PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces
The paper’s central claim is that deep-research agents perform better when their work is split into planned subquestions, stored state, and evidence checks before reuse.
PANINI-DR uses a shared Generative Semantic Workspace to track subquestions, dependencies, findings, and supporting evidence. Separate agents handle research and verification, with feedback from verification steering later investigation. In the authors’ tests on BrowseComp-Plus, FRAMES, and MoNaCo, it posted the strongest average performance among eight harnesses using the same model backbone and frozen corpora. The paper reports 73.3% accuracy on BrowseComp-Plus, 16 points ahead of the next-best harness, with the largest gains on hard questions. HF Daily Papers' note
PANINI-DR uses a shared Generative Semantic Workspace to track subquestions, dependencies, findings, and supporting evidence. Separate agents handle research and verification, with feedback from verification steering later investigation. In the authors’ tests on BrowseComp-Plus, FRAMES, and MoNaCo, it posted the strongest average performance among eight harnesses using the same model backbone and frozen corpora. The paper reports 73.3% accuracy on BrowseComp-Plus, 16 points ahead of the next-best harness, with the largest gains on hard questions. HF Daily Papers' note
score 5