Supervisor: Sashank Varma

Project:- Cognition as a Framework for Understanding AI Capability Profiles

1. Evaluating whether LLMs show intuitive (naive) but incorrect reasoning patterns that resemble common misconceptions held by people across domains like physics, teleological explanations, neuromyths, etc.

2.  We curate datasets of myths and misconceptions, then evaluate whether models can correctly identify and refute these errors or whether they themselves fall prey to them.

Supervisor: Dongwon Lee

Project:- A Comprehensive Taxonomy for Text watermarking (Accepted at NAACL '25)

1. We analyze research based on the specific intentions behind different watermarking techniques, evaluation datasets used, and watermarking addition and removal methods to construct a cohesive taxonomy. Building on this, we highlight the gaps and open challenges in text watermarking to promote research protecting text authorship.

2. Paper: Link

Supervisor: Hanspeter Pfister

Project:- Spot the lie: VLMs and the detection of misleading data visualizations.

1. Understanding whether VLMs are susceptible to different misleading questions/ contexts for charts and systematically investigating model sensitivity to different reasoning errors in visualization contexts, such as cherry-picking, causal inference, etc.

2. On track to be submitted to ARR 2026 cycle (January).

Supervisor: Xiang Ren

Project:- Generating free text explanations with LLMs

1. Building an attention pipeline to generate free text explanations for authorship verification.

2. Training the model to find similarities between the  two documents by classifying the documents into 5 feature vectors defining the 5 adjective styles of a document

3. Contributing towards building a training  dataset consisting of various adjective pairs for each feature vectors

Supervisor: Ali Emami

Project:- How Geographic Origin Influences LLMs Entity Deduction Capabilities (Presented at COLM 2025)

1. Awarded MITACS GRI’24 Scholarship at Brock University

2. Evaluating implicit geographical bias in LLMs through a multi-turn entity deduction game. We use notable people and culturally significant objects from diverse regions as entities and investigate the effect of the frequency of entities in pre-training datasets and their popularity on deduction capabilities.

3. Paper: Link

Supervisor: Sujith Thomas 

Project:- Effect of Active Recall and Personality Traits on Learning

1. Responsible for conducting experiments to capture participants' personality traits, performance on inference, and category learning with and without active recall.

2. With R language, statistical methods like multiple linear regression and individual sample t-tests are used.

Developed preprocessing utilities and evaluation scripts for JioBrain's Speech-to-Text module, automating dataset ingestion and inference-quality analysis across multilingual pilot datasets for Hindi, Gujarati, Marwari, Marathi, and Tamil.

A core member of the development club of BITS-Goa, whose main aim is to spread knowledge and learning related to Mobile, Web, App, and Game Development to students.

The group also organises hackathons as well as tech fests frequently within the College

Also a part of Open Source projects created by the group. 

Finished 3rd in Hackathon Organised by DevSoc for all the first-year undergraduate students at BITS Goa where I and my team made an app in which a person can book a cab from wherever he/she wants

Mentored newly inducted first-year students to get accustomed to the campus environment.

Guided them with their academics in their first year 

Aims to educate the underprivileged section of the society

I taught the workers in the mess subjects such as English and Mathematics, as well as applications like Excel, PowerPoint, and email writing.

Heading and managing the overall education of all the mess workers in BITS-Goa