Profile banner
Ahmad Beirami

Ahmad Beirami

AI Research Leader at Google DeepMind, Meta, and Electronic Arts

Joined June 2025
Favicon imagebeirami.github.io
Favicon imagescholar.google.com
Favicon imagelinkedin.com
Favicon imagex.com
Favicon imagegithub.com

Network

17.5K connections
💻
Google Software Engineers
🔬
Google Research Scientists
🧠
Google DeepMind AI/ML
🌟
Google AI/ML Leaders
♊
Google DeepMind Gemini Team
👑
Google DeepMind AI Leaders
🤖
AI/ML Rsch/Eng

Summary

Ahmad Beirami is a distinguished AI Research Leader with over 15 years of experience in developing innovative, efficient, and scalable AI products. He has held significant research roles at Google DeepMind, Google Research, Meta AI, and Electronic Arts, focusing on large language models (LLMs), reinforcement learning (RL), and conversational AI. github+2
His work has significantly advanced the capabilities of LLMs, including pushing the Gemini model's capabilities-compute Pareto frontier through projects like controlled decoding and InfAlign, and achieving 2x inference speedups for Gemini and Gemma via speculative decoding. He also led the development of CAIRaoke, Meta's first tool-using digital assistant. github+1
Academically, Ahmad is highly accomplished, with a PhD from Georgia Institute of Technology, where he received the Sigma Xi Best PhD Thesis Award for his work on network compression. He also held postdoctoral fellowships at MIT, Harvard, and Duke, focusing on the intersection of machine learning and information theory, and has a strong publication record with a high citation count (h-index 34, i10-index 73). github+3
He has been recognized with multiple awards, including Outstanding Paper Awards at the International Conference on Learning Representations (2025) and IEEE Computational Intelligence Society (2022). His expertise spans machine learning, natural language processing, statistics, information theory, optimization, federated learning, and multi-agent systems. github+2

Work

Education

Writing

Controlled Decoding from Language Models

January 1, 2024

Research on solving KL-regularized reinforcement learning (RL) problems in language model alignment while keeping the base model frozen through learning a prefix scorer.

Favicon imagebeirami.github.io

Speculative Decoding

January 1, 2023

Research achieving 2x inference speedups for Gemini and Gemma models.

Favicon imagebeirami.github.io

Network Compression via Network Memory: Fundamental Performance Limits

January 1, 2014

PhD thesis quantifiying data compression in finite-length data and proposing algorithms to approach the Shannon compression limit.

Favicon imageece.gatech.edu

InfAlign

Research on RL post-training and test-time compute scaling for Gemini model capabilities.

Favicon imagebeirami.github.io

CAIRaoke

A prototype of Meta's first tool-using digital assistant.

Favicon imagebeirami.github.io

Hobbies

Bronze Medalist in the Iranian Mathematics Olympiad (2002). github

Notes

KKKia Khezelisaid they've seen
ABAhmad Beirami
drive groundbreaking AI research or product development at a major tech company or top lab.

Nov 10, 2025