AnalysisAI ModelsSeptember 14, 2026

Lightning Weave composes reasoning capabilities into one efficient student model

Lightning Weave uses on-policy distillation to merge independently trained reasoning capabilities into a single student, improving both accuracy and token efficiency on math and code benchmarks. The method relies on log-ratio shifts and Tilted-Target DOPD to handle policy shift during composition.

1 source

More stories today

Open the live feed