ChainWorld (Multi Turn): leaderboard
ChainWorld (Siu et al., Scale AI and UC Berkeley, 2026) chains two to four compatible OSWorld desktop tasks into one long workload (347 chains) and keeps each task's original evaluator, so a computer-use agent must complete several objectives in one Ubuntu session whose state carries over, acting through pyautogui on full-screen screenshots. In this protocol the tasks are revealed one per turn, each with 75 steps, without resetting the desktop between turns. The score is chain completion, the fraction of chains (0 to 1) in which every task's evaluator passes; higher is better.
Metric: Chain completion (fraction of 347 chains with every task's evaluator passing; tasks revealed one per turn, 75 steps each, state kept across turns). Source: arxiv.org. Saturation forecast: Around October 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 0.31 | #48 |
| 2 | Claude Opus 4.7 | 0.17 | #75 |
| 3 | Kimi K2.5 | 0.1 | #252 |
| 4 | Gemini 3.1 Pro (Preview) | 0.02 | #91 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=chainworld-multi-turn · How It Works · Data refreshed daily, snapshot 2026-11-12.