Workflow Reconstruction on CHATDEV (test)
50.9SFEAgentXRay (Selected)
Evaluation Results
| Method | Links | |
|---|---|---|
| AgentXRay (Selected)Config=Selected primitive set2026.02 | 50.9 | |
| AgentXRay (All Tools)Config=Full primitive space2026.02 | 42.5 | |
| AgentXRay w/o ToolsAblation=Without Tools2026.02 | 41.3 | |
| AFlowProtocol=MCTS-based workflow search, Red-Black Pruning=Disabled2026.02 | 40.3 | |
| SFTProtocol=Direct input-output cloning2026.02 | 35.5 | |
| AgentXRay w/o PruningAblation=Without Pruning2026.02 | 28.6 | |
| ReAct (Claude Opus 4.5)Protocol=ReAct-style tool use2026.02 | 26.7 | |
| Claude Opus 4.5Protocol=Multi-turn self-refinement2026.02 | 25.6 |