University of Cambridge · Trinity College
PhD Researcher in Computer Science
I am a PhD researcher at the University of Cambridge, supervised by Amanda Prorok and Pietro Liò, and a member of Trinity College.
I work on graph and hypergraph neural networks and their application to multi-robot systems. More specifically, I have been exploring advanced representations and neuro-algorithmic methods for multi-agent pathfinding, to obtain powerful and scalable neural solvers, while retaining the guarantees of classical algorithms. In general, I am interested in exploiting external knowledge and structure to build powerful neural models.
Learns coordinated motion primitives for continuous multi-robot planning through imitation and reinforcement learning. Backup trajectories safeguard collision-free execution, with experiments scaling to over a thousand robots.
A dual-path attention mechanism that combines structural priors with learned attention to improve graph and hypergraph models, often with fewer parameters and faster inference.
PT-RAG learns which related perturbations are useful for a given cell state, bringing end-to-end differentiable retrieval to models of cellular responses.
Hypergraph attention captures group interactions in multi-agent pathfinding. A compact 1M-parameter model outperforms an 85M-parameter GPT-based baseline using 100× less training data.
LaGAT combines learned graph-attention guidance with classical search for scalable, high-quality pathfinding, preserving completeness and eventual optimality guarantees.
Differentiable priority-queue memory for graph neural networks, improving algorithmic reasoning and the modelling of long-range interactions. Developed during my master’s thesis.
Best Paper Award, ICML KLR Workshop · Spotlight, KDD DLG Workshop
An explainable image classifier that discovers visual concepts automatically, removing the need for manual concept annotations and enabling corrections at the concept level.
Text classifiers that explain predictions through human-readable logic rules. The explanations improve on LIME in faithfulness and sensitivity, and were preferred in a user study. Developed during my undergraduate thesis.
Get in touch
rishabh.jain [at] cst.cam.ac.uk
Department of Computer Science and Technology
University of Cambridge, United Kingdom