TEMPO switches the same model between actor and critic to value long-horizon rollouts

omarsar0 · x · 2026-08-20

TEMPO divides a long trajectory into macro-steps; at each step the same model switches from actor to critic. The Critic reasons over the current state, calls tools, and estimates the expected remaining return before the full task completes. This targets credit assignment in long-horizon tasks—tens-of-hours rollouts where a single terminal reward must be attributed across thousands of interactions, causing value-free RL methods like GRPO to struggle.

Related event: TEMPO Tops ARC-AGI-3 by Switching Model Between Actor and Critic Roles(3 posts)→

Original post →

More from Research

Research channel →