Add WAM paper: Understanding and Mitigating the Video-Action Generalization Gap via Temporal Ratio

[279object]

Paper

• Title: Understanding and Mitigating the Video-Action Generalization Gap via Temporal Ratio
• Short name: UAMTV
• arXiv ID: 2607.08127v1
• Paper: https://arxiv.org/pdf/2607.08127
• Authors: Utkarsh A. Mishra, Yongxin Chen, Danfei Xu, Yang Liu, Xi Chen, Jiayuan Mao
• Published: 2026-07-09
• arXiv categories: cs.CV, cs.RO
• Matched keywords: world-action model, world-action models

README Entry

- **UAMTV**: "Understanding and Mitigating the Video-Action Generalization Gap via Temporal Ratio", arXiv 2026. ![](https://img.shields.io/badge/Implicit-ea580c)
  [[📄 Paper](https://arxiv.org/pdf/2607.08127)]

Robot Decision

• Decision: accept
• Confidence: high
• Taxonomy: cascaded_implicit_latent
• Taxonomy confidence: medium
• README heading: ### Cascaded World-Action-Model
• Badges: https://img.shields.io/badge/Implicit-ea580c

Reason

Clearly studies and improves video/world-action models for robotic action prediction and control using future latent rollouts.

Evidence

• explicitly discusses World-Action Models (WAMs) and Video-Action Models (VAMs)
• action head uses predicted future video latents for robotic manipulation on LIBERO and real-world tasks

Taxonomy Evidence

• action head relies on future latent rollouts relative to current frame
• latent features from pre-trained video foundation models and learn a latent inverse-dynamics action head
• partially denoised latent features are projected into the action head for action denoising
• evaluated on LIBERO benchmark and real-world tasks

Human Review Checklist

• This paper belongs in the WAM survey
• README section is correct
• Badges are correct
• Short name is correct
• Paper link is correct
• Merge this PR if accepted