Web-based Sequential Decision Making on WebShop (hundred-product catalog setup)
90.2Mean RewardMAGE Sonnet–Qwen3-8B (n=5)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MAGE Sonnet–Qwen3-8B (n=5)2026.05 | 90.2 | 84 | |
| MAGE Opus–Qwen3-8B (n=5)2026.05 | 89.7 | 82 | |
| RetroAgent (RL-trained reflection)2026.05 | 88.9 | 82.3 | |
| MAGE Haiku–Qwen3-8B2026.05 | 88.3 | 77 | |
| RetroAgent (in-context, Qwen-7B reference)2026.05 | 87.6 | 78.9 | |
| MAGE (any teacher)–Llama-3.1-8B2026.05 | 0 | 0 |