Adrián Bazaga's Homepage

Welcome to my website!

Background:
Hi, I’m Adrián, a Senior Researcher, Tech Lead (SLM) and Post-Training Lead at Microsoft. I’m an established researcher with significant experience in model pre-training, post-training, reasoning, test-time scaling, tool-use, alignment, agents, LLM research, as well as reinforcement learning. I lead post-training for Aion Instruct, Microsoft’s foundational language model for Windows, unveiled by Satya Nadella at Microsoft Build 2026 and shipping on Windows devices, owning the end-to-end stack from data preparation → pretraining → mid-training → post-training → evaluation, and shipping agentic experiences directly on user devices for millions of users at global scale. My post-training work spans supervised fine-tuning (SFT), direct preference optimization (DPO), reinforcement learning (RL), verifiable rewards, reward modeling, and agentic tool-use, driving human alignment across response quality, conciseness, factuality, and safety.

I led post-training for Mu, Microsoft’s blazing-fast on-device SLM, driving the design and implementation of its training pipelines. I also led the development of the Windows Settings AI agent, already live on Windows Copilot+ devices. These efforts are part of my broader vision to enable seamless, deeply integrated AI experiences for everyone.

I hold a Ph.D. in Machine Learning from the University of Cambridge, where I conducted research under the supervision of Prof. Pietro Liò and Prof. Gos Micklem. My work has been published in leading Machine Learning conferences such as ICLR, ICML, ACL and EMNLP, as well as in Nature journals. Previously, I gained research experience through internships at Microsoft Research and Amazon AGI, where I explored novel training schemes to enhance few-step generation in diffusion models, and test-time scaling for temporal reasoning with Large Language Models (LLMs). Prior to that, I spent ~5 years in various startups, working at the intersection of AI and biology.

Research Interests:
My research focuses on advancing the capabilities of AI in the areas of foundational LLMs, reasoning, tool usage, and multimodality. Currently, I’m broadly interested in devising efficient SLM architectures, inventing data-efficient optimization techniques, advancing post-training with reinforcement learning (RL), verifiable rewards, and reward modeling, expanding agentic tool-use, and ensuring human alignment across response quality, conciseness, factuality, and safety.

Beyond Research:
In addition to my core research, I’d like to explore how generative models can improve education and governance. If you’re working on high-impact, real-world deployments in these areas, I’m always open to collaborate 👐.