Web Navigation and E-commerce on WebShop In-Distribution
94.73Average ScoreSkillFlow
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SkillFlowMethod Category=Ours2026.05 | 94.73 | 93.75 | |
| SkillRLMethod Category=Agent+RL2026.05 | 87.96 | 82.81 | |
| FlowSteerMethod Category=Agent+RL2026.05 | 85.43 | 79.2 | |
| v4-flashMethod Category=Baseline2026.05 | 85.17 | 67.97 | |
| AgentFlowMethod Category=Agent+RL2026.05 | 83.61 | 74.22 | |
| Task-Perturbed NLL OptimizationBackbone=Qwen3-8B2026.06 | 81 | — | |
| Task-Perturbed NLL OptimizationBackbone=Qwen3-4B2026.06 | 77 | — | |
| GRPOBackbone=Qwen3-8B2026.06 | 76 | — | |
| GRPOBackbone=Qwen3-4B2026.06 | 74 | — | |
| Qwen3.5-AFlowMethod Category=AFlow, Backbone=Qwen3.52026.05 | 71.09 | 59.38 | |
| Qwen3.5-9BMethod Category=Baseline2026.05 | 56.93 | 32.03 | |
| Qwen3.5-GRPOMethod Category=GRPO, Backbone=Qwen3.52026.05 | 54.1 | 36.71 | |
| Qwen3.5-SFTMethod Category=SFT, Backbone=Qwen3.52026.05 | 50.8 | 35.16 |