learning to Discard Useless Results
ERNIE Evolve builds a closed loop that connects code modification, experiment execution and performance gain validation. It has outperformed all public state-of-the-art records on NanoChat and NanoGPT Speedrun, and also boosted the inference speed of ERNIE-5.5-Flash. The core bottleneck of automated research arguably lies in result filtering.
As reported by Michal Sutter from MarkTechPost, Baidu has an automated research system called ERNIE Evolve. Its use case is very specific: modifying code, running experiments, and only keeping the changes that are verified to bring tangible gains. The report notes that it has exceeded the top publicly available scores on NanoChat and NanoGPT Speedrun, and accelerated ERNIE-5.5-Flash inference as well.
An automated research system sounds like a tool that lets AI write academic papers on its own. Let's look at its closed workflow first: generate code changes, run experiments, and filter out invalid attempts based on returned results. ERNIE Evolve's rule of "only keep verified gains" is its key highlight. The biggest pain point for automated experiments is producing a large number of seemingly runnable changes that bring no actual performance improvement. The ability to validate effectiveness and discard useless outputs is far more valuable than generating extra versions of code.
The system's achievements are divided into two categories. NanoChat and NanoGPT Speedrun are public tasks and benchmarks, suitable for horizontal cross-participant performance comparison; the inference acceleration of ERNIE-5.5-Flash falls into internal model optimization. The report does not give specific improvement percentages, nor does it elaborate on the full experiment setup. What is confirmed is that it has delivered valid results on these aforementioned scenarios.
For developers, this is more like a sample automated experiment framework that connects code search, experiment execution and result validation as a full pipeline. For general readers, you can understand it as an AI that runs trial and error within limited tasks, and decides what to keep based on experiment records. It is still far from the vision of a "general-purpose AI scientist", and its performance on public benchmarks cannot be directly extrapolated to all research scenarios. The real points worth tracking in follow-ups are: whether the validation process is rigorous, whether the performance gains are reproducible, and whether its advantages still hold when migrated to other tasks.
Original Article: MarkTechPost
发布时间: 2026-11-10 00:31