Ameya Jadhav

WebHierarch introduces a novel framework that combines large language model reasoning with hierarchical reinforcement learning to advance autonomous web automation. By integrating semantic understanding with experiential skill learning, the system achieves near state-of-the-art performance on complex MiniWoB++ benchmarks, with a 90% average success rate—outperforming both pure RL and LLM approaches. This work establishes a scalable and interpretable foundation for adaptable web agents capable of handling diverse and dynamic online environments.

This work presents a novel framework for reconstructing dynamic 3D meshes from monocular videos using Vision Transformers. The approach integrates D-NeRF scene representations, ViT-based spatial encoding, and Mesh R-CNN decoders, achieving measurable improvements in geometric accuracy and visual fidelity over prior methods. It establishes a proof of concept for applying transformer-based architectures to advance the state of the art in 3D reconstruction

This work introduces a methodology for augmenting mathematical reasoning in LLMs through the incorporation of axiomatic structures from Interactive Theorem Provers. Empirical evaluations reveal statistically significant, albeit modest, enhancement in proof-theoretic capabilities, suggesting this formalism-driven approach offers a promising vector for advancing machine cognition in mathematical domains.