List your company

Workflow Reconstruction on TEACHMASTER (test)

0.124SFESFT
0.1130.187250.26150.33575Feb 5, 2026

Evaluation results

MethodLinks
SFTProtocol=Direct input-output cloning2026.020.124
Claude Opus 4.5Protocol=Multi-turn self-refinement2026.020.303
ReAct (Claude Opus 4.5)Protocol=ReAct-style tool use2026.020.322
AFlowProtocol=MCTS-based workflow search, Red-Black Pruning=Disabled2026.020.348
AgentXRay w/o ToolsAblation=Without Tools2026.020.357
AgentXRay w/o PruningAblation=Without Pruning2026.020.378
AgentXRay (All Tools)Config=Full primitive space2026.020.39
AgentXRay (Selected)Config=Selected primitive set2026.020.399