GAIA: A Data Flywheel System for Training GUI Test-Time Scaling Critic Models

A team led by Shaojie Zhang has proposed GAIA, a data flywheel framework designed to train critic models that improve the test-time performance of graphical user interface agents, according to a preprint posted on arXiv [1]. The system, formally titled the GUI Action Critic's Data Flywheel System, addresses a persistent weakness in Large Vision-Language Model (LVLM) driven GUI agents: the irreversibility of their operations, where a single mistaken action can cause a task to fail catastrophically [1][2]. The preprint was first submitted to the open-access repository on 26 January 2026 and revised on 26 June 2026 [1]. arXiv, which began in 1991, hosts e-prints across physics, computer science, and other fields and is moderated but not peer-reviewed [7]. GAIA operates by first training an Intuitive Critic Model (ICM) on positive and negative action examples collected from a base GUI agent [1][2]. This critic evaluates an agent's intended next action and selects the operation with the higher probability of success. The initial critic then guides the agent to gather refined samples, which are fed back into the training loop to produce a second-round critic with stronger discernment [1][2]. The authors report that this self-improving cycle allows performance to rise incrementally as data is recycled through the system [1][2]. The framework was tested across multiple datasets and demonstrated the ability to boost test-time results for both closed-source and open-source models [1][2]. The researchers have stated that the code, dataset, and a datasheet will be publicly released on GitHub [2]. The submission history shows the original manuscript weighed 4,318 KB, while the revised version was 3,950 KB [1]. The paper appears on arXiv as a standalone preprint, meaning it has not yet undergone formal peer review. The repository's moderation process approves papers for posting based on subject-matter relevance, not methodological validation [7]. The preprint's abstract page also surfaces a suite of community-built tools through arXivLabs, a framework launched in 2020 that lets third-party developers integrate experimental features such as citation explorers and code finders directly on article pages [5][6].