Research 🔍

Agentic search systems iteratively interact with retrieval models to answer complex queries. Despite substantial progress, optimizing retrievers for agentic search remains challenging, often requiring heavy co-training or gold-standard annotations that limit real-world applicability. We propose Critic-R, a framework that explicitly closes the feedback loop between the reasoning agent and the retrieval model during both inference and training. Critic-R introduces a critic model that evaluates the agent’s introspective reasoning trace after consuming retrieved evidence to determine whether the retrieved context sufficiently supports the next reasoning step. Critic-R has two complementary mechanisms: Critic-R-Zero, an inference-time query refinement loop that iteratively rewrites queries and retrieval instructions, and Critic-Embed, an optimization approach for retrieval models that leverages successful and failed refinement trajectories as automatic supervision without requiring manual relevance annotation.

Investigated knowledge-intensive video question answering (KI-VideoQA) through the lens of multi-modal retrieval-augmented generation. We particularly focused on handling open-ended questions rather than just multiple-choice formats. We highlight the critical role of query construction and retrieval depth optimization in effective knowledge integration. Through our proposed approach, we achieve a substantial 17.5% improvement in accuracy on multiple choice questions in the KnowIT VQA dataset, establishing new state-of-the-art performance levels.

Worked on creating a geospatial language model that enhances the understanding of geo-entities in natural language (e.g. using landmarks and directions). This in turn helps create an address matching platform specially tuned for house addresses, where platforms like Google Maps fail (e.g. in places like Dhaka). This helps optimize delivery agents to deliver orders faster and more efficiently.

Built an annotation-efficient semi-supervised particle-picking framework for 3D object detection from macromolecular samples. The goal was to detect and classify particles from cryo-electron tomogramss having signal-to-noise ratio as low as 0.1, making it a very challenging problem

Developed PULSAR, an automated PD screening tool. It uses spatio-temporal graph neural network to detect PD from videos. We were also the first to explore positive unlabeled learning in this setting, addressing the lack of reliable labeled data, which is generally the case in medical data.

Our proposed method QFM-FI, a faster and improved version of the QFM algorithm, can amalgamate millions of quartets over thousands of taxa into a species tree with a great level of accuracy. This also achieves a speedup of 20,000× compared to its predecessor. Worked on providing a theoretical analysis of the running time and memory requirements of QFM-FI.