PeopleofSuperintelligence

Researcher

Arash Ahmadian

Arash Ahmadian is a research scientist at Google DeepMind who works on scalable reinforcement learning and post-training recipes. His research has covered quantization, reinforcement learning from human feedback, and multilingual alignment. He co-authored papers on a simpler REINFORCE-style approach, Aya Expanse, and Command A. His 2023 quantization study reported minimal performance loss on models with up to 52 billion parameters.

Works
18
Citations
123
h-index
6
Badges

Publication numbers from OpenAlex, as of 18 October 2026.

Timeline

  1. 2023

    Quantizing large language models

    In 2023, he co-authors “Intriguing Properties of Quantization at Scale,” which studies how pretraining choices affect large-model quantization. The team reports minimal performance loss when quantizing models from 410 million to 52 billion parameters.

  2. March 2024

    A Cohere research role

    In March 2024, he is a researcher at Cohere and Cohere for AI.

  3. August 2024

    A simpler RLHF recipe

    In August 2024, he and seven co-authors publish “Back to Basics,” which finds that REINFORCE-style optimization variants outperform PPO, DPO, and RAFT for RLHF.

  4. November 2024–December 2024

    Multilingual alignment and Aya Expanse

    In November 2024, he co-authors “The Multilingual Alignment Prism,” which studies global and local safety preferences and reports alignment across six languages with minimal performance loss. In December 2024, he co-authors the Aya Expanse report, which introduces multilingual 8B and 32B models.

  5. April 2025

    Command A technical report

    In April 2025, he co-authors the Command A technical report, which describes a multilingual model for enterprise use with retrieval, grounding, and tool use.

  6. As of October 2026

    Research scientist at DeepMind

    As of October 2026, he is a research scientist at Google DeepMind working on scalable reinforcement learning and post-training recipes.

Career

  1. NowGoogle DeepMindResearch Scientist
  2. 2022–2024University of Toronto
  3. CohereResearcher

From the biography and OpenAlex affiliations (the years of papers that list the organization), so some years are approximate.

Most-cited works

  1. Back to Basics: Revisiting REINFORCE-Style Optimization for Learning from Human Feedback in LLMs2024 · Annual Meeting of the Association for Computational Linguistics (ACL) · 41 citationsWith Sara Hooker
  2. Subspace learning using structure learning and non-convex regularization: Hybrid technique with mushroom reproduction optimization in gene selection2023 · Computers in Biology and Medicine · 17 citations
  3. Dual regularized subspace learning using adaptive graph learning and rank constraint: Unsupervised feature selection on gene expression microarray datasets2023 · Computers in Biology and Medicine · 16 citations
  4. Pushing Mixture of Experts to the Limit: Extremely Parameter Efficient MoE for Instruction Tuning2023 · 16 citationsWith Sara Hooker
  5. Subspace learning for feature selection via rank revealing QR factorization: Fast feature selection2024 · Expert Systems with Applications · 14 citations

From OpenAlex, as of 18 October 2026.

  1. The median h-index is 54 at universities and 20 at frontier AI labsIncludes Arash Ahmadian. The median h-index is 54 at universities and 20 at frontier AI labs. We compare publishing careers, citations and moves to estimate how universities and frontier labs contribute to AI research.
  2. 14% of AI career moves in the 2020s went from a university to a companyIncludes Arash Ahmadian. Our estimates of how university-to-company career transitions changed by decade, including roles added alongside an earlier job.
  3. Google has the highest combined h-index among AI organizationsIncludes Google. Our estimates of combined and median h-index at AI organizations, using current affiliations and OpenAlex records.
  4. Google alumni founded or led 82 other organizations, the most among the labs we compareIncludes Google. We trace 189 organizations founded or led by AI lab alumni, and 196 career moves between the labs.

Colleagues

People with dated positions at the same organization in the same years.

Sources and credits

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

  1. Intriguing Properties of Quantization at Scale arxiv.org. Checked 18 October 2026.
  2. Intriguing Properties of Quantization at Scale Paper cohere.com. Checked 18 October 2026.
  3. TalkRL: The Reinforcement Learning Podcast | Transcript: Arash Ahmadian on Rethinking RLHF talkrl.com. Checked 18 October 2026.
  4. Back to Basics: Revisiting REINFORCE-Style Optimization for Learning from Human Feedback in LLMs aclanthology.org. Checked 18 October 2026.
  5. The Multilingual Alignment Prism: Aligning Global and Local Preferences to Reduce Harm aclanthology.org. Checked 18 October 2026.
  6. Aya Expanse: Combining Research Breakthroughs for a New Multilingual Frontier arxiv.org. Checked 18 October 2026.
  7. Command A: An Enterprise-Ready Large Language Model arxiv.org. Checked 18 October 2026.
  8. Arash Ahmadian arashahmadian.io. Checked 18 October 2026.