OODA-Tool Splits State Tracking From Action Generation, With the Largest Multi-Turn Gains on Smaller Models

arXiv 2608.24368·medium signalXBlueskyLinkedInCopy link

The paper names the failure precisely as state-action competition: direct function-calling and ReAct-style policies learn state tracking and action generation in the same autoregressive trajectory, so pressure to emit the next tool call overwrites information accumulated earlier in the conversation. OODA-Tool routes each decision through controller-checked intermediate states following Boyd's cycle, where Observe reconstructs task state, Orient decides whether execution is warranted, Decide forms an admissible action structure and Act emits the external call. Evaluated on Qwen3 from 0.6B to 14B across multi-turn, multi-tool and incomplete-information settings, it consistently improves task success, with larger gains on smaller models and on tasks depending heavily on accumulated prior tool results.