Description
Open models doing deep search give up on hard questions that need multi-step digging. DeepDive from Tsinghua's THUDM synthesizes hard search questions from knowledge graphs and trains deep search agents with multi-turn RL.
Trained models improve markedly on BrowseComp and more, with open code and data.
Data synthesis:Hard questions from graphs.
Multi-turn RL:Learns to dig.
Open:Code and data.
Trained models improve markedly on BrowseComp and more, with open code and data.
Features
Data synthesis:Hard questions from graphs.
Multi-turn RL:Learns to dig.
Open:Code and data.
