The paper introduces dynamic task routing, a new evaluation protocol where an event stream revises which task is active. The authors repurpose six standard benchmarks into 31,119 dynamic episodes with 373,428 queries.
They propose Revision-Aware Independent Agent Graphs (RIAG), a multi-agent policy that separates temporal resolution from reasoning. RIAG caches solutions by document identity, starts each task with two unexposed attempts, and conditionally invokes audit and repair, using at most four calls per document version.
Homogeneous RIAG achieves 54.24% joint routing-and-answer accuracy at 0.62 calls per query, compared to 32.22% at 18.00 calls per query for the strongest baseline. This shows a dramatic improvement in efficiency and accuracy for dynamic reasoning tasks.