Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training
The paper argues that post-training systems need a gate for when old update evidence no longer applies.
The authors frame this as “conditional experience transfer,” where an update that worked for one parent model may fail after later training changes the model state. They introduce BCIT, which ties each observed update effect to its source context, checks for applicability, and vetoes hard conflicts before more weight-changing training. In tests on a 4B model across finance reasoning, text-to-SQL, and function calling, BCIT approved fewer harmful updates and produced better equal-budget final quality than the compared alternatives. HF Daily Papers' note
The authors frame this as “conditional experience transfer,” where an update that worked for one parent model may fail after later training changes the model state. They introduce BCIT, which ties each observed update effect to its source context, checks for applicability, and vetoes hard conflicts before more weight-changing training. In tests on a 4B model across finance reasoning, text-to-SQL, and function calling, BCIT approved fewer harmful updates and produced better equal-budget final quality than the compared alternatives. HF Daily Papers' note
score 4