Preprint2026
Adaptive Block Diffusion: Resolving Training-Inference Mismatch in Diffusion Language Models
Training over varied prefix-window configurations to support flexible decoding in diffusion language models.
01 / Gagan Jain
Research
My research spans efficient generative models, adaptive computation, and learning from feedback.
Generative modelingEfficient inferenceReinforcement learningPost-training
Publications & preprints
Google Scholar ↗ICML2025
Masked Generative Nested Transformers with Decode Time Scaling
Decode-time model scaling and computation reuse for efficient image and video generation.
AAMAS2025
Bayesian Collaborative Bandits with Thompson Sampling for Improved Outreach in Maternal Health Program
A Bayesian approach to collaborative bandits for more effective maternal health outreach.
NeurIPS2024
Mixture of Nested Experts: Adaptive Processing of Visual Tokens
Routing visual tokens through nested experts to allocate computation where it matters.
ECCV2024
LookupViT: Compressing visual information to a limited number of tokens
Compressing visual information into fewer tokens while retaining a high-resolution lookup representation.