ImproveAnyTask: An Autonomous Post-Training Harness for Iterative Model Self-Improvement

Source facts

The preprint describes an autonomous post-training harness that attributes errors, selects update directions, and executes model updates under a limited compute budget. It combines metric- and case-level analysis, consults research-backed strategies, checks small-scale execution before full post-training, and retains validated strategies and scripts. Across 11 tasks, the abstract reports mean gains of 18.29 and 11.97 percentage points for Base and Instruct models, with a maximum gain of 41.96 points under a 24-hour, eight-H20-equivalent budget.

Analysis

ImproveAnyTask treats post-training as a reusable control loop whose error distribution changes after each update. The durable artifact is not only the tuned model but also the validated strategy and execution script.

Uncertainty/Doubt

• This is an arXiv preprint. The task mix, compute accounting, strategy reproduction checks, and reported gains were not independently reproduced here.