Strong models can review agent steps instead of writing each

When an agent needs many tool calls, must a large model write every step to preserve task success? A smaller model may handle routine steps, but rejecting any step that differs from the larger model’s own choice can throw away valid progress. The paper asks whether the larger model can review a grounded draft and keep what works.

SpecAgent lets a smaller model draft grounded tool steps for a larger model to review, keep when valid, and repair when needed.

The tau-squared Airline test asks whether agents can complete multi-turn customer-service tasks involving database reads, changes, and simulated-user replies. Its Pass@1 score is the average success rate per trajectory across three runs for each task.

With the same model deployments and tasks, SpecAgent scored 86.0%, equal to Strong-only and above Weak-only’s 80.0%.

The paired uncertainty interval for SpecAgent versus Strong-only includes differences in either direction, so the measured success supports comparable performance in this test, not a proven tie.

The authors also estimate lower memory traffic and elapsed time than Strong-only under their shared serving setup.

The Airline success comparison is limited to that test and its reported runs; its paired uncertainty interval includes differences in either direction.

Memory traffic is an estimate, and the resource results depend on serving conditions, including whether the larger model is limited by memory bandwidth.