Ehsan Aghazadeh

Hi! I’m a PhD student at UMass Amherst in Computer Science, advised by Hossein Pishro-Nik and Hedyeh Beyhaghi.

My research spans both practical and theoretical aspects of Large Language Models (LLMs) and AI systems. I’m currently working on:

• Reliability of multi-agent AI systems — modeling agent networks as factor graphs with LDPC-inspired density evolution to predict failure modes and optimize architectures.
• Test-time scaling & self-consistency — CGES, a Bayesian framework that achieves 25–40% reduction in LLM inference cost with no accuracy loss.
• Ergodic trajectory design — provably coverage-optimal trajectory planning via conditional flow matching, with zero-shot deployment across multi-agent fleets.

Previously, I worked on mechanistic interpretability of language models (probing, token attribution, hallucination analysis in VLMs) and efficient training (gradient-based data pruning for NLP).

I did my undergraduate at the University of Tehran and have been a reviewer at NeurIPS, ACL, NAACL, and EACL.