Correction (2026-09-28). Two claims below did not survive later checks.
“Removing the world model helps by 45% on CheetahRun” came from our JAX re-implementation. On the official TD-MPC2 code with ordinary on-policy data collection, the full and stripped agents were within 5% of each other (1.04×, n = 3; Part 23). The analogous HopperHop “+150” was retracted in Part 21.
“Explicit structure (SE/glass, anti-collapse) buys sample-efficiency” is not supported for SE/glass or anti-collapse terms.
- In the matched 16-task benchmark, glass tied TD-MPC2 at about 1.35× the wall-clock (Part 5).
- Adding SE to the TD-MPC2 latent lowered return-AUC on all three tasks tested (e.g. CheetahRun 23.4 vs 58.9; Thread D).
The claim holds only for analytic-controller priors.
The full, detailed lab notebook (Parts 1–19+) now lives on the dedicated project page: suuttt.github.io/tdmpc-glass. This page keeps only the milestones.
What the project found#
- Abstraction is a speed lever, not a capacity lever. Across DMControl + Panda, explicit structure (consistency, anti-collapse, SE/glass, analytic controllers) buys sample-efficiency where the prior fits, not a higher ceiling versus a budget-matched baseline.
- A value-sufficiency instrument (VBN). Bottlenecking the value head fingerprints each task by how much of the latent the value function needs — three shapes (monotone / flat-high / ramp) that predict where structure has room to help.
- A world model can be harmful. Under planner-collection, removing the world model helps by 45% on CheetahRun (an inversion), while it’s load-bearing on Walker — one variance-inflation mechanism, two regimes.
- Anti-collapse regularizers are redundant for a value-based world model (uniformity ≈ VICReg ≈ vanilla).
Papers (drafts)#
Three write-ups in github.com/SuuTTT/wm-redundancy-paper: When Is Explicit Abstraction Redundant for a World Model?, The Anatomy of “Beating PPO”, and Abstraction is a Speed-of-Learning Lever, not a Capacity Lever.
→ Read the week-by-week detail at the project page.