Training models to fill in omitted reasoning from proprietary model outputs: a FIM-style distillation idea

OwariDa · x · 2026-11-23

OwariDa proposes training models specifically for fill-in-the-middle tasks that reconstruct omitted reasoning from proprietary model outputs. The reward function would check, in a separate pass, whether prefix + filled-in reasoning leads to the right answer (or best next step).

He notes the dataset must consist of problems, and flags a key technical pitfall: ensuring the model doesn't leak the answer inside the filled-in reasoning segment.

Original post →

More from Research

Research channel →