Adaptive-View Self-Distillation accepted to NeurIPS 2026

A researcher behind Adaptive-View Self-Distillation announced its acceptance to NeurIPS 2026. They say the method draws on multiple privileged-information views, including hints, partial or full solutions and execution outputs. A token-level gate balances what the resulting teachers agree on with signals unique to each view, aiming to provide better learning signals for math and code reasoning.