Action-Sequence Alignment for Multi-Turn On-Policy Distillation
November 11, 2026
This method improves multi-turn tool-use agents by selecting the most compatible successful trajectories as references during distillation. It uses longest common subsequence (LCS) as a proxy for latent state compatibility to avoid providing misleading supervision.
HOW THIS AFFECTS YOU
●
builderThis provides a way to better train small models to use tools by picking the right successful examples to learn from.
●
researcherYou can frame reference selection in on-policy distillation as an optimal transport problem using action sequences.