Ali Hatamizadeh.
Ali Hatamizadeh

Ali Hatamizadeh

LLM Tech Lead & Staff Research Scientist at NVIDIA

I am the LLM Tech Lead and a Staff Research Scientist at NVIDIA, where I lead research on large language models. My work spans two threads: the design of novel LLM architectures, and democratizing reinforcement learning for pre-training and post-training.

I co-invented Gated DeltaNet, a linear-attention model now adopted by several leading open-source LLMs, including Qwen3.5, Qwen3-Next, Qwen3-Coder-Next, and Olmo Hybrid.

I received my PhD and MSc in Computer Science from the University of California, Los Angeles (UCLA).

I am always glad to hear about new opportunities and collaborations — feel free to reach out.

Google Scholar GitHub LinkedIn X / Twitter Email

What I work on

Research

My research centers on large language models, organized around two complementary threads — architecture design and reinforcement learning.

01

Model Design

The goal of this thread is to invent new hybrid architectures for LLMs. I co-invented Gated DeltaNet, a linear-attention model that has been adopted by several leading open-source models — including Qwen3.5, Qwen3-Next, Qwen3-Coder-Next, and Olmo Hybrid — and has informed the design of related architectures such as Kimi Linear (KDA).

Selected publications

02

Democratizing RL for Pre-training & Post-training

This thread aims to make reinforcement learning a practical and scalable ingredient across the full LLM training pipeline — from pre-training through post-training — rather than a costly final-stage add-on. Work here includes RLP, which reframes reinforcement as a pre-training objective, and iGRPO, a self-feedback-driven approach to LLM reasoning.

Selected publications

Updates

News

  • May 21, 2026 Gated DeltaNet-2 released — decoupling erase and write in linear attention.
  • Mar 10, 2026 Released NVIDIA Nemotron 3 Super — an open MoE hybrid Mamba-Transformer model for agentic reasoning.
  • Jan 26, 2026 RLP accepted at ICLR 2026.
  • Jan 22, 2025 Gated DeltaNet accepted at ICLR 2025.

Get in touch

Contact

Open to new opportunities, discussions and collaborations.

ahatamiz123@gmail.com