Robot world models grow up: small, unified, and measured

Small models, longer memory

In a preprint, Marrie and colleagues introduce REKI, a vision-language-action model that claims a new Pareto frontier between performance and model size, outperforming models twice its size on LIBERO-Plus, RoboTwin Clean2Rand and RoboCasa365. The mechanism: register tokens in the action expert summarize the vision-language model's output at each frame, giving the model a memory mechanism for recent visual history — and the authors report this reduces pose oscillations by approximately 4x. m:…7297

The other notable detail: REKI is pretrained exclusively on public data — robot, egocentric human and vision-language datasets — a strategy that reduces proprietary-data dependency. For developers working on real-world deployment, that combination of small size, temporal stability and public-data training is the actual news — architecture doing the work that scale used to do. m:…7297

One transformer for the whole world

In a preprint, Gao and colleagues present Thinker-WM, a unified world model whose Unified Dynamics Transformer supports joint video-action generation, world prediction, policy inference, action-conditioned generation, video-to-action generation and inverse dynamics. The authors report first place on EWMBench, WorldOlympiad and WorldArena. m:…5838

The conceptual bet is that fragmentation — separate models for perception, prediction and control — is the bottleneck in embodied AI. Supporting evidence from the paper: a lightweight motion probe can recover motion information from the model's frozen visual representations, suggesting the shared representation is genuinely carrying dynamics, not just pixels. If that holds up under independent testing, the unified approach simplifies the embodied stack considerably. m:…5838

The architecture verdict

In a preprint, Ma and colleagues do the field a service that individual model papers cannot: a systematic comparison of world-modeling paradigms in vision-language-action models. Their finding: world modeling consistently improves performance over action-only baselines, with notable gains on long-horizon and dynamic tasks, and the decoupled MoT Future Expert paradigm offers competitive overall performance with superior out-of-domain robustness. m:…4831

The second finding is arguably more actionable: world representation effectiveness is paradigm-dependent, whereas 2D semantic representations offer robust, cross-paradigm gains. For a developer choosing an architecture this month, the paper reads as a decision guide — decouple the future-prediction expert, keep the representation semantic and 2D, and expect the biggest wins where the horizon is longest. m:…4831

Measuring what models still cannot do

In a preprint, the WorldCast authors introduce a benchmark for multi-step world-state prediction in multimodal generative models: 320 manually designed cases across eight task categories, with 12 representative image and video generation systems evaluated. The gap is stark: GPT-Image-2.5 scores 72.8, Nano Banana Pro reaches 42.8, and the best open-source model manages only 20.1. The authors attribute the deficit to complex context handling, entangled references and reasoning-heavy tasks — multi-step prediction requires models to forecast successive future world states, and strong generative models effectively borrow the reasoning capacity of their underlying language models. m:…5609

Finally, in a preprint, OmniVerifier-C1 tackles the evaluation side: a decomposed graded verifier for rule-governed generative world modeling. Complex prompts and environmental rules are decomposed into fine-grained components, and the verifier assigns continuous pointwise scores instead of binary True/False judgments. A decoupled reinforcement learning strategy jointly optimizes sub-question-level binary judgment and global alignment scoring, and the authors report that structured, interpretable outputs improve generalization to out-of-domain multimodal critic tasks. Pairwise verification paradigms yield more accurate critiques than the pointwise formulation when ground-truth images are available. As world models move toward structured agents, the verifier may matter as much as the generator. m:…3652