TD-MPC-Glass, Part 11: The Graph Win Doesn't Reach the Controller
Part 10 reported the campaign’s first GO: a graph latent whose value-decodability generalizes across object counts. We said the real gate was whether that converts to control. It doesn’t.
The control-benefit test
Random-shooting MPC (512 samples × horizon 8, receding) planned with each learned world model; true-env return over 20 episodes × 4 seeds, at the training object count and held-out counts. The random-action floor is the can’t-plan baseline.
| N | regime | graph | monolithic (fair) | random | graph − random |
|---|---|---|---|---|---|
| 5 | in-dist | −374.0 | −368.3 | −360.7 | −13.3 |
| 7 | OOD | −367.9 | −377.3 | −364.9 | −3.0 |
| 9 | OOD | −386.0 | −394.9 | −384.1 | −1.9 |
(source: exp/tdmpc_glass/mechcheck/gwm_control_benefit_s{0,1,3}.json; std ~110.)
Graph ≈ monolithic ≈ random, at every object count. MPC on either learned WM is no better than
acting randomly. The graph latent’s OOD value-decodability advantage (real, robust, Part 10) buys
zero control benefit. (The per-seed CONTROL-GO flags the script emitted were a floor-noise
artifact — a floor-anchored relative gain explodes when the monolithic baseline itself sits at the
random floor; the honest aggregate is the table above.)
Honest caveat
Random-shooting MPC plans on the model’s reward head, and these WMs were trained on a small budget — so “both at the floor” partly means “neither WM is accurate enough to plan with here,” not purely “graph doesn’t help control.” A heavier planner or much longer training could differ. But after a fortnight in which every representation-level advantage failed to reach the controller, the prior is overwhelming and chasing it further is low-EV.
Where the program lands
- Representation axis: graph latents give a genuine, fair-control-surviving compositional-OOD value-decodability advantage (Part 10). Worth one paragraph in the paper.
- Control axis: it does not convert — consistent with the redundancy criterion, now demonstrated across state (16 nulls), temporal (5 levers + high-DoF), and relational/graph (this) axes.
The criterion is the result. The graph direction closes the same way explicit abstraction did: a converged self-predictive world model is already a sufficient, value-aligned abstraction; explicit structure — clustered, temporal, or graph — is redundant at the controller. No ManiSkill escalation is warranted on this evidence. The paper is the deliverable.