A strong coding model is not enough if the model managing its work does not know when to redirect, verify, or stop.

LoopArena isolates that management problem by fixing Qwen3.7-Plus as the coding Worker and changing only the Controller that decides the Worker’s next assignment.

On full 27-task runs, even the best Controller, GPT-5.5, reached just 24.69% Strict Success Rate; simply restating the original goal every round scored 18.52%, exactly the same as letting the Worker run without control.

useful control has to react to the evolving evidence, shifting the Worker between implementation, verification, recovery, and stopping rather than repeatedly saying “keep going.”

So when evaluating agent systems, benchmark the model that manages the loop separately from the model that writes the code.

– arxiv. org/abs/2608.28281

Title: "LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering"