Agentic Orchestration on GAIA
80AccuracyAORCHESTRA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| AORCHESTRALM=Gemini-3-Flash2026.02 | 80 | 0.79 | |
| AORCHESTRA (ICL)Optimization=In-context Learning, LM Pool=Mixed2026.02 | 75.15 | 0.57 | |
| AORCHESTRA (Gemini-3-Flash Orchestrator)LM Pool=Mixed2026.02 | 72.12 | 0.7 | |
| AORCHESTRALM=Claude-4.5-sonnet2026.02 | 71.52 | 0.91 | |
| AORCHESTRA (Qwen3-8B SFT Orchestrator)Training=Supervised Fine-Tuning (SFT), LM Pool=Gemini-3-Flash2026.02 | 68.48 | 0.68 | |
| AORCHESTRALM=Deepseek-v3.22026.02 | 67.87 | 0.14 | |
| AORCHESTRALM=GPT-5-mini2026.02 | 67.27 | 0.28 | |
| AORCHESTRA (Qwen3-8B Orchestrator)LM Pool=Gemini-3-Flash2026.02 | 56.97 | 0.36 | |
| ReActLM=GPT-5-mini2026.02 | 54.55 | 0.052 | |
| ReActLM=Claude-4.5-sonnet2026.02 | 53.93 | 0.19 | |
| ReActLM=Gemini-3-Flash2026.02 | 49.09 | 0.07 | |
| ReActLM=Claude-4.5-haiku2026.02 | 47.88 | 0.066 | |
| ReActLM=Deepseek-v3.22026.02 | 46.7 | 0.027 |