World-Language-Action Model for Unified World Modeling, Language Reasoning, and Action Synthesis

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

World-Language-Action (WLA) models are introduced as a new class of embodied foundation models that unify world modeling, language reasoning, and action synthesis. These models accept textual instructions, images, and robot states to jointly predict textual subtasks, subgoal images, and robot actions. Unlike previous World-Action Models (WAMs) that use bidirectional diffusion Transformers, WLA employs an autoregressive (AR) Transformer backbone to predict a "next state" comprising both semantic-level textual intention and fine-grained physical dynamics. The WLA-0 prototype, with 2B active parameters, achieves 40 ms inference latency on an NVIDIA RTX 5090. Evaluations show WLA-0 achieves state-of-the-art multi-task and long-horizon learning, including a 92.94% success rate on RoboTwin2.0 Clean and 56.5% on RMBench. It also demonstrates the ability to learn novel tasks directly from cross-embodiment robot videos without action annotations.

Key takeaway

For robotics engineers developing embodied AI, WLA models offer a compelling architecture for real-time control and complex task execution. You should consider WLA's autoregressive design for its efficiency and ability to handle long-horizon tasks through language-based planning and memory. Its capacity to learn from action-free, cross-embodiment videos could significantly reduce your data collection burden for novel skills.

Key insights

WLA models unify world modeling, language reasoning, and action synthesis for robust embodied AI.

Principles

• Next state prediction should combine high-level textual intention and low-level physical dynamics.
• Autoregressive Transformers can unify language generation and physical dynamics modeling.
• Implicit parameter updates for world prediction allow disabling it during inference for efficiency.

Method

WLA uses an AR Transformer backbone, a World Expert for future visual state prediction (via VAE features), and an Action Expert for action generation, trained end-to-end with meta-queries.

In practice

• Use lightweight diffusion Transformers for World Expert.
• Predict static future frames, not full video clips, for physical dynamics.
• Employ test-time scaling (TTS) with value models for improved control.

Topics

• Embodied AI
• World Models
• Language Reasoning
• Action Synthesis
• Robot Learning
• Real-time Control

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• World-Language-Action Model for Unified World Modeling, Language Reasoning, and Action Synthesis — WLA models unify world modeling, language reasoning, and action synthesis using an autoregressive Transformer for embodied AI.
• World Action Models give robots the ability to simulate consequences before they move — World Action Models enable robots to simulate action consequences, improving generalization and learning from unlabeled video data.
• Chain of World: World Model Thinking in Latent Motion — CoWVLA (Chain-of-World VLA) introduces a novel "Chain of World" paradigm to address limitations in existing Vision-Language-Action (VLA) models, which either waste capacity on redundant backgrounds or…
• The Sequence Knowledge #842: Everything You Need to Know About World Models — World models represent the shift from text-based AI to physical simulation, enabling AI to understand and operate within dynamic reality.
• Robbyant Open Sources LingBot World: a Real Time World Model for Interactive Simulation and Embodied AI — LingBot World enables long-horizon, interactive video simulations for embodied AI using a 28B parameter diffusion transformer.
• Roundtables: Can AI Learn to Understand the World? — World models are emerging as a critical AI frontier to ground systems in real-world understanding, addressing LLM limitations in common sense and causality.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.