About

I’m Sanket, an LLM Research Engineer at Multiverse Computing. These days most of my time goes into post-training LLMs with reinforcement learning — shaping training data mixes and reward environments, running multi-node GPU training runs, and building evaluation pipelines to figure out what actually improved (and what the model quietly learned to game). I also work close to real-world use cases, adapting efficient models to domain-specific tasks.

Earlier, I finished my PhD at PAVIS, Italian Institute of Technology under the supervision of Dr. Alessio Del Bue. During my research, I primarily focused on social interaction analysis for first person videos.

Previously, I have worked with Nokia Bell Labs , SmartVizx , Ernst and Young with their research teams. Besides, I had two startups during my college and filed two provisional patents. I have also been an active participant in technical communities and presented multiple technical talks and workshops at various level(s).

Areas of interest: RL for LLMs (RLVR, GRPO), reward modelling, preference tuning (RLHF, DPO), evaluation and efficient models.
PS: Also exploring mechanistic interpretability — and writing up what I learn along the way on Medium.

Checkout my latest podcast !

Hit me up and lets make another podcast.

News [Updates]:

*Mar’26 : Joined Multiverse Computing as LLM Research Engineer — working on post-training RL.

*Jan’25 : Joined Frontera Health as Applied Research Scientist.

Papers:

*Leveraging Next-Active Objects for Short-Term Anticipation in Egocentric Videos
Thakur S., Beyan C., Morerio P., Murino V., Del Bue A.
WACV ‘24
Project Page, Paper

*Guided Attention for Next Active Object@ EGO4D STA Challenge
CVPRW’23
Thakur S., Beyan C., Morerio P., Murino V., Del Bue A.
CVPRW Paper - GANOv2, Code, GANOv2 - CVPR23 submission
Winner CVPR23 EGO4D STA challenge ★

*Enchancing Next-Active-Object based Egocentric Action Anticipation with Guided Attention
ICIP’23
Thakur S., Beyan C., Morerio P., Murino V., Del Bue A.
Project Page, Paper

*Anticipating Next Active Object for egocentric videos
Thakur S., Beyan C., Morerio P., Murino V., Del Bue A.
Journal Review
Project Page, Paper

*Audio-Visual Inpainting: Reconstructing Missing Visual Information with Sound
ICASSP’23
Sanguineti V., Thakur S., Morerio P., Del Bue A., Murino V.
Project Page, Paper

*Predicting Gaze from Egocentric Social Interaction Videos and IMU Data
ACM ICMI’21
Thakur S., Beyan C., Morerio P., Del Bue A.
Project Page, Paper, Code

[Other]:

Stable Diffusion Concept : Pok-Logo < gan-pok> Hugging face

Generative AI based : Guess The YouTube game Project Page

Poster Award, Synapse AI symposium 2023: (Poster), Sept ‘23

Technical Talk at Dolby Labs : Next-Active-Objects for egocentric videos, Sept ‘23

Visiting CSIC-IRI(UPC), Barcelona to work with Mariella Dimiccoli, May ‘22

Invited Technical Talk at SVNIT,India : Egocentric Video Understanding , Aug ‘22

Attended VS3’22 Summer school , July ‘22

Accepted for ELLIS Doctoral symposium ‘21 , September ‘21

Website Chair, ETRA ‘22 , July ‘21

Started PhD at PAVIS/VGM, IIT , November ‘20

---

You can find me on Twitter.
If you’re oldfashioned you can send me an email

Checkout some of my VR / AR tutorials from previous life

PS: I am still updating this site, few commits a day.