Large language models (LLMs) provide broad capabilities, yet adapting them to shifting domains, tools, and requirements often requires repeated post‑training. Autonomous systems can automate parts of this workflow by proposing updates, training candidate models, and using evaluation feedback to steer subsequent proposals. As evidence accumulates, a central question arises: after later training has altered the parent model, which past update evidence remains actionable? An update’s effect depends on its parent, the data, and the training stage. Treating past success as context‑free permission wastes compute and, if the child model is promoted, can degrade the future training trajectory.
We formalize this as conditional experience transfer and introduce Boundary‑Calibrated Intervention Transfer (BCIT). BCIT authorizes experience reuse before any weight‑changing training step. Its main components are:
- Bind an observed effect to its source context;
- Verify applicability conditions;
- Veto candidates that exhibit named hard conflicts;
- When needed, obtain current‑state evidence through a bounded training trial. Even fully trained candidates are subject to a shared adoption rule, and only observed events are stored in memory.
We evaluated BCIT on a 4‑billion‑parameter model adapted across three domains: financial reasoning, text‑to‑SQL, and function calling. Candidate updates showed heterogeneous target and retention effects across the evaluated contexts. Under matched candidates, evidence, and compute budget, BCIT authorized fewer harmful updates and achieved higher final‑model quality than the alternatives. These results support treating experience authorization as a distinct problem in autonomous post‑training.
Blogger's Review: BCIT’s explicit context binding and conflict veto provide a safer reuse mechanism, making it a promising addition to autonomous fine‑tuning pipelines.