List your company

Workflow Reconstruction on SCI (test)

0.139SFESFT
0.128280.200640.2730.34536Feb 5, 2026

Evaluation results

MethodLinks
SFTProtocol=Direct input-output cloning2026.020.139
Claude Opus 4.5Protocol=Multi-turn self-refinement2026.020.299
ReAct (Claude Opus 4.5)Protocol=ReAct-style tool use2026.020.305
AgentXRay w/o PruningAblation=Without Pruning2026.020.312
AFlowProtocol=MCTS-based workflow search, Red-Black Pruning=Disabled2026.020.373
AgentXRay w/o ToolsAblation=Without Tools2026.020.378
AgentXRay (All Tools)Config=Full primitive space2026.020.395
AgentXRay (Selected)Config=Selected primitive set2026.020.407