University of California, Los Angeles
*Equal contribution
Deep research requires connecting multiple steps of search, reasoning, and computation, with intermediate findings guiding what to investigate next. As investigations grow, agents must maintain research state across model calls and check findings before errors propagate into subsequent work. We introduce PANINI-DR, a multi-agent harness that separates coordination of the overall investigation from the research and verification of individual subquestions. A shared workspace, built on the Generative Semantic Workspace representation, records subquestions, their dependencies, findings, and supporting evidence. Researchers investigate individual subquestions, while a separate verification agent checks proposed answers against those subquestions and their evidence. Verification feedback guides further investigation, and recorded findings inform subsequent assignments. We evaluate PANINI-DR on BrowseComp-Plus, FRAMES, and MoNaCo against seven state-of-the-art deep-research harnesses, all using a shared language-model backbone and frozen corpora. PANINI-DR achieves the strongest average performance across the three benchmarks. On BrowseComp-Plus, it achieves 73.3% accuracy, outperforming the next-best harness by 16.0 points. Its advantage is strongest on hard questions, where it answers nearly twice as many as the next-best harness. Ablations further show that structured planning and persistent state improve performance, as does checking each finding against its cited evidence before it is reused. Because a model can answer correctly from memory alone, we also analyze the evidence retained by each harness. PANINI-DR achieves the highest answer recoverability from saved evidence on all three benchmarks, while complementary claim-support audits show stronger evidence-backed answers. Accounting for this evidence changes how harnesses compare and further strengthens PANINI-DR's advantage.
PANINI-DR is built on two ideas. An Orchestrator coordinates the investigation; Researchers and a Verifier work on one subquestion at a time.
A shared workspace, built on the Generative Semantic Workspace representation, records the subquestions, how they depend on each other, the findings, and the evidence behind them. The Orchestrator reads it to decide what is ready to investigate.
Researchers propose answers to single subquestions with cited evidence. A separate Verifier checks each proposal against its subquestion and evidence; rejections send the Researcher back to search. Only admitted findings feed the subquestions that depend on them, and calculations run in code.
We compare PANINI-DR with seven state-of-the-art deep-research harnesses. All of them run on the same language model (gpt-oss-120b), the same frozen corpora and the same retrieval service, so the differences come from the harnesses.
| BCP CF-150 accuracy (%) | FRAMES | MoNaCo | Evidence-backed, Obs. | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Harness | Easy | Med. | Hard | Overall | Acc. | F1 | Avg. | BCP | FRAMES | MoNaCo | Avg. |
| PyRAG | 56 | 20 | 6 | 27.3 | 61.9 | 0.446 | 44.6 | 26.0 | 46.6 | 0.332 | 35.3 |
| Re-TRAC | 70 | 32 | 16 | 39.3 | 64.9 | 0.544 | 52.9 | 14.0 | 46.1 | 0.396 | 33.2 |
| Omni-Decision | 84 | 48 | 10 | 47.3 | 64.6 | 0.359 | 49.3 | 45.3 | 53.9 | 0.275 | 42.2 |
| HyMem | 82 | 52 | 12 | 48.7 | 70.8 | 0.612 | 60.2 | 46.7 | 36.7 | 0.388 | 40.7 |
| Open Deep Search | 82 | 50 | 28 | 53.3 | 76.0 | 0.571 | 62.1 | 51.3 | 47.6 | 0.325 | 43.8 |
| IterResearch | 84 | 60 | 20 | 54.7 | 70.5 | 0.564 | 60.5 | 51.3 | 45.9 | 0.390 | 45.4 |
| SLEUTH | 86 | 58 | 28 | 57.3 | 77.2 | 0.556 | 63.4 | 54.0 | 55.2 | 0.396 | 49.6 |
| PANINI-DR | 92 | 76 | 52 | 73.3 | 79.1 | 0.549 | 69.1 | 70.7 | 66.1 | 0.458 | 60.9 |
Accuracy is official accuracy on BrowseComp-Plus (BCP, CF-150 split; 50 questions per difficulty) and FRAMES-824, and mean item-level F1 on MoNaCo-1315. Obs. keeps official credit only when the credited answer can be recovered from the evidence the harness saved. Averages weight the three benchmarks equally, with MoNaCo rescaled to 0-100. Bold: best in each column.
