New Method Improves Autonomous LLM Post-Training by Conditional Experience Transfer

Large language models (LLMs) require frequent adaptation to new domains, tools, and requirements, often through repeated post-training. Autonomous systems can automate parts of this process by proposing updates, training candidates, and using feedback for selection. A core problem arises: determining which past successful updates remain relevant and beneficial after the parent model has evolved through subsequent training. Reusing outdated or incompatible experience can waste computational resources and degrade the model's performance trajectory. This research frames the challenge as "conditional experience transfer" and introduces Boundary-Calibrated Intervention Transfer (BCIT). BCIT is designed to authorize experience reuse *before* weight-changing training occurs. It achieves this by binding observed effects to their original context, checking specific applicability conditions, vetoing candidates with known conflicts, and conducting bounded training trials when necessary to gather current-state evidence. Experiments on a 4B model adapted across finance reasoning, text-to-SQL, and function calling demonstrated that candidate updates have varied effects depending on the context. Under matched conditions, BCIT authorized fewer detrimental updates and achieved higher final model quality with the same computational budget compared to alternative methods. This highlights the importance of treating experience authorization as a distinct and critical problem in autonomous post-training workflows.