नमस्ते · bonjour · hi

Soumya Sharma

PhD student · EQUAL Lab, McGill & Mila · Montréal

I study privacy leakages in multi-agent LLM systems and how to prevent them.

I’m a PhD student at McGill and Mila in the EQUAL Lab, advised by Golnoosh Farnadi and Adriana Romero Soriano. I research safety, privacy and alignment in multi-agent LLM systems.

Before this I spent two years as a predoctoral researcher on the AQUA team at Google DeepMind India with Dr. Aravindan Raghuveer, working on reward modeling, evaluation and self-improving agents. Earlier: an M.S. (Research) at IIT Kharagpur with Prof. Pawan Goyal and Prof. Niloy Ganguly.

email cv scholar github linkedin

Lately

2026Our privacy-leakage benchmark for multi-agent systems is an oral at the COLM 2026 HAIPS workshop. I also presented it at WiML @ ICML 2026, South Korea..

2026IDEAFix, on creativity and hive-mind behaviour in LLMs, was presented at the Human CoCreativity workshop at ICML 2026.

Fall 2026TA-ing Computational Linguistics at McGill.

2026IDP-Bench is out on arXiv!

June 2026Attended CIFAR AI Safety School in Toronto.

May 2026Attended BlueDot AI Safety School.

Winter 2026TA for Natural Language Processing at McGill.

Aug 2025Started my PhD at McGill & Mila in the EQUAL Lab.

Research

all papers & tl;drs →

NOW

Privacy in multi-agent LLMs

Benchmarks and adversarial probing for contextual-integrity violations in agentic dialogue.

BEFORE

Financial & clinical NLP

Datasets and benchmarks for financial documents with Goldman Sachs and Samsung Research, and knowledge-graph-grounded medical NLI, at ACL, EMNLP and WWW.