Add WAM paper: HALO-WA: Hybrid-Attention Latent-Guided Online Reinforcement Learning for World-Action Models

[279object]

Paper

• Title: HALO-WA: Hybrid-Attention Latent-Guided Online Reinforcement Learning for World-Action Models
• Short name: HALO-WA
• arXiv ID: 2607.04265v1
• Paper: https://arxiv.org/pdf/2607.04265
• Authors: Angen Ye, Weijie Ke, Xiaofeng Wang, Xinze Chen, Chaojun Ni, Guosheng Zhao, Boyuan Wang, Zheng Zhu, Junjie Xie, Dapeng Zhang
• Published: 2026-07-05
• arXiv categories: cs.RO, cs.AI
• Matched keywords: world-action model, world-action models

README Entry

- **HALO-WA**: "HALO-WA: Hybrid-Attention Latent-Guided Online Reinforcement Learning for World-Action Models", arXiv 2026.
  [[📄 Paper](https://arxiv.org/pdf/2607.04265)]

Robot Decision

• Decision: accept
• Confidence: high
• Taxonomy: world_model_for_reinforcement_learning
• Taxonomy confidence: medium
• README heading: ### World Model for Reinforcement Learning
• Badges: None

Reason

The paper explicitly develops and evaluates an online RL adaptation framework for world-action models in robotic manipulation.

Evidence

• World-action models generate long-horizon action chunks for robotic manipulation
• HALO-WA refines a frozen WA backbone using actor-critic online reinforcement learning
• Evaluated on real-world precision manipulation and RoboTwin simulation tasks

Taxonomy Evidence

• hybrid-attention latent-guided online reinforcement learning framework
• freezes the pretrained WA backbone and trains a lightweight actor-critic adapter
• leverages latent features and action priors from the WA generation process
• producing refined action chunks for real-world precision manipulation

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