AI Paper Atlas

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

Learning from Failure: Inference-Time Self-Improvement for Computer-Use Agents

Xueqiao Sun, Yuhui Zhang, Xiaohan Wang, Ludwig Schmidt, Serena Yeung-Levy

ECCV 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 7 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add inference-time failure-derived patches (full method)OpenCUA-72B · OSWorld (100 steps)success rate (%): 42.3 -> 48.9Improved
add visual search post-action verification (ablation)OpenCUA-72B · OSWorld small setsuccess rate (%): 41.67 -> 47.22Improved
add terminal execution (ablation)OpenCUA-72B · OSWorld small setsuccess rate (%): 41.67 -> 47.19Improved
add knowledge support (manual retriever and search) (ablation)OpenCUA-72B · OSWorld small setsuccess rate (%): 41.67 -> 44.44Improved
add repetition detection and recovery mode (ablation)OpenCUA-72B · OSWorld small setsuccess rate (%): 41.67 -> 44.40Improved
combine all four strategies (ablation)OpenCUA-72B · OSWorld small setsuccess rate (%): 41.67 -> 52.74Improved
apply same loop to OpenCUA-32BOpenCUA-32B · OSWorld (100 steps)success rate (%): 34.5 -> 38.2Improved
apply same loop to GUI-Owl-32BGUI-Owl-32B · OSWorld (100 steps)success rate (%): 19.0 -> 21.3Improved
transfer OSWorld-derived patches to AndroidControlQwen3-VL-32B-Instruct · AndroidControlscore: 28.37 -> 36.23Improved
transfer OSWorld-derived patches to WebVoyagerQwen3-VL-32B-Instruct · WebVoyagersuccess rate: 23.80 -> 27.90Improved
transfer OSWorld-derived patches to ScreenSpotProQwen3-VL-32B-Instruct · ScreenSpotProaccuracy: 27.50 -> 30.74Improved

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