Zarif Ikram (জারিফ ইকরাম)

I am a first year Annenberg PhD fellow and research assistant at USC Viterbi, advised by Paria Rashidinejad in close collaboration with Stephen Tu & Mahdi Soltanolkotabi.

I am an incoming research scientist intern at Microsoft Research AI Frontiers, NYC under John Langford and Riashat Islam.

Previously, I studied planning and learning with world models at the Artificial Scientific Intelligence Lab within the NUS AI Institute under Dianbo Liu and Ling Pan. In Summer'25,I led the AI consultancy efforts at Superobservation. I hold a degree from CSE, BUET (Class of '25); my undergraduate thesis on discrete generative models was advised by M Saifur Rahman. Prior to that, I studied pedestrian behavior modeling under Jim Whitehead at the Augmented Design Lab, UC Santa Cruz.

I perform as a tenor with the USC Thornton Concert Choir under the direction of Tram Sparks. My musical foundation is rooted in Indian classical music, having trained in the Indore gharana under Asit Dey. Additionally, I earned a first-class specialization in Nazrul Sangeet from Chhayanaut Music School. As a multi-instrumental vocalist, I accompany myself on harmonium, guitar, and piano.

Contact: zikram[at]usc.edu
GitHub  /  Google Scholar  /  X  /  YouTube  /  Anonymous Feedback   

Research

I am interested in fundamental AI problems in world modeling and continual improvement.

CrispEdit is a scalable second-order LLM editing method that enforces capability preservation via curvature-aware constrained optimization, achieving high edit success with <1% capability degradation across benchmarks.

A powerful method for the GFlowNets training using Evolutionary algorithms that enhances its robustness to long trajectory and sparse rewards

We find an interesting performance assymtry in Atari100k benchmark and propose a novel latent denoising world model solution for it.

Incorporating MaskGIT as a dynamics prior to enhance sequence modeling in world models, we address the limitation of transformer-based world models in continuous action environments and the inadequacies of prior methods, like STORM, in capturing effective state representations, suggesting that Masked Generative Priors could be a powerful inductive bias for world models, supporting broader applicability across diverse RL tasks and environments.

Taking inpiration from animal brain’s ability to perform temporal prediction in the latent space, we propose a simple unsupervised augmentation to existing methods for better downstream sample efficiency of world models.

Combining denoising score model and discriminator model for discretized Langevin MCMC, this method generates high quality biomolecular sequences that optimize a given objective.

GFlowNets enable diverse generation of HD maps that are structurally valid.

We design and propose a grid-based simulator that is lightweight and can be used to model pedestrian behavior in a variety of scenarios.

Procedural content generation method to generate infinite complex roundabout HD maps that can be used in autonomous vehicle testing simulators.

Selected Talks/Lectures

I love to share my experiences with the audience and learn from them.

Academic Services

Music

My music interest is shaped by music lessons from the age of five till today and mostly comprises of classical and comtemporary. I listen to everything.

Selected honors

Selected Television and Stage Appearances

this guy's website is kinda cool.