Soumya Suvra Ghosal

sghosal@umd.edu

I am a final-year Ph.D. student in the CS dept at the University of Maryland, College Park. I am fortunate to be advised by Prof. Dinesh Manocha. Prior to this, I completed my Master's in Computer Science at the University of Wisconsin-Madison, where I was lucky to be advised by Prof. Sharon Li. I am interested in the broad domain of AI Safety, Reasoning, and Agentic AI.

I am actively exploring Industrial Research Scientist or Academic positions in generative modeling, reasoning, agentic systems, and AI safety. If you have any relevant openings, please feel free to reach out.

Education

University of Maryland, College Park

PhD in Computer Sciences
August 2023 -

University of Wisconsin, Madison

MS in Computer Sciences
September 2021 - May 2023

National Institute of Technology Durgapur, India

Bachelor of Technology (B.Tech)
Electronics and Communications Engineering
August 2016 - May 2020

News

• 08/2023    Joined CS PhD program at University of Maryland College-Park.
• 08/2023    Joined GAMMA lab under supervision of Prof. Dinesh Manocha.
• 09/2023    Paper on "Are Vision Transformers Robust to Spurious Correlations?" accepted by International Journal of Computer Vision (IJCV).
• 10/2023    Our latest survey paper on "Towards Possibilities & Impossibilities of AI-generated Text Detection: A Survey" is on arXiv.
• 12/2023    Our latest work on Out-of-Distribution got accepted to AAAI 2024 . Thanks to my amazing collaborators Yiyou Sun and Sharon Li !
• 12/2023    Our survey paper "A Survey on the Possibilities & Impossibilities of AI-generated Text Detection" got accepted to TMLR .
• 05/2024    Joining Adobe as PhD Research Intern for Summer 2024.
• 06/2024    Our recent paper "Transfer Q Star: Principled Decoding for LLM Alignment" on inference-time alignment of large language models is now available on arXiv.
• 09/2024    Our latest work on Vision-Language prompt tuning got accepted to EMNLP 2024.
• 09/2024    Transfer Q-Star got accepted to NeurIPS 2024 .
• 01/2025    Our work on improving ICL performance in low resource settings, PromptRefine got accepted to NAACL 2025 .
• 02/2025    Our work on inference-time safety against jailbreak attacks in Multi-modal LLMs, Immune got accepted to CVPR 2025 .
• 05/2025    Our work on user-specified decoding on multiple reward objectives got accepted to ICML 2025 .
• 05/2025    Joining IBM Research as a Research Scientist Intern at Thomas J. Watson Research Center for Summer 2025.
• 09/2025    Our study on understanding test-time scaling in reasoning models got accepted to NeurIPS 2025 .
• 09/2025    Our work on reward modeling in low-resource Indic languages got accepted to EMNLP 2025 .
• 09/2025    Joining Amazon Science as a Applied Scientist Intern for Fall 2025.

Publications

Preprints

SafeThink: A Key to Safety in Multi-Modal Large Reasoning Models

Soumya Suvra Ghosal, Vaibhav Singh, Souradip Chakraborty, Mengdi Wang, Furong Huang, Dinesh Manocha, Amrit Singh Bedi [ Paper ]

Conference Proceedings

Does Thinking More always Help? Understanding Test-Time Scaling in Reasoning Models

Annual Conference on Neural Information Processing Systems (NeurIPS 2025).

Soumya Suvra Ghosal, Souradip Chakraborty, Avinash Reddy, Yifu Lu, Mengdi Wang, Dinesh Manocha, Furong Huang, Mohammad Ghavamzadeh, Amrit Singh Bedi [ Paper ]

Relic: Enhancing Reward Model Generalization for Low-Resource Indic Languages with Few-Shot Examples

Empirical Methods in Natural Language Processing (EMNLP) 2025 (Findings)

Soumya Suvra Ghosal, Vaibhav Singh, Akash Ghosh, Soumyabrata Pal, Subhadip Baidya, Sriparna Saha, Dinesh Manocha [ Paper ]

