Agents stop learning from scratch: code worlds plus a real-time neural renderer build shelter by round 4

State and appearance are separated: a scene program runs in a simulator or game engine that maintains persistent state and executes explicit interaction rules, while a shared neural renderer generates what the agent sees from structured conditions the engine exports. Simulators and game engines already offer explicit state and programmable rules, but expanding their visual diversity requires substantial 3D asset creation and complex rendering pipelines; video world models synthesize rich interactive observations from data, yet task-relevant state and physical transition rules stay implicit in generated histories and learned representations, precluding direct inspection, editing, and testing. The renderer observes state only through conditions, and the transition takes no rendered observation as input, so rendering errors cannot accumulate in the state; a recorded state trajectory can be rendered again under a different appearance reference or camera without altering the underlying events.

The renderer is built from a pretrained bidirectional video model: it is adapted to geometry conditions, converted to block-causal generation, and distilled into a few-step real-time renderer. Operating a video model as an interactive renderer requires it to strictly follow updated engine conditions, keep visual consistency across extended rollouts, and synthesize observations in fine-grained short blocks so agents get immediate feedback after brief actions. Adversarial Forcing matches distributions on the model's own rollouts and uses exact replay so later losses update how history is encoded; replay runs block by block and is bitwise identical to the sampled trajectory, whereas full-sequence FlexAttention replay differs by 3.99% relative error.

Agents perceive the world only through rendered first-person frames and, after each round, write inheritable skill files that agents in fresh conversations read next round. The 2019 hide-and-seek study learned from random weights with self-play reinforcement learning, reporting shelter construction after roughly 25 million episodes and ramp use after another 75 million, with agents observing object state directly. Perceiving only rendered observations, with no object coordinates or hidden world state, hiders built shelter with panels by round 4 and seekers crossed walls with a ramp by round 10; four further worlds each ran four rounds, with companion-dog engagement rising from 13 to 19, one-lane bridge time falling from 71 to 41 seconds, herding penning all four sheep in every round after the first, and the quarry loader clearing, delivering, and parking in 329 of its 360 seconds by round 4.

The reported results cover the renderer's visual quality, replay accuracy, and throughput under one NVIDIA H100, BF16 computation, a four-step sampler, and blocks of four latent frames, plus multi-round scores for pretrained agents in five code worlds. The intended readers are researchers and engineers building embodied-agent training environments or interactive world models. The geometry conditions the renderer receives describe visible geometry only; color, material, and object identity are carried by the appearance reference and visual history.

The authors note that depth and surface normals are a redundant conditioning interface, and that spatial downsampling can remove thin structures, narrow gaps, and small contact changes; geometry alone also leaves some state unspecified, since a rotationally symmetric object can spin without changing its depth or normals, and material, color, and object identity are not determined by geometry, with visual history possibly losing these attributes after long occlusions or revisits beyond the memory window. They list more abstract and compact condition representations, letting coding agents write and check worlds automatically, and deciding which lessons to keep, merge, or drop across worlds as open problems.