Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies

Think Like a World Model, Act Like a VLA. Distilling World-Model Representations into Compact Robot Policies. By Trung Dao and four others, from the University of Wisconsin-Madison and the University of Illinois Urbana-Champaign. Published on the twenty-first of September. Its code repository is worked out: about one hundred and thirty source files, with a readme and a licence.

A vision-language-action model maps observations to actions, and nothing in its objective accounts for how the world responds to those actions. Its robustness is therefore bounded mostly by how much its data covers. World models carry exactly that missing objective, and they are better grounded because of it. But using one means rolling the future forward, which costs seconds for every decision. That keeps world models out of the control loop. So the choice has looked like one between a fast policy that knows little about physical consequences, and a grounded model too slow to act with.

The paper's claim is that the useful part and the expensive part can be separated. What a world model knows about physical scenes lives in its internal features. Generating the future is only the objective that produced those features. So a policy can inherit the grounding, and leave the generative machinery behind.

The mechanism is a single addition to ordinary VLA training: one feature-alignment term. A frozen world model is run over the training frames once, and its features are cached. The student policy then learns to agree with that cache, while it goes on learning to produce actions as it normally would. The teacher is consulted exactly once per training frame, in advance, and never while the student is being optimized.

No teacher model is loaded during training. The projector used for alignment is discarded afterwards. The deployed policy is identical to the baseline without distillation, running in thirty-two milliseconds and one point eight six gigabytes on a consumer RTX 5090. That identity is deliberate. With no added capacity and no extra compute at test time, every gain can be put down to the representation.

A student of zero point eight billion parameters reaches ninety-seven point nine percent on LIBERO. On RoboCasa-GR1 humanoid manipulation, it improves from forty-eight point two percent to fifty point five percent. The same objective carries over to real hardware, on a single-arm platform and on a bimanual one.

The gain also survives changes to the setup: a different student scale, a different backbone, a different alignment layer, a different teacher. The authors take that as a sign of a broad representational prior, rather than a fragile alignment between two particular networks. In other words, what transfers is not a quirk of one teacher and one student, but something about the world model's features in general.

Read the title as a division of labor. Think like a world model, meaning absorb what its features know about how scenes behave. Act like a VLA, meaning keep the speed and footprint of a policy that never simulates anything at run time. The world model itself is paid for once, offline, when the cache is built, and never again.