Nano World Models Puts a Reproducible Baseline Under World-Model Research
World-model research has no common baseline across datasets, training recipes, and metrics; NanoWM gives teams a single, reproducible pipeline to isolate which design decisions actually drive planning success. Its ablation on semantic latents, in particular, warns that representation quality does not translate to controllable dynamics under a diffusion objective.
Nano World Models (NanoWM) is an open-source, minimal codebase for building world models, built around Diffusion Forcing. It unifies multiple predictive objectives (x, epsilon, v, and flow matching), five action-injection mechanisms, three latent encoders, and four model scales into one Hydra-based pipeline. The project's stated goal is to shift world-model research away from inventing new architectures and toward rigorous, reproducible ablation of design choices that actually matter.
The finding that semantic latents (Web-DINO, V-JEPA) drive action embeddings to near zero under diffusion training means latent space choice cannot be separated from the training objective; a representation that is good for recognition is not automatically good for controllable generation.
Cross-attention, the most expressive action-injection mechanism, performed worst on RT-1 because the model could predict adjacent frames without using the action signal, a symptom of overfitting to temporal correlation rather than a failing of the mechanism itself.
The paper's decision to standardize six datasets and a fixed evaluation recipe is a direct response to the fragmentation of world-model research, where every lab uses its own data, training scheme, and metric, making scientific comparison nearly impossible.