Mohali, India · research in model-based RL
I study agents that learn a model of their world — and are curious about the parts of it that actually matter for their decisions.
↳ move around. the agent and you are both clearing the fog.
01 — currently exploring
Research InternGroupBrainAISep 2026 — present
I’m investigating how reinforcement-learning agents can use learned world models, uncertainty and information gain to choose informative experience — and become more sample-efficient because of it.
The question underneath: is general model uncertainty the right curiosity signal, or only the uncertainty that is relevant to the agent’s decisions? I’m designing controlled experiments to tell them apart.
02 — manuscripts
accepted Physica Scripta · co-author
Where does prediction stop being possible? We probed the limits of photonic quantum reservoir computing on financial time-series that behave almost like a random walk.
under revision co-author
We propose EviBench, a seven-axis benchmark that asks whether vision-language models are right for the right reasons — scoring evidence alignment, confidence calibration and counterfactual robustness across open and closed state-of-the-art models.
03 — project · 2026
A latent neural ODE learns a two-body orbit from only 20% of the trajectories.
Roll forward past the data and a plain MLP forgets the physics — its orbit flies apart.
A PINN holds on longer, but the error compounds and the orbit drifts.
The physics-informed world model keeps tracking at 25× the training window.
View on GitHub ↗schematic illustration — see the repo for real results
04 — project · 2026
Investigating reward design and the alignment problem: with PPO I studied reward misspecification and specification gaming — and demonstrated agents hacking their rewards in MountainCar and LunarLander.
View on GitHub ↗05 — episode log