Benchmarks & research
A Doom study asks when a fast decision model should hand control to a slower reasoning model.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Bailo’s offline results improved when a slower model took over the least-confident decisions. Yet none of the thirty-three closed-loop Doom runs reached the exit. Better answers in isolation and faster choices did not translate into successful navigation.
Try the handoff as part of a whole game, rather than judging the two models separately. The fast model chooses an allowed move to meet the immediate need; low confidence or repeated actions can call in a slower model for advice. The failed runs show why those separate decisions are not enough to judge how the pair plays.