AI Paper Atlas

Research map / Multi-agent systems and games / LLM agents

PSBench: Editing Image via GUI Agents in Photoshop

Yinuo Zhang, Zian Cheng, Ziya Zhao, Zongyu Li, Bingshuo Liu, Qingbin Liu, Junxian Cai, Xi Chen, Zhiying Tu, Dianhui Chu, Xiaoyan Yu, Dianbo Sui

ICML 2026 · Multi-agent systems and games › LLM agents · approach: GUI agent memory and trajectories

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces PSBench, a benchmark for GUI agents performing image editing tasks inside Adobe Photoshop, targeting multi-step, layer-based professional workflows.

Problem

Existing GUI agent benchmarks and methods focus on web interfaces and short, low-complexity tasks, and do not model the multi-level menus, fine-grained operations, and layer-based non-destructive editing required by professional image editing software like Photoshop.

Approach

A Photoshop-specific GUI agent benchmark that emphasizes long-horizon, multi-step professional editing with layer-based operations and semantic understanding, rather than web-based short tasks.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use hierarchical agentic framework (GPT-5 planning plus UI execution) instead of a single general MLLMGPT-4o / Agent S3 · PSBench hard taskssuccess rate (%): 18.09 on hard tasks for Agent S3Improved
use GUI-specialized models instead of general-purpose MLLMsgeneral-purpose MLLMs · PSBench non-layer-related taskssuccess rate (%): 17.46 -> 42.06Improved
increase task difficulty from easy to hardGPT-4o · PSBenchsuccess rate (%): easy LR 16.82, hard LR 0.00Hurt
layer-related instead of non-layer-related tasksGPT-4o · PSBench overallsuccess rate (%): 17.46 -> 3.80Hurt
use specialized GUI models instead of general MLLMs, effect on NDECUI-TARS-2-2509 / OpenCUA-7B · PSBenchNDEC (%): 63.22 and 61.78 vs 78.24 for GPT-4oHurt
increase task difficulty from easy to hardMLLMs · PSBenchNDEC (%): 95.83 -> 50-67Hurt
GUI assistant mode (GPT-4o gives instructions, human executes) instead of autonomous GPT-4o agentGPT-4o · PSBench easy tasks (human-in-loop)success rate (%): 17.5 -> 100Improved
GUI assistant mode instead of autonomous GPT-4o agentGPT-4o · PSBench hard tasks (human-in-loop)success rate (%): 0 -> 75Improved
GUI assistant mode instead of unassisted novicenovice user · PSBench easy tasks (human-in-loop)success rate (%): 85 -> 100Improved
GUI assistant mode instead of unassisted novicenovice user · PSBench medium tasks (human-in-loop)success rate (%): 45 -> 85Improved
GUI assistant mode instead of internet accessnovice user · PSBench hard tasks (human-in-loop)completion time (s): 720 -> 330Improved
GUI assistant mode instead of unassisted novicenovice user · PSBench easy tasks (human-in-loop)completion time (s): 180 -> 45Improved

Experimental setup

Benchmarks: PSBench

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy