# Xiao Ma

> Alignment researcher at Meta Superintelligence Labs since April 2026. Recently led alignment of the image and video planner agent for Meta models as technical lead. Her work spans post-training, reinforcement learning, and scalable safety evaluations. Critical contributor to the July 2026 Muse Image launch. Previously a Staff Software Engineer at Google DeepMind (2024–April 2026). Co-invented and empirically validated RLGF (grading feedback), launching the method in late 2024 to unblock Gemini 2.5 Pro and beyond, then scaling adoption across Gemini. Co-author of an ICLR 2025 Outstanding Paper on safety alignment. Her human–AI interaction and social-trust research informs her interest in security and social engineering risks in emerging hybrid human–agent networks, and in establishing protocols for trust within them.

## Highlights

- Recently led alignment of the image and video planner agent for Meta models as technical lead
- Personally launched alignment training jobs incorporating data and improved rewards
- Established safety evaluations and onboarded the full safety-evaluation suite onto scalable infrastructure
- Critical contributor to the [Muse Image launch](https://about.fb.com/news/2026/07/introducing-muse-image-meta-ai/) in July 2026, with no content-safety escalations at launch
- Core contributor to Gemini from its inception
- Co-invented, empirically validated, and launched RLGF (grading feedback) to unblock Gemini 2.5 Pro and beyond
- Scaled RLGF adoption across Gemini through technical documentation, a playbook, and a summit
- ICLR 2025 Outstanding Paper Award (AI safety alignment)
- Multiple best paper awards
- Patent granted for ExploreLLM (task decomposition for agentic behavior)
- Board member, National Sawdust (art and technology)

## Research Areas

- Recent work: alignment of the image and video planner agent for Meta models, in a technical lead capacity
- Research interests: agent security, emerging hybrid human–agent network dynamics, and human–agent alignment
- Research concerns: security and social engineering risks
- Post-training and instruction following for large language models
- RLGF (grading feedback): method co-invention, empirical validation, launch, and adoption across Gemini
- Long-horizon reinforcement learning
- Reinforcement learning from human and critic feedback (RLHF / RL*F)
- Reward modeling and critic feedback at scale
- AI safety and alignment
- Foundation model development (Gemini)
- AI-mediated communication and societal trust (prior work)
- Human-computer interaction and computational social science (prior work)

## Technical Keywords

Large language models, LLM post-training, reinforcement learning, RLHF, reward modeling, instruction following, scalable safety evaluations, agent safety, AI alignment, Muse Image, Gemini, multimodal models, human–agent alignment, hybrid human–agent networks, agent security, security risks, social engineering risks, human–AI interaction, computational social science, trust

## Selected Publications

- [Gemini 2.5 Technical Report](https://arxiv.org/pdf/2507.06261) (2025): Pushing the frontier with advanced reasoning, multimodality, long context, and agentic capabilities
- [Safety Alignment Should Be Made More Than Just a Few Tokens Deep](https://arxiv.org/abs/2406.05946) (ICLR 2025, Outstanding Paper Award)
- [Beyond ChatBots: ExploreLLM for Structured Thoughts and Personalized Model Responses](https://arxiv.org/abs/2312.00763) (CHI 2024, Patent Granted): Early work on task decomposition, leading up to agentic behavior
- [Gemini: A Family of Highly Capable Multimodal Models](http://maxiao.info/research) (2023)

## Education

- PhD in Information Science, Cornell Tech — focused on networked trust among humans
- BS in Microelectronics, Peking University (2010–2014) — focused on chip design and manufacturing

## Art & Angel Investing

Runs a mini venture studio exploring AI agents and human agency through installations and selective angel investments. Investments include [alphaXiv](https://www.alphaxiv.org/) and [Design Arena](https://www.designarena.ai/). Her arts work includes board service at National Sawdust and digital preservation at Rhizome.

## Availability

- Potentially open to: speaking engagements, advisory roles, research collaborations
- Contact: x8.assist@gmail.com — please include "Brian Eno" in the email subject line
- Location: Palo Alto, CA

## Pages

- [Home](http://maxiao.info/): Bio, contact information, and background
- [Publications](http://maxiao.info/research): Full list of academic publications
- [Media & Talks](http://maxiao.info/media): Media coverage and speaking engagements
- [CV (PDF)](http://maxiao.info/xm-cv.pdf): Curriculum vitae
- [CV (Markdown)](http://maxiao.info/xm-cv.md): Machine-readable curriculum vitae

## Links

- [Google Scholar](https://scholar.google.com/citations?user=xLPxJsYAAAAJ&hl=en)
- [alphaXiv profile](https://www.alphaxiv.org/@xiao-ma)
- [LinkedIn](https://www.linkedin.com/in/xiaom)
- [Twitter/X](https://twitter.com/infoxiao)
- [Substack](https://xiao.substack.com)