Deep research requires connecting multiple steps of search, reasoning, and computation, maintaining research state across model calls, and checking findings before errors propagate. Panini-DR is 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, while a separate verification agent checks proposed answers against that evidence. Evaluated on BrowseComp-Plus, FRAMES, and MoNaCo against seven baselines with a shared language-model backbone and frozen corpora, Panini-DR achieves the strongest average performance: 73.3% accuracy on BrowseComp-Plus, 16.0 points above the next-best baseline at higher inference cost, 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.
@inproceedings{turali2026paninidr,title={{PANINI-DR}: Controlled Deep Research Through Generative Semantic Workspaces},author={Turali, M. Y. and Rajesh, S. and Lorasdagi, M. E. and Duan, C. and Sharma, K. and Roychowdhury, V.},booktitle={Submitted to the Fifteenth International Conference on Learning Representations (ICLR 2027)},year={2026},note={Under review.},}
LLM-based decision support for anti-seizure medication management in Ugandan pediatric epilepsy care must adapt to local prescribing practice and know when to defer. MANANA is a non-parametric prompt-learning framework, with single- and multi-agent variants, that converts observed prescription errors into auditable prompt memories learned from a small patient-level training set. Bayesian prompt averaging turns the learned prompt trajectory into prescription likelihoods and an uncertainty-based deferral signal: on an independently collected held-out cohort, it improves visit-level top-3 prescription accuracy by 4-8 percentage points over prompt-optimization baselines and auto-handles the most confident half of cases at 95% precision while deferring the rest for specialist review.
@inproceedings{rajesh2026manana,title={Teaching LLMs to Recommend and Defer in Underrepresented Epilepsy Care},author={Rajesh, S. and Sharma, K. and Monsoor, T. and Turali, M. Y. and Idro, R. and Kayaga, J. and Sebunya, R. and Namata, T. T. and Pasqua, J. N. and Roychowdhury, V. and Mazumder, R.},booktitle={Advances in Neural Information Processing Systems (NeurIPS)},year={2026},note={*Equal contribution by Rajesh and Sharma.},}
PANINI is a non-parametric continual-learning framework that encodes documents into Generative Semantic Workspaces (GSW)—entity- and event-aware networks of atomic question-answer pairs. A beam-search retrieval procedure (Reasoning Inference Chain Retrieval, RICR) follows multi-hop inference chains through GSWs to assemble compact evidence, achieving top average performance across six QA benchmarks while using 2-30x fewer answer-context tokens than competitive baselines.
@inproceedings{rajesh2026panini,title={Panini: Continual Learning in Token Space via Structured Memory},author={Rajesh, S. and Holur, P. S. and Turali, M. Y. and Duan, C. and Roychowdhury, V.},booktitle={Proceedings of the 43rd International Conference on Machine Learning (ICML)},year={2026},note={Accepted as poster. *Equal contribution by Rajesh and Holur.}}
Human brain assembloids offer a powerful platform for modeling neurological diseases, yet comprehensive methods for analyzing their complex network dynamics are lacking. We develop a time-resolved network analysis pipeline that extracts quantitative biomarkers from two-photon calcium imaging and apply it to assembloids carrying a pathogenic MAPT p.R406W variant, clinically associated with an Alzheimer’s disease-like phenotype, and their isogenic controls. Mutant networks exhibit significantly increased degree variance and clustering, a hub-like topology prone to hypersynchrony, and a Random Forest classifier trained on the dynamic network features distinguishes diseased from control states with an F1 score of 0.90.
@inproceedings{mccrimmon2026organoid,title={Time-Resolved Neuronal Network Dynamics Distinguish Pathological States in Organoid Models},author={McCrimmon, C. M. and Sinha, P. and Cao, Q. and Monsoor, T. and Sharma, K. and Turali, M. Y. and Samarasinghe, R. A. and Roychowdhury, V.},booktitle={2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},pages={7787--7791},year={2026},note={*Equal contribution by McCrimmon, Sinha, Cao, and Monsoor.},doi={10.1109/ICASSP55912.2026.11461496},}
We customize open-source LLMs to extract quantitative medication attributes (dose, frequency, duration) from heterogeneous electronic health record systems, demonstrating robustness across formats relevant to addiction-medicine deployments.
@inproceedings{fei2025ehrllm,title={Demo: Customizing Open-Source LLMs for Quantitative Medication Attribute Extraction across Heterogeneous EHR Systems},author={Fei, Z. and Turali, M. Y. and Rajesh, S. and Dai, X. and Pham, H. and Holur, P. S. and Zhu, Y. and Mooney, L. J. and Hser, Y.-I. and Roychowdhury, V.},booktitle={NeurIPS 2025 Workshop on GenAI for Health: Potential, Trust, and Policy Compliance},year={2025},note={Accepted as poster. *Equal contribution by Fei, Turali, and Rajesh.}}
@article{turali2025bfmo,title={Binary Feature Mask Optimization for Feature Selection},author={Lorasdagi, M. E. and Turali, M. Y. and Kozat, S. S.},journal={Neural Computing and Applications},volume={37},number={6},pages={5155--5167},year={2025},doi={10.1007/s00521-024-10913-9},}
@article{ozturk2025hydravit,title={HydraViT: Adaptive Multi-Branch Transformer for Multi-Label Disease Classification from Chest X-ray Images},author={Öztürk, Ş. and Turalı, M. Y. and Çukur, T.},journal={Biomedical Signal Processing and Control},volume={100},pages={106959},year={2025},doi={10.1016/j.bspc.2024.106959},}
@article{turali2024sgdfilter,title={Optimal Stochastic Gradient Descent Algorithm for Filtering},author={Turali, M. Y. and Koc, A. T. and Kozat, S. S.},journal={Digital Signal Processing},volume={155},pages={104731},year={2024},issn={1051-2004},doi={10.1016/j.dsp.2024.104731},}
@article{turali2024afsbm,title={AFS-BM: Enhancing Model Performance through Adaptive Feature Selection with Binary Masking},author={Turali, M. Y. and Lorasdagi, M. E. and Kozat, S. S.},journal={Signal, Image and Video Processing (SIViP)},volume={18},number={11},pages={7571--7582},year={2024},doi={10.1007/s11760-024-03411-x},}
@article{ilhan2023gbmma,title={Gradient Boosting With Moving-Average Terms for Nonlinear Sequential Regression},author={Ilhan, E. and Turali, M. Y. and Kozat, S. S.},journal={IEEE Signal Processing Letters},volume={30},pages={1182--1186},year={2023},doi={10.1109/LSP.2023.3309577},}
@article{turk2023jointnet,title={JointNET: A Deep Model for Predicting Active Sacroiliitis from Sacroiliac Joint Radiography},author={Turk, S. and Demirkaya, A. and Turali, M. Y.},journal={arXiv preprint},year={2023},doi={10.48550/ARXIV.2301.10769},}