Research papers.

Public research manuscripts and preprints from the lab. Open a paper for its details and reading links.

2026ARXIV PREPRINT

One Lens, Many Worlds: A Capability-Typed Interface for World-Model Interpretability

A shared interpretability interface for world models built on different architectures, with an end-to-end analysis of I-JEPA.

Bhavith Chandra Challagundla, Sanskar Pandey, Param Thakkar, Rishikesh Mallagundla, Yugandhar Reddy Gogireddy, Wenhao Lu, Hindol Roy Choudhury, Shravani Challagundla, Mohamed Deraz Nasr, Spursh Deshpande

Explore the paper

WorldModelLens defines a capability-typed adapter for recurrent, token-based, and joint-embedding world models. One hook and cache layer supports time-indexed activations, imagined rollouts, and intervention replay. The paper demonstrates the analysis suite on I-JEPA and describes adapters across several model families.

Paper record Read PDF

2026ICLR 2027 SUBMISSION

WorldModelLens: From Attention to Rollouts - A Framework for World Model Interpretability

A shared adapter and hook-and-cache toolkit for tracing world-model predictions from individual attention heads through full rollout behavior.

Explore the paper

WorldModelLens brings recurrent state-space, token-based, and joint-embedding world models behind one typed adapter interface. Its shared analysis layer supports probing, activation patching, sparse autoencoders, intervention replay, and imagined rollouts across model families.

In an I-JEPA ViT-H/14 study, Integrated Gradients showed a modest but statistically robust localization signal. The paper reports that latent attribution recovered the same ranking with about 180 times fewer model evaluations than exhaustive patching and correlated with head- and block-level interventions.

Paper record Read PDF

For ongoing work before public release, see working papers .