Tsinghua University proposes ActionOPD, an action-level online distillation technique to transfer large model decision-making capabilities.

Tsinghua University proposes ActionOPD, an action-level online distillation technique, which transfers large model capabilities by learning Agent decision actions rather than copying Tokens. This technique is suitable for multi-round tool call scenarios, achieving a 3.5x reduction in inference cost on 175B parameter models, and supports complex environmental interactions with a context length of 8k.