Soumyajit Karmakar

I am a CSE Master's student at UC San Diego, advised by Prof. Mikhail (Misha) Belkin and Prof. Sanjoy Dasgupta. My research focuses on studying the activation geometry of Large Language Models (LLMs). Currently, I am working on extending Recursive Feature Machines (RFM) to uncover universal linear transformations within the concept space of LLMs.

Prior to this, I worked at Fujitsu Research, Bengaluru. During this time I collaborated closely with Prof. R. Venkatesh Babu at the Vision and AI Lab (VAL), Indian Institute of Science. I worked on Multi-Modal Learning and Vision-Language Models. I completed B.Tech in CSE from IIIT Guwahati, where I secured departmental rank 2. I did my Bachelor's thesis under the co-supervision of Prof. C. V. Jawahar on the topic of Semantic segmentation for the urban road settings using Diffusion Models.

In the summer of 2022, I participated in the Parameterized Algorithms and Computational Experiments (PACE) competition under the guidance of Prof. Srinibas Swain. Our solver was awarded the best student submission award, in the heuristic track.

Email  /  CV  /  LinkedIn  /  Google Scholar  /  Github

profile photo
News

  • Dec 2025 - Attending NeurIPS 2025 in San Diego!
  • Sept 2024 - Started my master's degree at UCSD.
  • Oct 2023 - Paper on "A Study on ViTs Augmented by Masked Autoencoders" accepted at WACV 2024.
  • Jan 2023 - Joined CVIT, IIITH as a Research Fellow for my Bachelor's thesis under Prof. C. V. Jawahar.
  • Oct 2022 - Paper on "Convolutional Ensembling based Few-Shot Defect Detection Technique" accepted at ICVGIP 2022.
  • Aug 2022 - Started a project under Dr. Srijan Das at UNCC.
  • July 2022 - Got the Best student submission award at the PACE 2022 competition in the heuristics track.
  • May 2022 - Joined AITG, CSIR-CEERI as a Research Intern under Dr. Sanjay Singh.
  • Aug 2019 - Started BTech CSE at IIIT Guwahati.

Publications

2025
Prompt-ReID: Prompt Based Inference for Robust Re-Identification
Soumyajit Karmakar, Shrinivas Ramasubramanian, Harsh Rangwani, Yasuhiro Aoki, Genta Suzuki, and Venkatesh Babu Radhakrishnan.

Technical Report
paper / code

We developed a prompt-based inference mechanism using CLIP for robust person re-identification, eliminating the need for extensive fine-tuning.


2024
Limited Data, Unlimited Potential: A Study on ViTs Augmented by Masked Autoencoders
Srijan Das, Tanmay Jain, Dominick Reilly, Pranav Balaji, Soumyajit Karmakar, Shyam Marjit, Xiang Li, Abhijit Das, and Michael S. Ryoo.

Winter Conference on Applications of Computer Vision (WACV), 2024
paper / code

We developed a joint ViT training pipeline which uses a Self-Supervised Auxiliary Task alongside the primary task, for when the training data is limited.


2022
Convolutional Ensembling based Few-Shot Defect Detection Technique
Soumyajit Karmakar, Abeer Banerjee, Prashant Sadashiv Gidde, Sumeet Saurav, and Sanjay Singh.

Indian Conference on Computer Vision, Graphics and Image Processing (ICVGIP) 2022
arXiv / code

We developed a novel convolution stacking based ensemble learning technique that can efficiently extract and process the knowledge base of various pretrained CNN networks.

Feedback vertex set using Edge Density and REmove Redundant (FEDRER): A heuristic solver for finding a feedback vertex set in a directed graph
Aman Jain, Sachin Agarwal, Nimish Agrawal, Soumyajit Karmakar, Srinibas Swain.

Presented at the poster session of the IPEC 2022
paper / code

Our submission for the PACE 2022 competition. We developed a divide and conquer heuristic algorithm to obtain the Directed Feedback Vertex Set of a graph.

Academic Activities and other Projects

  • Standardized Exam: 331/340 in GRE (162 Verbal Reasoning, 169 in Quantitative Reasoning). 112/120 in TOEFL.
  • Contributed to open source project CompilerGym. CompilerGym is a open source library of reinforcement learning environments for compiler tasks maintained by Facebook Research.
  • Served as reviewer at ICVGIP 2022.


Thanks to Jon Barron for the template.