Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2505.21964

UI-Evol: Automatic Knowledge Evolving for Computer Use Agents

Ziyun Zhang·Xinyi Liu·Xiaoyi Zhang·Jun Wang·Gang Chen·Yan Lu
View original source ↗Read source document ↗Export metadata ↓

Abstract

External knowledge has played a crucial role in the recent development of computer use agents. We identify a critical knowledge-execution gap: retrieved knowledge often fails to translate into effective real-world task execution. Our analysis shows even 90% correct knowledge yields only 41% execution success rate. To bridge this gap, we propose UI-Evol, a plug-and-play module for autonomous GUI knowledge evolution. UI-Evol consists of two stages: a Retrace Stage that extracts faithful objective action sequences from actual agent-environment interactions, and a Critique Stage that refines existing knowledge by comparing these sequences against external references. We conduct comprehensive experiments on the OSWorld benchmark with the state-of-the-art Agent S2. Our results demonstrate that UI-Evol not only significantly boosts task performance but also addresses a previously overlooked issue of high behavioral standard deviation in computer use agents, leading to superior performance on computer use tasks and substantially improved agent reliability.

Explore related subjects

cs.HCcs.CL

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Ziyun Zhang, Xinyi Liu, Xiaoyi Zhang, Jun Wang, Gang Chen, Yan Lu. 2025-11-03. UI-Evol: Automatic Knowledge Evolving for Computer Use Agents. https://arxiv.org/abs/2505.21964

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

TrialCompass: Visual Analytics for Enhancing the Eligibility Criteria Design of Clinical Trials

Eligibility criteria play a critical role in clinical trials by determining the target patient population, which significantly influences the outcomes of medical interventions. However, current approaches for designing eligibility criteria have limitations to support interactive exploration of the large space of eligibility criteria. They also ignore incorporating detailed characteristics from the original electronic health record (EHR) data for criteria refinement. To address these limitations, we proposed TrialCompass, a visual analytics system integrating a novel workflow, which can empower clinicians to iteratively explore the vast space of eligibility criteria through knowledge-driven and outcome-driven approaches. TrialCompass supports history-tracking to help clinicians trace the evolution of their adjustments and decisions when exploring various forms of data (i.e., eligibility criteria, outcome metrics, and detailed characteristics of original EHR data) through these two approaches. This feature can help clinicians comprehend the impact of eligibility criteria on outcome metrics and patient characteristics, which facilitates systematic refinement of eligibility criteria. Using a real-world dataset, we demonstrated the effectiveness of TrialCompass in providing insights into designing eligibility criteria for septic shock and sepsis-associated acute kidney injury. We also discussed the research prospects of applying visual analytics to clinical trials.

cs.HC↗
arXiv2026

PRIMMDebug: Teaching Secondary School Students a Reflective Approach to Debugging

Debugging is a challenging and infuriating experience for many secondary school students learning their first text-based programming language. One frequent problem is the lack of reflection in students' debugging strategies, which makes error resolution unlikely and teacher reliance common. Tools that encourage more reflective and teacher-independent debugging may foster more success with fixing errors, but are lacking. This paper presents PRIMMDebug, an approach for teaching the debugging process to secondary school students. PRIMMDebug consists of an online tool that takes students through the steps of a pedagogical process based on PRIMM, a framework for teaching programming. The tool encourages written articulation throughout the debugging process and limits students' ability to run and edit code at certain stages. A classroom study with PRIMMDebug found a general reluctance among students to engage with the reflection it promotes, despite teachers appreciating this emphasis on reflection. We end by suggesting three considerations for future pedagogical debugging research and tooling: balance structure and flexibility, teach shorter debugging heuristics, and use tooling early on in students' programming journey.

cs.HC↗
arXiv2026

Mirror Skin: In Situ Visualization of Robot Touch Intent on Robotic Skin

Effective communication of robot touch intent is essential for safe and predictable physical human-robot interaction. While intent communication has been widely studied, existing approaches lack the spatial specificity and semantic depth necessary to efficiently convey robot touch intent. We present Mirror Skin, a cephalopod inspired concept that mirrors in-situ visual representations of a human's body parts onto the corresponding robot's touch region to communicate who shall initiate touch, where it will occur, and when it is imminent. We informed the design of Mirror Skin through a structured design exploration with experts and demonstrate the real-world feasibility of Mirror Skin with a proof-of-concept prototype. User studies in VR and with the physical prototype showed that Mirror Skin significantly improves accuracy and response times for interpreting touch intent and improves the user experience during physical human-robot interactions.

cs.HC↗