Event
An Experimental Study on EFE-Controlled Planning Compute in World Models
1 reports · 1 sources
Tong Zhang et al. study using expected free energy (EFE) to decide when to stop thinking during planning based on a latent-space world model; their objective function combines goal distance with disagreement among stochastic predictions. Experiments find that on tasks where success rate plateaus with depth, stopping when the elite cost stagnates can reduce refinement, with average success rate close to a fixed-depth reference. However, this stopping rule has failure cases: further optimization may exploit model error, and predicted cost continues to decline after task success rate has plateaued. Under a frozen model, a larger-weight disagreement bonus lowers success rate; when the training data lack the region needed to reach the goal, online updates can recover part of the loss.
Report timeline
- Controlling Test-Time Compute for World-Model Planning with Expected Free Energy: When to Stop Thinking
The study examines whether expected free energy (EFE) can determine how much computation to allocate to planning based on a learned world model, with an objective combining goal distance and disagreement among stochastic predictions. Experiments found that stopping when elite cost stalls reduces refinement while mean success rates stay close to fixed-depth references; with a frozen model, rewarding disagreement lowers success, while online updates partially recover losses caused by regions missing from the training data. The cost-stall rule fails when optimization exploits model errors, and predicted cost keeps falling after task success rates peak.
arXiv cs.RO ·