Dimensions overall score 9.0

GitHub Code Pulse

No public code linked for this paper yet.

• Environment Pranjal Aggarwal CMU Graham Neubig CMU Sean Welleck CMU https://cmu-l3.github.io/gym-anything Abstract Computer-use agents hold the promise of assisting in a wide range of digital eco- nomic activities. Howeverclaim 72%
• We useGym-Anythingto construct CUA-World, a collection of over 10,000 tasks across 200 software applications. CUA-Worldspans domains such as medical science, astronomy, engineering, finance, enterprise systems, and educational platforms. It includes tasks on three operating systemsclaim 72%
• hundreds of steps; (3) training and test-time scaling results, including distillation to a 2B model that outperforms models 2× its size, and a test-time audit agent that improves long-horizon performance onCUA-World-Long; and (4) a full release of all code, infrastructure, and benchmark dataclaim 72%
• discovery 3,400 Selectable sandbox-ready softwares Filter: free, GUI, self-hostable, no hardware 500 Selected across 22 SOC groups 200 Built compute budget k2b: STEM & Research k2a: Strategic Health, Ed, Safetyclaim 72%
• mapping Cleanup + Enrichment 5-tier Selection Build Environments Figure 3:GDP-grounded software selection pipeline.Starting from U.S. occupational data, we estimate per-software GDP, filter to sandboxable candidates, and apply tiered selection to yield 200 software applications. 2 Methodology In this sectionclaim 72%
• occ GDPocc ×p computer ×s category ×s software (1) Here, pcomputer is the fraction of the occupation’s work that involves computers (available from occupational surveys), scategory is the share of that computer work attributed to a software category (e.g., “spreadsheets” for an accountant)claim 72%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card