# VisCritic: Visual State Comparison as Process Reward for GUI Agents

Source: [VisCritic: Visual State Comparison as Process Reward for GUI Agents](<https://www.sota2.com/research/paper/6a3b64d8206542251d7c993a>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-06-25T05:00:34.132Z

Authors: Jiachen Qian

Published: 2026-06-23T12:57:02.000Z

## Abstract

GUI agents powered by vision-language models show strong potential for automating digital tasks, yet frequently fail in long-horizon scenarios due to the absence of step-level verification. Existing process reward models verify actions through textual reasoning alone, missing the visual nature of GUI state changes. We introduce VisCritic, a visual process reward framework that verifies agent actions by directly comparing pre-action and post-action screenshots in visual feature space. VisCritic employs a Siamese vision transformer to extract change-aware representations, coupled with an Action-Aware Critic Head that jointly evaluates action success, task progress, and error type. A critic-training data construction pipeline generates weakly supervised samples from existing trajectories without additional human labels for critic training. Experiments and offline analyses across five benchmarks demonstrate that VisCritic serves as a plug-and-play enhancement for diverse GUI agents, generally improving benchmark metrics while providing visual diagnostic cues.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Offline Action Consistency Diagnostic on AITW](<https://www.sota2.com/research/sota/offline-action-consistency-diagnostic-on-aitw.md>): Offline Action Consistency Diagnostic; AITW; Action Accuracy: 74.2.

- [Action Success Prediction on AndroidWorld (smallest held-out split)](<https://www.sota2.com/research/sota/action-success-prediction-on-androidworld-smallest-held-out-split.md>): Action Success Prediction; AndroidWorld (smallest held-out split); Precision: 87.1.

- [Web Agent on WebArena](<https://www.sota2.com/research/sota/web-agent-on-webarena.md>): Web Agent; WebArena; Success Rate: 29.4.

- [Mobile UI Control on AndroidWorld](<https://www.sota2.com/research/sota/mobile-ui-control-on-androidworld.md>): Mobile UI Control; AndroidWorld; Overall Task Success Rate: 29.8.

- [Web Navigation on Mind2Web](<https://www.sota2.com/research/sota/web-navigation-on-mind2web.md>): Web navigation; Mind2Web; Overall Success Rate: not recorded.

- [GUI Agent Interaction on OSWorld (Step Success Rate)](<https://www.sota2.com/research/sota/gui-agent-interaction-on-osworld-step-success-rate.md>): GUI Agent Interaction; OSWorld; Success Rate (Max Steps: 15): not recorded.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2606.24525>)

[PDF](<https://arxiv.org/pdf/2606.24525.pdf>)