Experiments, ablations, and what would count as a result

The scientific content of this project is in the comparisons, not in any single run. The following grid isolates each of WAM-RL's mechanisms on Fast-WAM.

Reward ablation.

• (R0) Success only. This is the SimpleVLA-RL / RLinf baseline applied to Fast-WAM.
• (R1) Success plus reconstruction from the frozen world model.
• (R2) Success plus reconstruction from the live (fine-tuned) world model. This is WAM-RL proper.

If R1 and R2 do not beat R0, the reconstruction reward does not transfer to an actor that never reads the imagination, which is a clear and publishable negative result about Fast-WAM's design.

World-model update ablation (crossed with R0, since it is independent of the reward).

• (W0) Video expert frozen; actor-only RL.
• (W1) Video SFT without the feature KL.
• (W2) Video SFT with the feature KL.
• (W3) Video SFT with $\lambda _{\mathrm{action}} = 1$ (self-imitation on successful actions).

WAM-RL's claim is that W2 beats W0. W1 versus W2 tests whether the KL is doing anything. W3 tests whether a much simpler mechanism captures most of the benefit.

Inference-mode question specific to Fast-WAM. After RL with the Optional-IDM model, evaluate both first_frame and idm modes. If the gap between them shrinks, RL has made test-time imagination less necessary, which speaks directly to the Fast-WAM paper's question.

Where to look for headroom. LIBERO averages are near 98%; report LIBERO-10 separately, and consider RoboTwin (lower base success, three cameras, 384 resolution; the same code applies with configs/sim_robotwin.yaml) or the unseen-instruction setting as the primary test bed.

What a result looks like. A reportable positive is a consistent gain on LIBERO-10 or RoboTwin under the deterministic 20-step evaluation over at least three seeds, with the group-relative advantages of Chapter 7 and no degradation on the suites that were already saturated.