Bounded Rationality for LLMs: Satisficing Alignment at Inference-Time

International Conference on Machine Learning 2025 (ICML 2025)

Mohamad Chehade*, Soumya Suvra Ghosal*, Souradip Chakraborty, Avinash Reddy, Dinesh Manocha, Hao Zhu, Amrit Singh Bedi [ Paper ]

Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment

The IEEE/CVF Conference on Computer Vision and Pattern Recognition 2025 (CVPR 2025)

Soumya Suvra Ghosal*, Souradip Chakraborty*, Vaibhav Singh*, Tianrui Guan, Mengdi Wang, Alvaro Velasquez, Ahmad Beirami, Furong Huang, Dinesh Manocha, Amrit Singh Bedi [ Paper ] [ Project Page ]

PromptRefine: Enhancing Few-Shot Performance on Low-Resource Indic Languages with Example Selection from related Example Banks

Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics (NAACL 2025) (Main Track)

Soumya Suvra Ghosal, Soumyabrata Pal, Koyel Mukherjee, Dinesh Manocha [ Paper ] [ Code ]

Transfer Q Star: Principled Decoding for LLM Alignment

Neural Information Processing Systems (NeurIPS) 2024

Souradip Chakraborty*, Soumya Suvra Ghosal*, Ming Yin, Dinesh Manocha, Mengdi Wang, Amrit Singh Bedi, Furong Huang (* denotes Equal Contribution) [ Paper ] [ Code ]

IntCoOp: Interpretability-Aware Vision-Language Prompt Tuning

Empirical Methods in Natural Language Processing (EMNLP) 2024 (Main Track)

Soumya Suvra Ghosal, Samyadeep Basu, Soheil Feizi, Dinesh Manocha [ Paper ]

How to Overcome Curse-of-Dimensionality for Out-of-Distribution Detection?

AAAI Conference on Artificial Intelligence (AAAI) 2024

Soumya Suvra Ghosal*, Yiyou Sun*, Yixuan Li (* denotes Equal Contribution) [ Paper ] [ Code ]

Distributionally Robust Optimization with Probabilistic Group

AAAI Conference on Artificial Intelligence (AAAI) 2023 (Oral)

Soumya Suvra Ghosal, Yixuan Li [ Paper ] [ Code ]

A Novel Data Augmentation Technique for Out-of-Distribution Sample Detection using Compounded Corruptions

European Conference on Machine Learning and Principles \& Practice of Knowledge Discovery in Databases (ECML-PKDD) 2022. (Research Track)

Ramya S. Hebbalaguppe*, Soumya Suvra Ghosal, Jatin Prakash, Harshad Khadilkar, Chetan Arora. [ Paper ] [ Code ]

Journal Proceedings

A Survey on the Possibilities & Impossibilities of AI-generated Text Detection

Transactions on Machine Learning Research (TMLR)

Soumya Suvra Ghosal, Souradip Chakraborty, Jonas Geiping, Furong Huang, Dinesh Manocha, Amrit Bedi [ Paper ]

Are Vision Transformers Robust to Spurious Correlations?

International Journal of Computer Vision (IJCV)

Soumya Suvra Ghosal, Yixuan Li [ Paper ] [ Code ]

Awards and Fellowships

Reviewing Roles

JOY GURU !

I am a disciple of Sree Sree Thakur Anukul Chandra . Here are some inspiring quotes of Thakur Anukul Chandra:

  • " You are for the Lord
    Not for others,
    You are for the Lord
    and so for others. "
    -- Sree Sree Thakur Anukulchandra
    •  
  • " Serve as much as you are able,
    But take care you have no desire
    to be served. "
    -- Sree Sree Thakur Anukulchandra
    •  
  • " Do never die
    nor cause death
    but resist death
    to death. "
    -- Sree Sree Thakur Anukulchandra

Here is an English translated version of the Satyanusaran (Pursuit of Truth) book written by Thakur Anukul Chandra.