One bit to rule the planner: SIML FEPGate + V-JEPA 2

A small intervention in the planning loop

This experiment adds a SIML surprise signal to the candidate-selection loop of a frozen V-JEPA 2 planner. The encoder and predictor are left unchanged; the intervention operates on planning candidates rather than retraining the model.

The lab reports comparisons of final reaching error, per-step latency, and estimated energy use under the same planning budget. The method, setup, and implementation links are available in the original report.

How to read the result

These are reported lab measurements from a particular experimental setting, not an independently replicated claim about all world models. The interesting research question is whether a compact feedback signal can help a planner use computation more selectively.

The experiment extends the earlier technical note on predictive representations and forms part of Deep SIML's continuing work on simulation and efficient decision systems.