I've been blessed with the opportunity to mentor many passionate undergraduate, predoctoral, and masters students during my time at UCSB. Here's a little rundown:
Since 2021 I have mentored at least one group of young researchers a year.
I mentored two Early Research Scholars Program groups, in a direct supervisory role from 2021-2022, and as a sort of "secondary" supervisor, helping guide the junior grad student mentors of the following group from 2022 onward. Additionally, I led a Fatima Al-Fihri Predoctoral Fellowship group from 2023-2024.
Each of these groups has published at least one top conference paper with the mentees as lead authors, an ICLR 2023 Oral, and EMNLP 2023 poster, and a NeurIPS 2024 spotlight poster respectively.
With mentees Matthew Ho, Aditya Sharma, and Justin Chang we released WikiWhy, a challenge QA dataset containing "why" questions which requires that LMs successfully produce rationales alongside each answer. We devised a creative pipeline to produce good why questions with nontrivial rationales by crawling Wikipedia for sentences indicating complex causal relations. I coached the students in designing and scoping a project, data collection, crowdworker annotation, and baseline model training and assessment. The students put forth exceptional effort and polish in codebase and resources. Our paper appeared as an oral at ICLR 2023.
Primary mentors Yujie Lu and Alex Mei led this group of students Vaishnavi Himakunthala, Andy Ouyang, Danny Rose, and Ryan He. I was a sort of meta-mentor who participated heavily in the early stages of scoping the project. The group devised a "visual chain of thought" method for VLM reasoning in videos, where the "chain of thought" entails infilling the sequence of frames, and the VIP dataset to further evaluation of vision-language models, which they presented at EMNLP 2023.
All four are currently in their senior year at UCSB and plan to graduate in 2025.
I worked with Fatima Jahara and Mahsa Khoshnoodi, members of a Fatima Al-Fihri Fellowship group remotely from Bangladesh and Iran respectively. Our project rescoped a few times based on surprising findings they made over the course of the investigation, until we landed on meta-evaluation of text-to-image faithfulness metrics. We produced the T2IScoreScore meta-evaluation of faithfulness benchmarks and found that current evaluation practices are inadequate. This work was a spotlight paper at NeurIPS 2024.
In addition to mentorship groups, I have worked with many students on 1-1 projects, including:
Avani Tanna (UCSB MS 2023): as a part of her MS thesis project we trained a more robust voice conversion system for accent robustness which she presented at Interspeech 2023.
Aditya Sharma (UCSB MS student 2023-2024): after the ERSP program, I continued working with Aditya, and as a part of his MS thesis project we produced "LoCoVQA", a work where we demonstrated failure cases of VLMs in long-context reasoning under visual distractors which we presented at EMNLP 2024.
Siddharth Yayavaram and Arnav Yayavaram (BITS Pilani undergrads, 2024-2025): with primary mentor Simran Khanuja we produced the CAIRe tool, a method for performing cultural attribution of images which is currently under review.