Correction (2026-09-28).
- The CheetahRun TD-MPC2 cell (“strip +45%”) came from our JAX re-implementation. On the official code with on-policy collection, the full and stripped agents were within 5% (1.04×, n = 3; tdmpc-glass Part 23).
- The claim that a cheap checkpoint-time probe tells you, per task, whether the world model helps did not hold in a later pre-registered test on ten held-out tasks. Its predictions were at chance (rank ρ ≈ 0.30, p = 0.24; not yet published).
Short version: whether a learned world model helps a planner or hurts it is a property of the task, not of the world-model architecture. We now see the same tasks flip load-bearing vs redundant across two independent WM families — TD-MPC2 (latent-consistency) and Dreamer (reconstruction RSSM).
The result#
We ablate the world model (strip its forward-dynamics learning) and measure the change in return, per task, in each framework:
| task | TD-MPC2 | Dreamer |
|---|---|---|
| cheetah-run | stripping the WM under planner-collection helps +45% (the “inversion”) | vanilla beats stripped by +9.7% (WM helps) |
| walker-run | null — imposed structure is redundant | vanilla ≈ stripped (null) |
Cheetah is the task where the value function needs a large fraction of the latent (low compressibility); walker is value-sufficient (a tiny bottleneck already recovers most of the return). The world model matters on the former and is redundant on the latter — and this ordering is identical in both frameworks.
Why it matters#
Most “does a world model help” debates are argued at the level of architectures. This says the more useful question is at the level of tasks: a cheap, checkpoint-time probe (a value-sufficiency bottleneck) tells you, per task, whether the learned WM has room to help — before you spend the compute. And on the tasks where the value head is already sufficient, the WM is not just neutral: under planner-collection it can inflate the planner’s own target variance (~3×) and actively destabilize learning.
What’s next#
The diagnostic is prescriptive, not just descriptive: it points at a gated world model — one that measures its own value-sufficiency and rollout variance and down-weights itself on the tasks (and states) where it would otherwise hurt.
The detailed lab notebook lives on the project page; the underlying numbers and commit hashes are in the campaign’s public issues and ledger.