MM-ToolSandBox partial credit (LFM3 report)
Agents
MM-ToolSandBox (Apple, 2026) evaluates visual tool-calling agents on human-verified multi-turn scenarios with several images each and a simulated user, scoring a binary Agent Success Rate with a rubric judge and graded Entity F1 partial credit from environment state changes. Liquid AI's LFM3 card reports a 'partial credit' score without defining it or the scenario set and uses GPT-6 Luna as its judge, so the rows are kept as a source-specific diagnostic.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.5 Flash-Lite | 62.4 |
| GPT-5.6 Luna | 56.1 |
| Gemma 4 12B IT | 43.6 |
| LFM3-VL-2B | 42.5 |
| Qwen3.5-4B | 31.3 |
| Gemma 4 E2B IT | 23.8 |
| Qwen3.5-2B | 15.2 |