# No Need to Imagine the Future at Inference—Robot Still Hits 91.8%! Fast-WAM Debunks WAM's Core Assumption

Training Video Objectives Matter More Than Imagining the Future at Test Time

Full article: https://haiguangboy.com/posts/fast-wam

## Core Method

The experimental design is clean—four variants change only one factor:
· Fast-WAM: Keeps video co-training objective during training, generates no future frames at inference, outputs actions in a single forward pass
· Fast-WAM-Joint: Jointly denoises video and actions during both training and inference
· Fast-WAM-IDM: First generates future video, then predicts actions
· No video co-training: Removes the video training objective, everything else unchanged

In other words, it doesn't ask "are world models useful?" but instead separates video modeling during training from future imagination at inference, isolating which one truly contributes.

## Key Results

RoboTwin results: 91.8% vs 90.6% / 91.3% vs 83.8%.
LIBERO results: 97.6% vs 98.5% / 98.0% vs 93.5%.

↳ Whether you imagine the future at inference makes negligible difference; but removing the video training objective causes a drop several times larger.

Real-world evidence is even more straightforward. In the towel folding task on Galaxea R1 Lite, the variant without video co-training achieves only 10% success and takes the longest; differences among the three variants with video co-training are limited. Not learning video prediction during training is the real fatal flaw.

Not imagining the future brings an extra bonus: Fast-WAM has 190ms inference latency, while Fast-WAM-IDM, which generates video before deciding, needs 810ms. A 4x+ speed gap is enough to support real-time control.

## Related Work Comparison

- No structured associations yet.

## Limitations

This is not a conclusion that "future imagination at inference is never useful." Fast-WAM omits the outer autoregressive rollout loop for controlled comparison, focusing on single action chunk generation; current experimental platforms are mainly LIBERO, RoboTwin 2.0, and one real towel folding task. Whether this conclusion holds with large-scale embodied pretraining data and larger models remains unanswered.

## Author's Take (Not from the paper; cross-paper synthesis)

Fast-WAM advances the question from "should robots first imagine future frames?" to "when does the video prediction objective actually add value?" Its answer is practical: what truly matters may not be generating future frames at inference, but using video objectives during training to shape world representations. For engineering systems, this means you can retain the representation benefits of world modeling while dropping expensive test-time video generation, reducing latency and deployment complexity.