Large language models often require repeated post-training to adapt to evolving domains, tools, and requirements. Autonomous systems can automate parts of this process by proposing updates, training candidates, and using evaluation feedback to guide subsequent proposals. However, as evidence accumulates, a key challenge arises: past update evidence may no longer be valid after the parent model, data, or training stage changes.
To address this, researchers from Hugging Face propose a new method called Boundary-Calibrated Intervention Transfer (BCIT). BCIT treats experience reuse as a conditional problem, binding each observed effect to its source context. It checks applicability conditions, vetoes candidates with named hard conflicts, and runs bounded training trials on the current parent model when applicability is uncertain.
In experiments with a 4B model (Qwen3-4B) adapted across finance reasoning, text-to-SQL, and function calling, BCIT demonstrated significant improvements. It achieved the highest equal-budget final-model quality among evaluated shared-model alternatives, improving the mean task score by 2.63 points over a Flat-Additive baseline (95% CI [2.10, 3.16]) while satisfying prespecified retention margins.
The authors emphasize that treating experience authorization as a distinct problem in autonomous post-training is crucial. Their findings support the idea that not all past successes should be reused indiscriminately, and that a transparent Reject / Validate / Train policy can lead to better outcomes.