Sources
about 2 minutes
Primary paper
- Fast-WAM: Do World Action Models Need Test-time Future Imagination? (2603.16666) · HTML · project page · code
Masking and attention design
- UNIVERSE: Unified Video Action Models for Autonomous Driving with Flexible Mask-Modulated Modality Generation (2607.05133)
- NoiseGate: Learning Per-Latent Timestep Schedules as Information Gating in WAMs (2605.07794)
- MaskWAM: Unifying Mask Prompting and Prediction for World-Action Models (2606.13515)
- Next Forcing: Causal World Modeling with Multi-Chunk Prediction (2606.11187)
- Discrete-WAM: Unified Discrete Vision-Action Token Editing (2606.05645)
- Understanding and Mitigating the Video-Action Generalization Gap via Temporal Ratio (2607.08127)
Architecture and efficiency
- Faster-WAM: Do World Action Models Need Deep Action Modules? (2608.02365)
- Faster-WAM: Efficient Inference-Time Future Conditioning for Robust WAMs (2608.04404)
- Enfold: Folding World Model Imagination into Predictive Representations (2607.26657)
- ImageWAM: Do WAMs Really Need Video Generation, or Just Image Editing? (2606.19531)
- World Tokens: Enhancing Embodied Policies with Training-Time World Modeling (2608.09730)
- C³ache: Accelerating WAMs with Cross Inference Chunk Cache (2606.08962)
- Efficient-WAM: A 1B-Parameter WAM with Low-Cost Future Imagination (2606.10040)
Cheap, latent or adaptive imagination
- ForeWAM / Foresight Without Seeing: Latent Futures for WAMs (2608.11605)
- LAWA: Latent Action as Intention Enables Efficient Future Imagination (2608.24882)
- GlanceWAM: Sparse Test-Time Imagination for World-Action Models (2608.23927)
- RISE: Adaptive Imagination for World Action Models (2608.20430)
- When to Trust Imagination: Adaptive Action Execution for WAMs (2605.06222)
Diagnostics, robustness and critique
- Beyond Task Success: Behavioral and Representational Diagnostics for WAM and VLA (2606.01095)
- BadWAM: When World-Action Models Dream Right but Act Wrong (2607.15207)
- Mask World Model: Predicting What Matters for Robust Robot Policy Learning (2604.19683)
- Do World Action Models Generalize Better than VLAs? A Robustness Study (2603.22078)
- S-VAM: Shortcut Video-Action Model by Self-Distilling Geometric and Semantic Foresight (2603.16195)
Surveys, tutorials and community resources
- From World Models to World Action Models: A Concise Tutorial for Robotics (2607.00836)
- World Model for Robot Learning: A Comprehensive Survey (2605.00080)
- Video Generation Models as World Models: Efficient Paradigms, Architectures and Algorithms (2603.28489)
- OpenMOSS Awesome-WAM reading list
- EmergentMind: Necessity of Test-Time Future Imagination in WAMs