Hugging Face@article{turali2026paninidr,
title = {{PANINI-DR}: Controlled Deep Research through Generative Semantic Workspaces},
author = {Turali, M. Yigit and Rajesh, Shreyas and Lorasdagi, M. Efe and Duan, Chenda and Sharma, Kartik and Roychowdhury, Vwani},
journal = {arXiv preprint arXiv:2610.15208},
url = {https://arxiv.org/abs/2610.15208},
year = {2026}
}
PANINI-DR is a multi-agent deep-research harness for answering hard multi-step questions over a document collection. An Orchestrator coordinates the investigation through a shared research workspace, built on the Generative Semantic Workspace (GSW) representation, that records subquestions, their dependencies, findings and evidence. Researcher agents investigate individual subquestions, and a separate Verifier checks each proposed answer against its subquestion and cited evidence before it is reused.
No. They are related work from the same lab. GSW (Generative Semantic Workspaces, AAAI 2026) is the structured representation that PANINI-DR's research workspace is built on. Panini (arXiv:2602.15156, ICML 2026) uses GSW for continual learning in token space; the code for GSW and Panini is at github.com/roychowdhuryresearch/gsw-memory. PANINI-DR (arXiv:2610.15208) is a separate deep-research harness for multi-step question answering; its code is at github.com/roychowdhuryresearch/panini-dr.
On BrowseComp-Plus (CF-150 split) PANINI-DR reaches 73.3% accuracy, 16.0 points above the next-best of seven state-of-the-art deep-research harnesses (SLEUTH, 57.3%), with all harnesses using the same language model (gpt-oss-120b), frozen corpora and retrieval service. On the hard questions it answers 52%, against 28% for the next best.
PANINI-DR has the best average accuracy across BrowseComp-Plus, FRAMES and MoNaCo (69.1, against 63.4 for SLEUTH). It also has the highest observed support, the share of answer credit that can be recovered from the evidence each harness saved, on all three benchmarks (average 60.9, against 49.6).
Two ideas: research state is kept explicitly in a shared workspace instead of only in a growing conversation, and every intermediate finding is verified against its evidence before later steps use it. Ablations in the paper show that both the plan-and-state structure and verification improve accuracy.
Instead of relying on a long, growing conversation, PANINI-DR keeps the investigation in a shared workspace that pairs the plan (the subquestions and how they depend on each other) with the admitted findings, each with its evidence and a review status (verified, warned or derived). The Orchestrator reads this workspace to decide which subquestions are ready, and only admitted findings are passed to the subquestions that depend on them.
Run panini-dr verify. It replays recorded model, reranker and embedding responses from FRAMES questions through your installation, on a CPU with no network, and requires its answers and traces to match those of the code behind the paper exactly. Reproducing the paper's numbers themselves needs the model servers and the benchmark data; REPRODUCE.md in the repository gives the commands, and live runs vary slightly because the LLM server is not bit-deterministic.
Write your questions as questions.jsonl and your documents as docs.jsonl, run panini-dr prepare to build the corpus and its search indexes, then panini-dr run on the prepared folder and panini-dr score to grade the answers. The repository README has the exact formats and commands.
Yes. The code is open source under the Apache-2.0 licence at github.com/roychowdhuryresearch/panini-dr. It includes the paper configuration, the graders, benchmark preparation scripts, support for running on your own dataset, and panini-dr verify, a CPU-only check that an installation reproduces the paper code's decisions exactly.
M. Yigit Turali and Shreyas Rajesh (equal contribution), M. Efe Lorasdagi, Chenda Duan, Kartik Sharma and Vwani Roychowdhury, University of California, Los Angeles. The paper is arXiv:2610.15208.