Ashwinee Panda

I am a Postdoctoral Fellow at UMD working with Tom Goldstein on LLMs.

I received my PhD from Princeton University working with Prateek Mittal. During my PhD, I received the OpenAI Superalignment Fast Grant for our work showing that current safety alignment is shallow, which received the Outstanding Paper Award at ICLR 2025.

If you are interested in working with me, send me a DM on Twitter.

CV  /  Google Scholar  /  Twitter  /  Github

Research

I am currently working on a number of topics in LLMs. Right now I'm trying to understand how reasoning improves LLMs. Another big direction is in identifying what elements of the LLM pipeline are bottlenecking our scaling growth. A lot of my research is conducted through the lens of sparsity. I was a lead organizer for the ICLR 2025 workshop on Sparsity in LLMs.

We analyze safety alignment and show that it is largely shallow. We propose methods for making it deeper.

We present a lightweight approximation method that gives the MoE router a dense gradient update while continuing to sparsely activate its parameters.

We conduct an in-depth analysis of attention in video diffusion transformers (VDiTs) and report a number of novel findings.

We present the first method for doing privacy auditing of LLMs.

We release the Gemstone model suite for open-source scaling laws.

We propose a new definition for attention heads dominated by attention sinks, known as dormant attention heads.

Lottery Ticket Adaptation (LoTA) is a new adaptation method that handles challenging tasks, mitigates catastrophic forgetting, and enables model merging across different tasks.

We propose the first method for performing differentially private fine-tuning of large language models without backpropagation. Our method is the first to provide a nontrivial privacy-utility tradeoff under pure differential privacy.

We find that using scaling laws for Differentially Private Hyperparameter Optimization significantly outperforms prior work in privacy and compute cost.

We propose a new practical data extraction attack that we call "neural phishing". This attack enables an adversary to target and extract sensitive or personally identifiable information (PII), e.g., credit card numbers, from a model trained on user data.

We propose the first method for performing differentially private in-context learning. Our method generates sentences from in-context learning while keeping the in-context exemplars differentially private, that can be applied to blackbox APIs (ex RAG).

We propose the first method for generating visual adversarial examples that can serve as transferrable universal jailbreaks against aligned large language models.

We pretrain networks with synthetic images that have strong performance on downstream private computer vision tasks.

We generate differentially private images from non-privately trained diffusion models by analyzing the inherent privacy of stochastic sampling.

Neurotoxin is a novel model poisoning attack for federated learning that stays present in the system for up to 5X longer than the baseline attack.

SparseFed is a provably robust defense against model poisoning attacks in federated learning that uses server-side sparsification to avoid updating malicious neurons.

FetchSGD is a communication-efficient federated learning algorithm that compresses gradient updates with sketches.

Not Research

WeChat  /  LinkedIn  /  Instagram  /  Yelp  /  Goodreads /  Spotify

I've lived all over America; feel free to ask me for food recs in NYC, LA, the SF Bay Area, or the DMV.

While at Berkeley I founded DiscreetAI (now inactive), a venture-backed startup building privacy-preserving machine learning as-a-service. You can check out our ProductHunt launch or our GitHub for more information. Among other things we won the first YCombinator Hackathon and built federated learning solutions for Fortune 500 companies.

I read a lot and post my thoughts on Twitter and Goodreads.

Website template from Jon Barron.