PANINI-DR / PANINI / Generative Semantic Workspace

Generative Semantic Workspaces (GSW) are entity- and event-aware structured memories built from atomic question–answer (QA) pairs. This research thread uses them for deep-research agents (PANINI-DR) and for non-parametric continual learning (PANINI).

PANINI-DR: controlled deep research

Paper (arXiv:2610.15208) · Code · Project page

PANINI-DR is a multi-agent deep-research harness that separates coordination of the overall investigation from the research and verification of individual subquestions.

• Shared GSW workspace: records subquestions, their dependencies, findings, and supporting evidence, so agents work from explicit research state instead of long chat histories.
• Verification agent: checks proposed answers against their subquestions and evidence; verification feedback guides further investigation.
• Results: strongest average performance on BrowseComp-Plus, FRAMES, and MoNaCo against seven baselines with a shared language-model backbone and frozen corpora. 73.3% accuracy on BrowseComp-Plus, 16.0 points above the next-best baseline, and 26 of 50 hard questions answered correctly, nearly twice the next-best baseline. It also achieves the highest answer recoverability from saved evidence on all three benchmarks.

Under review at ICLR 2027. Co-first authors M. Y. Turali and S. Rajesh, with M. E. Lorasdagi, C. Duan, K. Sharma, and V. Roychowdhury.

PANINI: continual learning in token space

Paper (arXiv:2602.15156) · Code

PANINI is a non-parametric continual-learning framework that encodes documents into GSWs, enabling reasoning-grounded inference over stored experience without continual parameter updates.

• Reasoning Inference Chain Retrieval (RICR): a beam-search procedure that performs one-shot query decomposition and follows multi-hop inference chains through GSWs to assemble compact evidence.
• Dual indexing for scalable memory access: a BM25 index over entity + role/state descriptors paired with a dense vector index over QA pairs, enabling targeted retrieval of QA evidence rather than long text chunks.
• Results: top average performance across six QA benchmarks with 2–30× fewer answer-context tokens than competitive baselines, and fewer unsupported answers on curated unanswerable queries.

Accepted as a poster at ICML 2026. Co-authored with S. Rajesh, P. S. Holur, C. Duan, and V. Roychowdhury.