List your company

Workflow Reconstruction on CHATDEV (test)

50.9SFEAgentXRay (Selected)
24.58831.41938.2545.081Feb 5, 2026

Evaluation results

MethodLinks
AgentXRay (Selected)Config=Selected primitive set2026.0250.9
AgentXRay (All Tools)Config=Full primitive space2026.0242.5
AgentXRay w/o ToolsAblation=Without Tools2026.0241.3
AFlowProtocol=MCTS-based workflow search, Red-Black Pruning=Disabled2026.0240.3
SFTProtocol=Direct input-output cloning2026.0235.5
AgentXRay w/o PruningAblation=Without Pruning2026.0228.6
ReAct (Claude Opus 4.5)Protocol=ReAct-style tool use2026.0226.7
Claude Opus 4.5Protocol=Multi-turn self-refinement2026.0225.6