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

Excerpt

arXiv:2610.06347v1 Announce Type: new Abstract: Adapting general-purpose large language models to specific tasks requires substantial human effort in designing data and training strategies. Sustaining improvement is especially challenging because model updates change the error distribution, requiring strategies to be continually refined. We introduce ImproveAnyTask, an autonomous post-training harness that improves task performance under a limited compute budget. Drawing inspiration from gradien