Varun Reddy

Hello. I’m Varun, and this is the part of the internet I am personally responsible for. Most of it is research. The rest is not, and that is on purpose.

I think about learning as something that unfolds in time, in neural networks and in brains. I want to know how a representation gets built, how it changes as the system sees more of the world, and whether those two stories are secretly the same one.

I am at the Harvard Kempner Institute, working with Bernardo Sabatini. I studied Computer Science at the University of Virginia, where I worked with Yen-Ling Kuo on mechanistic interpretability and knowledge editing, and with Laura Barnes on AI for healthcare.

In undergrad I was also on the Fair Play R&D team at Chess.com, building models to catch overperformance in online games, and I spent time at Caltech in Dean Mobbs’ group, which is where I first sat down with the brain as a research object.

I help make AI and math videos at Welch Labs. We have a new book.

If you want to keep going, here is what is actually on this site:

The research, in longer form, including the parts I have not figured out.
Class notes I took for myself, starting with Susskind’s classical mechanics. Unofficial, and occasionally wrong.
Resources I keep going back to: lectures, papers, tools. Not a curriculum.
Photos and life outside work, the one section with no citations.
Philosophy, which is currently three sentences and a promise.

The non-research half is the reason I wanted a site in the first place. Notes, photos, and a philosophy page can sit two links away from a Hessian, and nobody can tell me those belong on separate websites.

My email is varunreddy [at] g [dot] harvard [dot] edu. I actually like email, and I write back.

news

May 15, 2025	I’ll be joining Chicago Trading Company this summer as a Quantitative Trading Intern.
May 10, 2025	I have graduated from the University of Virginia with a B.S. in Computer Science in 3 years.
Apr 05, 2025	Our team placed 3rd at the Point72 Cubist Hackathon, my friend wrote a bit about our project here.

selected papers

CaseEdit: Enhancing Localized Commonsense Reasoning via Null-Space Constrained Knowledge Editing in Small Parameter Language Models

2025

Abs PDF

Large language models (LLMs) exhibit strong performance on factual recall and general reasoning but struggle to adapt to user- specific, commonsense knowledge, a challenge particularly acute in small- parameter settings where computational efficiency is prioritized. We in- troduce CaseEdit , a new dataset and generation pipeline for evaluating localized, personalized commonsense knowledge editing in small LLMs to address this. Built upon the ATOMIC20/20 commonsense graph, CaseEdit uses a multi-stage inference process to generate both typical and atypi- cal contextual edits for household objects, paired with targeted evaluation questions across four axes: reliability, generalization, locality, and portabil- ity. We evaluate established knowledge editing methods using CaseEdit and demonstrate that AlphaEdit, a technique employing null-space projec- tion to minimize interference with unrelated knowledge, consistently out- performs other methods when applied to an LLaMA 3.2 3B model, even in scalability tests, showing minimal ripple effects. Our results indicate that using CaseEdit with effective editing techniques like AlphaEdit al- lows small models to internalize high-quality, context-sensitive common- sense knowledge, paving the way for lightweight, personalized assistants.
ACII

AudioInsight: Detecting Social Contexts Relevant to Social Anxiety from Speech

2024

Abs PDF

During social interactions, understanding the intricacies of the context can be vital, particularly for socially anxious individuals. While previous research has found that the presence of a social interaction can be detected from ambient audio, the nuances within social contexts, which influence how anxiety provoking interactions are, remain largely unexplored. As an alternative to traditional, burdensome methods like self-report, this study presents a novel approach that harnesses ambient audio segments to detect social threat contexts. We focus on two key dimensions: number of interaction partners (dyadic vs. group) and degree of evaluative threat (explicitly evaluative vs. not explicitly evaluative). Building on data from a Zoom-based social interaction study (N=52 college students, of whom the majority N=45 are socially anxious), we employ deep learning methods to achieve strong detection performance. Under sample-wide 5-fold Cross Validation (CV), our model distinguished dyadic from group interactions with 90% accuracy and detected evaluative threat at 83%. Using a leave-one-group-out CV, accuracies were 82% and 77%, respectively. While our data are based on virtual interactions due to pandemic constraints, our method has the potential to extend to diverse real-world settings. This research underscores the potential of passive sensing and AI to differentiate intricate social contexts, and may ultimately advance the ability of context-aware digital interventions to offer personalized mental health support.