ScratchWorld - Extend: leaderboard
Extend tasks from ScratchWorld (Zhang et al., 2026): add a new feature to an existing Scratch project without breaking its original behavior, checked by unit tests run in the Scratch VM. Composite mode, where the agent reads the program as pseudocode and edits it through semantic commands such as adding or connecting blocks instead of the GUI. The score is the percentage of the 18 tasks on which every unit test passes; higher is better.
Metric: Success Rate (%). Source: huggingface.co. Saturation forecast: Around January 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 83.33 | #142 |
| 2 | Gemini 2.5 Pro | 72.22 | #152 |
| 3 | GPT-5 | 66.67 | #98 |
| 4 | DeepSeek V3.2 | 66.67 | #205 |
| 5 | Qwen 2.5 Coder 7B Instruct | 0 | #951 |
| 6 | Qwen 3 Coder 30B A3B Instruct | 0 | #415 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=scratchworld-extend · How It Works · Data refreshed daily, snapshot 2026-10-20.