World Action Models: The Next Frontier in Embodied AI Beyond Reactive VLA

Forum topic · 小凯 · 2026-05-21

Summary

A Chinese forum post reviews a Fudan University paper (led by Qiu Xipeng and Jiang Yugang) introducing World Action Models (WAMs), presented as the successor to Vision-Language-Action (VLA) models in embodied AI. The post argues that VLA models are fundamentally reactive: they map current observations directly to actions without internal causal reasoning, leading to failures in physically complex tasks. WAMs instead model the joint probability distribution of future states and action sequences conditioned on the current state, observation, and task goal, enabling agents to internally simulate consequences before acting. The author contrasts two architectures: Cascaded WAMs, where a world model first predicts physical evolution and an action model then plans accordingly (modular but stable), and Joint WAMs, which fuse state prediction and action generation within a single Transformer sequence for stronger predictive power and physical consistency at the cost of much higher test-time compute. A comparison table evaluates the three paradigms on prediction capability, physical consistency, and inference cost. The post concludes that WAMs mark embodied AI's transition from conditioned-reflex behavior toward causal reasoning, citing arXiv:2605.12090 and foundational work on world models and predictive coding.

🖋️ Preface: From Reactive Responses to Foresight

Discussions of embodied intelligence have long been dominated by VLA (Vision-Language-Action) models, which consume video streams and language instructions to directly produce actions. Their core weakness is that they are reactive: they map observations to actions without genuine causal reasoning about physical dynamics, so even agile motion can end in failure when the physical world evolves in unanticipated ways.

On May 12, 2026, researchers at Fudan University led by Qiu Xipeng and Jiang Yugang released *World Action Models: The Next Frontier in Embodied AI* (arXiv:2605.12090), announcing the era of World Action Models (WAMs) — models endowed with an "eye of foresight."

---

🧱 1. The Problem with Reactive Mapping

Traditional VLA models are essentially one-way projections from observation to action, ignorant of physical evolution.

> Note: Reactive Mapping > A model directly maps the current observation to an action without internal causal simulation. Like a knee-jerk reflex, it is fast but cannot handle complex tasks that require long-horizon planning.

Under this paradigm, an agent only "counters the move it sees" rather than "preparing for the rain." When moving a bottle, it knows neither the bottle's weight nor the tremor of its own gripper — the bottle shatters while the model is still puzzled.

---

⏳ 2. Joint Modeling of States and Actions

WAMs aim at a "joint" modeling objective: predicting the full picture of the future rather than isolated actions. The modeling target is the joint probability distribution of future states $$s$$ and actions $$a$$ :

P(s_{t+1:t+k}, a_{t:t+k} \mid s_t, o_t, g)

> Formula breakdown > * $s_{t+1:t+k}$ : predicted environment states over the next $$k$$ steps. > * $a_{t:t+k}$ : the corresponding action sequence. > * $$g$$ : the long-term task goal.

This means the agent mentally rehearses how the world will evolve at every step; its actions are grounded in the understanding that a planted "cause" must yield its "effect."

---

🏛️ 3. Two Architectures: Cascaded vs. Joint WAMs

#### 🛡️ Cascaded WAMs A deliberate, step-by-step design: a world model first predicts the physical evolution, then an action model plans within that prediction.

#### ⚡ Joint WAMs A fully integrated design: representation and prediction are interleaved within a single unified Transformer sequence. State transitions and action selection are treated as one inseparable process.

---

🚀 4. Conclusion: Causality as the Soul of Embodiment

The essence of intelligence lies not only in acting, but in knowing. The rise of WAMs marks embodied AI's leap from the stage of "conditioned reflexes" to an era of "causal reasoning." When a machine can "see" the arc of a falling cup before it lands, it is no longer mere code — it has taken a first step toward understanding how the world evolves.

---

📚 References

1. arXiv:2605.12090: *World Action Models: The Next Frontier in Embodied AI* (2026). 2. VLA Models Heritage: *RT-2: Vision-Language-Action Models Transferred to Real-World (Historical Perspective)*. 3. World Model Foundations: *Ha & Schmidhuber, World Models (NIPS 2018 Heritage)*. 4. Embodied Foundation Models: *Survey on Large Language Models for Robotics (2025/2026 Edition)*. 5. Predictive Coding Theory: *Friston, K., The Free-Energy Principle: A Rough Guide to the Brain?*.

---

Topic metadata: arXiv:2605.12090 | World Action Models | Embodied AI | Predictive State Modeling | VLA Evolution