Agentic Research at Harvard

• Intro
• The research
• The stay

Intro

From May 13th to Aug 17th 2026 I was research fellow at Harvard Engineering School hosted by Prof. Flavio du Pin Calmon. And I will be returning to Harvard in September for the Fall.

I would like to thank Dr. Flavio du Pin Calmon and Dr. Sajani Vithana for mentorship. I would also like to thank all my ASU collaborators, Dr. Lalitha Sankar, Dr. Oliver Kosut and Dr. Atefah Gilani for our discussions.

The research

I will have a detailed website for our work during this period soon and we are submitting to a conference (double blind), so I won’t go into details. Essentially I worked on the problem of “covert agentic communication”, where we wish to embed hidden messages into LLM generated text in “inference-time” without encoder and decoer sharing prompt or model weights. This is often referred to as “black-box” steganography. This is the same setting as the LLM watermarking problem recently got internet attention because of a post by antropic. The main difference is that we attempt to embed multiple bits instead of a one-bit watermark, for which the “rate” of embedding is an additional metric of concern.

Such a problem has been studied under multi-bit watermarking for past year but we have been able to improve the performance of embedding by essentially two orders of magnitude, compared to the latest baseline, via information-theoretic feedback coding. On C4 dataset with average 50 tokens, and 8-bit payload, we were able to reduce to the error from around 15% (most competitive baseline) to essentially error-free. We constructed a two-way communication protocol that identifies a new LLM security threat, that human actors may leverage independently deployed LLM agents to embed hidden messages with high rate and reliability.

The stay

I was residing on the Allston side of the campus, with a daily 4min-5min compute to work and overall I find the environment very enjoyable (pictures under media section of my website).