Logit-Lens Layer Probing for InternVL3
Jul 15, 2026A random mini experiment applying logit-lens probing to InternVL3 on our aerial imagery dataset
Hi, I'm a researcher, data scientist.
I love working on problems with social impact. Research Interests: Generative and Responsible AI, Low-resource NLP, LLMs, Computational linguistics, education-tech and pedagogy. I did my BS in Data science from IIT Madras. I'm currently working on hallucination in Visual Language Models (VLMs) at IIT Hyderabad. Prev. research internship experience: IIT Roorkee, IISER Kolkata.
News
What I'm up to?
September 2026
Our paper on Hallucination detection in aerial imagery was accepted at ACCV 2026!
January 2026
Our IITR AI4Fashion work has been published in multiple newsletters including The Times of India, PIB!!
January 2026
Will be joining the NLIP lab at IIT Hyderabad during the final terms of my BS!!
Publications
Conference proceedings — see Scholar for shared task system papers
Dressing the Imagination: A Dataset for AI-Powered Translation of Text into Fashion Outfits and A Novel NeRA Adapter for Enhanced Feature Adaptation
Gayatri Deshmukh, Somsubhra De, Chirag Sehgal, Jishu Sen Gupta, Sparsh Mittal
WACV 2026, 2026
Abstract:
TL;DR: We present FLORA (Fashion Language Outfit Representation for Apparel Generation), the 1st curated dataset of fashion outfit sketches paired with rich, industry-grade textual descriptions. FLORA aims to advance AI-driven fashion design and assist designers and end-users in bringing creative ideas to life. We also propose NeRA (Nonlinear low-rank Expressive Representation Adapter), a novel parameter-efficient adapter based on Kolmogorov-Arnold Networks (KANs).
@InProceedings{Deshmukh_2026_WACV,
author = {Deshmukh, Gayatri and De, Somsubhra and Sehgal, Chirag and Gupta, Jishu Sen and Mittal, Sparsh},
title = {Dressing the Imagination: A Dataset for AI-Powered Translation of Text into Fashion Outfits and A Novel NeRA Adapter for Enhanced Feature Adaptation},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
month = {March},
year = {2026},
pages = {2094-2103}
}
LLMs – the Good, the Bad or the Indispensable?: A Use Case on Legal Statute Prediction and Legal Judgment Prediction on Indian Court Cases
Shaurya Vats, Atharva Zope, Somsubhra De, Anurag Sharma, Upal Bhattacharya, Shubham Kumar Nigam, Shouvik Guha, Koustav Rudra, Kripabandhu Ghosh
EMNLP 2023 (Findings), 2023
Abstract:
The Large Language Models (LLMs) have impacted many real-life tasks. To examine the efficacy of LLMs in a high-stake domain like law, we have applied state-of-the-art LLMs for two popular tasks: Statute Prediction and Judgment Prediction, on Indian Supreme Court cases. We see that while LLMs exhibit excellent predictive performance in Statute Prediction, their performance dips in Judgment Prediction when compared with many standard models. The explanations generated by LLMs (along with prediction) are of moderate to decent quality. We also see evidence of gender and religious bias in the LLM-predicted results. In addition, we present a note from a senior legal expert on the ethical concerns of deploying LLMs in these critical legal tasks.
@inproceedings{vats-etal-2023-llms,
title = "{LLM}s {--} the Good, the Bad or the Indispensable?: A Use Case on Legal Statute Prediction and Legal Judgment Prediction on {I}ndian Court Cases",
author = "Vats, Shaurya and
Zope, Atharva and
De, Somsubhra and
Sharma, Anurag and
Bhattacharya, Upal and
Nigam, Shubham Kumar and
Guha, Shouvik and
Rudra, Koustav and
Ghosh, Kripabandhu",
editor = "Bouamor, Houda and
Pino, Juan and
Bali, Kalika",
booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2023",
month = dec,
year = "2023",
address = "Singapore",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.findings-emnlp.831/",
doi = "10.18653/v1/2023.findings-emnlp.831",
pages = "12451--12474",
abstract = "The Large Language Models (LLMs) have impacted many real-life tasks. To examine the efficacy of LLMs in a high-stake domain like law, we have applied state-of-the-art LLMs for two popular tasks: Statute Prediction and Judgment Prediction, on Indian Supreme Court cases. We see that while LLMs exhibit excellent predictive performance in Statute Prediction, their performance dips in Judgment Prediction when compared with many standard models. The explanations generated by LLMs (along with prediction) are of moderate to decent quality. We also see evidence of gender and religious bias in the LLM-predicted results. In addition, we present a note from a senior legal expert on the ethical concerns of deploying LLMs in these critical legal tasks."
}