A proposed AI-agent controller evaluates alternatives before choosing what to explore next

Yann LeCun argued that chain-of-thought prediction is not search because it does not evaluate multiple configurations. Replying to him, a researcher described a controller that assesses alternative next computations, then chooses branches to explore, revisit or stop. The paper reports 71.5% on ProgramBench with GPT-5.5 versus 58.0% for Codex, but says the controller’s overhead can hurt at small compute budgets.