A coding agent can tune training recipes while checks guard correctness

Training teams often have to retune a recipe for each model and hardware setup. But can an automated search improve speed while ensuring that a successful run still trains the model as intended? ATTUNE addresses both concerns by pairing recipe search with checks outside the agent.

ATTUNE uses a diagnosis-led coding agent to search training settings, while a separate control plane checks candidates before a recipe ships.

For a given model and training environment, can an automated tuner improve a working recipe without changing what the model learns? ATTUNE has a coding agent diagnose runs and search settings, while a separate control plane checks candidates and the final recipe before shipping.

Researchers compared its shipped recipes with adapted public recipes on each model's workload. Final measurements showed higher mean throughput for GLM-4.5 and Llama-3.1-405B; for Mixtral-8x22B, the best candidate's gain stayed within the noise band, so the public recipe remained. The evidence supports gains in two tested cases, not a universal improvement.

The evaluation held-out models from ATTUNE's development, but used the same software stack, accelerators and agent model. ATTUNE searches settings exposed by the framework and training image, and the study does not show that it finds the fastest possible recipe or works in other environments.