System Switch: When Should a Fast Decision Model Stop and Think?
Deferring low-confidence actions helped offline, but the live Doom agent still could not finish a run.
The paper tests a fast actor that hands control to a reasoning vision-language model only when a gate opens. On held-out questions, sending the least confident 30% of decisions to the reasoner beat random deferral, with gains tied to the actor’s confidence sensitivity. In closed-loop Doom, none of the tested variants reached the exit across 33 games. Reasoning or fixed exploration opened more doors, but also exposed limits: the reasoner confused ordinary doors for locked ones when told some doors require keys. HF Daily Papers' note
The paper tests a fast actor that hands control to a reasoning vision-language model only when a gate opens. On held-out questions, sending the least confident 30% of decisions to the reasoner beat random deferral, with gains tied to the actor’s confidence sensitivity. In closed-loop Doom, none of the tested variants reached the exit across 33 games. Reasoning or fixed exploration opened more doors, but also exposed limits: the reasoner confused ordinary doors for locked ones when told some doors require keys. HF Daily Papers' note
score 4