Robotics' new bet: models that predict the world instead of reacting to it

One model to imagine and act

GreenWAM is the week's flagship bet. In a preprint, the authors describe a world foundation model for robotics that does double duty: it is both an action-conditioned world model (predict what happens if the robot does X) and a policy (decide what the robot should do), unified across 20 robot families from bimanual arms to dexterous humanoids via a shared anatomical action space. They report top-tier trajectory, depth and 3D accuracy on WorldArena 1.0 and the highest zero-shot LIBERO-plus success among compared methods — and, notably, better real-humanoid performance than pi0.5 and Cosmos 3 Edge. m:…4581

The design detail founders should note: robot specialization improved world prediction and control together, and keeping the world-model modes during policy adaptation preserved robustness. In plain terms, the model that keeps imagining the world while learning to act in it breaks less. If you are building a robotics startup, the architectural question of 2026 is whether your policy and your world model are the same thing — GreenWAM says yes. m:…4581

The recipe details matter for anyone tempted to copy it. In a preprint, the authors describe training on three modes — visual, action-conditioned, and world-action — unified through a shared anatomical action space expressed in common joint coordinates, and creating policies by splitting an action expert off the generator. They report the model is distillable into a single-step policy for real-time control, sits among the top entries on the RoboTwin 2.0 leaderboard, and stays competitive with dedicated vision-language-action policies on CALVIN and LIBERO. It is a foundation-model playbook applied to robots: one trunk, many bodies. m:…4581

Dense rewards for long horizons

Tev attacks the training-data side of the same thesis. In a preprint, the authors introduce Tev-Gym: 1,170 verified task environments across 7 domains, built with a latent-world-first approach — render the environment from a complete hidden specification of the task — plus progress-potential reward shaping, which converts task progress into a dense per-step reward. The payoff they report: Tev-27B matches GPT-6 Astra and Claude Opus 5 on the Agents' Last Exam benchmark at Pass@5, and the larger Tev-320B beats open-source models on Terminal-Bench subsets. m:…9996

The founder-relevant point is leverage: a 27B model matching frontier labs' best on agentic science benchmarks, attributed by the authors to better environments and denser rewards rather than scale. If the recipe holds, the moat in agent training shifts from compute to environment design — which is good news for startups and bad news for anyone whose strategy was 'wait for bigger GPUs.' m:…9996

The scoreboard around Tev is aggressive. In a preprint, the authors report Tev-27B surpassing GLM 5.2 and Gemini 3.7 Flash on Terminal-Bench subsets, reaching Claude Opus 4.8-level performance on middle-horizon tasks like Terminal-Bench 2.1, and generalizing to Terminal-Bench 3.0 and 4.0 subsets — while the larger Tev-320B tops open models on Terminal-Bench 4.0-CM. The authors' attribution is consistent throughout: the gains come from the environments and the reward shaping, not from scale. If you believe them, the scarce input in agent training is no longer FLOPs — it is well-designed tasks with dense feedback. m:…9996

Worlds with more than one actor

Multics takes the world-model idea multi-player. In a preprint, the authors note that existing approaches either pay quadratic attention costs (fully connected) or smear unrelated events together (centralized). Their fix is hierarchical communication — player, group, global levels — with shared group tokens that aggregate locally relevant information. They report state-of-the-art results in 4- and 6-player settings and, more strikingly, generalization from 4-player training to 6-, 8- and 10-player inference without retraining. m:…8022

For founders, the pattern across all three preprints is the strategic signal: prediction is becoming the product. Whether it is a robot imagining its next action, an agent earning dense rewards from a latent world, or a game scaling to ten players without retraining, the winning move is a model of what happens next. The companies that own the best world simulators may end up owning the agents trained inside them — worth thinking about before you rent yours. m:…4581 m:…9996 m:…8022

Multics' engineering is worth a paragraph on its own. In a preprint, the authors describe budget-constrained dynamic grouping based on recent camera trajectories — players near each other in the scene share more communication — and group tokens whose parameters are shared across group slots, so the module accommodates varying player counts without new parameters. They report reduced self-attention computation versus fully-connected baselines alongside the accuracy gains. The design moral: in multi-actor worlds, most communication is waste, and the winners will be the architectures that spend it only where the action is. m:…8022