# World Action Verifier: Self-Improving World Models via Forward-Inverse Asymmetry

**Authors:** Yuejiang Liu, Fan Feng, Lingjing Kong, Weifeng Lu, Jinzhou Tang, Kun Zhang, Kevin Murphy, Chelsea Finn, Yilun Du  
**arXiv:** [2604.01985](https://arxiv.org/abs/2604.01985)  
**Date:** 2026-04-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2604.01985v2/

## Key points

- Development of general-purpose world models in AI and robotics enables scalable policy evaluation, optimization, and planning
- Challenge of achieving necessary robustness for world models due to vast array of suboptimal actions
- Introduction of World Action Verifier (WAV) framework to address critical issue
- WAV empowers world models to identify prediction errors and self-improve over time
- Key concept behind WAV is decomposition of action-conditioned state prediction into verifiable factors: state plausibility and action reachability
- WAV demonstrates improved tractability compared to direct forward prediction methods by verifying factors separately
- Success attributed to inherent imbalances: wider availability of action-free data and lower dimensionality of action-relevant features
- Framework enhances world model with diverse subgoal generator and sparse inverse model for improved performance
- Extensive testing across nine diverse tasks showcases remarkable performance improvements achieved by WAV, including doubling sample efficiency and boosting downstream policy performance by over 22%
- Research represents significant advancement in AI-driven robotics field by addressing challenges related to robustness and reliability in world modeling systems

## Summary

In the realm of artificial intelligence and robotics, the development of general-purpose world models has shown great promise in enabling scalable policy evaluation, optimization, and planning. However, a significant challenge lies in achieving the necessary level of robustness for these models. While traditional policy learning methods tend to concentrate on identifying optimal actions, world models must be able to navigate through a vast array of suboptimal actions that are often overlooked in action-labeled robot interactions. To address this critical issue, a groundbreaking framework known as World Action Verifier (WAV) has been introduced. This innovative approach empowers world models to not only identify their own prediction errors but also self-improve over time. The key concept behind WAV is the decomposition of action-conditioned state prediction into two distinct and independently verifiable factors: state plausibility and action reachability. By verifying these factors separately, WAV demonstrates significantly improved tractability compared to direct forward prediction methods. The success of WAV can be attributed to two fundamental asymmetries present in the process: the wider availability of action-free data and the lower dimensionality of action-relevant features. Leveraging these inherent imbalances, the framework enhances a world model with a diverse subgoal generator sourced from video corpora and a sparse inverse model capable of inferring actions from select state features. Through enforcing cycle consistency among proposed subgoals, inferred actions, and forward rollouts, WAV establishes an effective verification mechanism particularly beneficial in under-explored regimes where existing methodologies often fall short. Extensive testing across nine diverse tasks spanning MiniGrid, RoboMimic, and ManiSkill showcases the remarkable performance improvements achieved by WAV. Notably, the method demonstrates twice the sample efficiency while boosting downstream policy performance by an impressive margin exceeding 22%. This research represents a significant advancement in the field of AI-driven robotics by offering a solution that addresses crucial challenges related to robustness and reliability in world modeling systems.