research

Research projects at Mila and selected past work. † denotes ongoing projects.

  • 2026
    ONGOING

    Discrete Diffusion for Molecular Generation with RL and Natural Language Steering

    Kains Kavuluri

    Université de Montréal, April 2026

    Learns drug-like molecule representations through discrete masked diffusion over SELFIES, with 100% chemical validity by construction. REINFORCE with per-step gradient accumulation steers generation toward desired properties, improving drug-likeness by 12.8%. Natural language conditioning via SciBERT cross-attention reaches 76% of SOTA at 15% of the parameter count.

  • 2025
    ONGOING

    Reinforcement Learning in Non-Stationary Environments: Loss of plasticity and evolutionary remedies

    Kains Kavuluri

    Ongoing research · Mila, 2025–present

    Investigating the loss of plasticity phenomenon in deep RL agents and experimenting with evolutionary algorithms to overcome it.

  • 2021
    THESIS

    Progressive Conformer for Image Classification

    Kains Kavuluri

    NIT Karnataka, 2021

    A hybrid architecture combining Transformers and CNNs at the attention level for progressive feature learning in image classification.