PeopleofSuperintelligence

Researcher

Ashwinee Panda

Ashwinee Panda is an AI researcher whose work spans federated learning, private machine learning, and language-model safety. He coauthored research on compressed federated updates, durable backdoors, private image classification, and extracting sensitive information from language models trained on private data. In October 2023, he was a fourth-year Princeton PhD student working with Prateek Mittal; he had earned B.S. and M.S. degrees at UC Berkeley's RISE Lab, where he studied federated learning. Safety Alignment won an ICLR 2025 Outstanding Paper Award, and he coauthored DynaGuard, a system that evaluates text against user-defined policies and supports chatbot self-correction. As of October 2026, he builds reinforcement-learning systems and algorithms at Together AI.

Badges

Timeline

  1. 2020

    FetchSGD reduces communication

    In 2020, he coauthors FetchSGD, which compresses model updates with Count Sketch and merges sketches across workers to reduce federated learning's communication bottleneck.

  2. 2022

    SparseFed and Neurotoxin

    In 2022, he coauthors SparseFed, which uses global top-k update sparsification and device-level gradient clipping to mitigate model-poisoning attacks. He also coauthors Neurotoxin, a backdoor attack on parameters changed less during training that makes backdoors more durable.

  3. 2023

    DP-RandP for private images

    In 2023, he coauthors DP-RandP, which learns priors from images generated by random processes and transfers them to private data to improve the privacy-utility tradeoff of DP-SGD.

  4. October 2023

    Berkeley degrees and Princeton PhD

    In October 2023, he is a fourth-year PhD student at Princeton working with Prateek Mittal. He holds B.S. and M.S. degrees from UC Berkeley, where he researches federated learning in the RISE Lab with co-advisers Joey Gonzalez and Raluca Ada Popa.

  5. 2024

    Neural phishing for LLMs

    In 2024, he coauthors Teach LLMs to Phish, a practical data-extraction attack that targets language models trained on private data and extracts sensitive or personally identifiable information.

  6. 2025

    Safety Alignment wins an ICLR award

    During his Princeton PhD with adviser Prateek Mittal, he coauthors Safety Alignment, which argues that current safety alignment is shallow. The paper receives an Outstanding Paper Award at ICLR 2025.

  7. 2025

    DynaGuard at UMIACS

    In 2025, he is a postdoctoral fellow at the University of Maryland, working with Tom Goldstein on large language models through UMIACS and the Institute for Trustworthy AI in Law & Society. He coauthors DynaGuard, which evaluates text against user-defined policies and gives chatbots explanations they can use to self-correct.

  8. As of October 2026

    Reinforcement learning at Together AI

    As of October 2026, he builds systems and algorithms for reinforcement learning at Together AI.

Career

  1. NowTogether AI
  2. From 2025University of MarylandPostdoctoral Fellow
  3. University of California, BerkeleyResearcher

Education

  1. Princeton UniversityPhD
  2. University of California, BerkeleyB.S. and M.S.

From the biography.

  1. Among employers with 30+ full-time cases, OpenAI leads primary technical role H-1B pay at $310,000, and Amazon, Google and Microsoft file 57% of all casesIncludes Together AI. Across 440,669 certified US H-1B filings from 44 employers (FY2020–FY2026), OpenAI leads primary technical role base pay at $310,000 and xAI leads overall base pay at $290,000.
  2. The company share of AI researchers’ top papers was 31% in 2020–2024 and 8% in 1970–1974Includes Together AI. We estimate where the people shaping AI worked when their most-cited papers appeared: the company share was 31% in 2020–2024 and 8% in 1970–1974, and Google led all organizations in 2020–2024.
  3. 11% of AI doctorate holders from 2010 to 2016 founded or led a company or lab by 10 years after graduationIncludes Together AI. We estimate how often AI doctorate holders founded or led a company or lab by 10 years after graduation, how waiting times compare between founders and top jobs, and which universities trained the most.
  4. More people shaping AI studied at Massachusetts Institute of Technology than at any other institutionIncludes University of California, Berkeley. We estimate where the people shaping AI studied and where their careers led. Massachusetts Institute of Technology leads with 467 former students, including 291 recorded doctorates.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Ashwinee Panda yet. Send us a free-licensed photo.

  1. FetchSGD: Communication-Efficient Federated Learning with Sketching proceedings.mlr.press. Checked 21 October 2026.
  2. SparseFed: Mitigating Model Poisoning Attacks in Federated Learning with Sparsification proceedings.mlr.press. Checked 21 October 2026.
  3. Neurotoxin: Durable Backdoors in Federated Learning proceedings.mlr.press. Checked 21 October 2026.
  4. Differentially Private Image Classification by Learning Priors from Random Processes arxiv.org. Checked 21 October 2026.
  5. Challenges in Augmenting Large Language Models with Private Data talks.cs.umd.edu. Checked 21 October 2026.
  6. Teach LLMs to Phish: Stealing Private Information from Language Models proceedings.iclr.cc. Checked 21 October 2026.
  7. Ashwinee Panda kiddyboots216.github.io. Checked 21 October 2026.
  8. Announcing the Outstanding Paper Awards at ICLR 2025 blog.iclr.cc. Checked 21 October 2026.
  9. Day 1 Program — International Symposium on Trustworthy Foundation Models istfm.github.io. Checked 21 October 2026.
  10. Panda — NSF Institute for Trustworthy AI in Law & Society (TRAILS) trails.umd.edu. Checked 21 October 2026.
  11. DynaGuard: A Dynamic Guardian Model With User-Defined Policies arxiv.org. Checked 21 October 2026.
  12. DynaGuard: A Dynamic Guardrail Model With User-Defined Policies taruschirag.github.io. Checked 21 October 2026.