# PANINI-DR

> PANINI-DR is a multi-agent deep-research harness for answering hard multi-step questions over a document collection. It keeps the investigation's state in an explicit shared workspace (built on the Generative Semantic Workspace representation) and has a separate Verifier check every intermediate finding against its cited evidence before it is reused. Paper: "PANINI-DR: Controlled Deep Research through Generative Semantic Workspaces" (arXiv:2610.15208), University of California, Los Angeles.

Key results (all harnesses use gpt-oss-120b, the same frozen corpora and the same retrieval service):

- BrowseComp-Plus (CF-150): 73.3% accuracy, 16.0 points above the next-best of seven state-of-the-art deep-research harnesses (SLEUTH, 57.3%); 52% on hard questions against 28% for the next best.
- Best average accuracy across BrowseComp-Plus, FRAMES and MoNaCo: 69.1 against 63.4.
- Highest observed support (answer credit recoverable from the saved evidence) on all three benchmarks: average 60.9 against 49.6.

Not to be confused with Panini (arXiv:2602.15156, continual learning in token space, ICML 2026) or GSW (Generative Semantic Workspaces, AAAI 2026), related work from the same lab whose code is at https://github.com/roychowdhuryresearch/gsw-memory. PANINI-DR builds its research workspace on the GSW representation.

## Links

- [Project page](https://roychowdhuryresearch.github.io/PANINI-DR_Project_Page/): interactive Figure 1, results, citation
- [Paper on arXiv](https://arxiv.org/abs/2610.15208)
- [Paper page on Hugging Face](https://huggingface.co/papers/2610.15208)
- [Code on GitHub](https://github.com/roychowdhuryresearch/panini-dr): Apache-2.0; paper configuration, graders, benchmark preparation, bring-your-own-dataset, and `panini-dr verify`

## Authors

M. Yigit Turali and Shreyas Rajesh (equal contribution), M. Efe Lorasdagi, Chenda Duan, Kartik Sharma, Vwani Roychowdhury (University of California, Los Angeles)

## Technical stack

- Language model, all roles: gpt-oss-120b served with vLLM 0.24.0 (OpenAI-compatible API; one 96 GB GPU at a 65,536-token context, or two for throughput)
- Reranker: jinaai/jina-reranker-v3.5 served with vLLM 0.26.0 (non-commercial licence)
- Dense retrieval: Qwen/Qwen3-Embedding-8B query encoder with FAISS; lexical retrieval: BM25 (Pyserini/Lucene for BrowseComp-Plus)
- Python 3.12, installed with uv; command-line tool `panini-dr` with `run`, `prepare`, `score` and `verify`

## Documentation

- [Install and serve the models](https://github.com/roychowdhuryresearch/panini-dr#install)
- [Run on your own data](https://github.com/roychowdhuryresearch/panini-dr#run-on-your-own-data)
- [Reproduce the paper](https://github.com/roychowdhuryresearch/panini-dr/blob/main/REPRODUCE.md)
- [Check an installation against the paper's code (`panini-dr verify`)](https://github.com/roychowdhuryresearch/panini-dr#check-this-code-against-the-papers-runs-panini-dr-verify)