Workflow Reconstruction on METAGPT (test)
55.7SFEAgentXRay (All Tools)
Evaluation Results
| Method | Links | |
|---|---|---|
| AgentXRay (All Tools)Config=Full primitive space2026.02 | 55.7 | |
| AgentXRay (Selected)Config=Selected primitive set2026.02 | 47 | |
| AgentXRay w/o PruningAblation=Without Pruning2026.02 | 33.4 | |
| ReAct (Claude Opus 4.5)Protocol=ReAct-style tool use2026.02 | 33.1 | |
| Claude Opus 4.5Protocol=Multi-turn self-refinement2026.02 | 32.2 | |
| AgentXRay w/o ToolsAblation=Without Tools2026.02 | 30.1 | |
| AFlowProtocol=MCTS-based workflow search, Red-Black Pruning=Disabled2026.02 | 28 | |
| SFTProtocol=Direct input-output cloning2026.02 | 27.2 |