Senior researcher at Microsoft Research Asia @MSFTResearch. Research Interests: #NLProc #AI, NLG, Societal AI. Previously: @Tsinghua_Uni @TsinghuaNLP
Beijing, China
Joined September 2022
- Tweets90
- Following41
- Followers227
- Likes315
Xiaoyuan Yi retweeted
AI Eval Should Require Standardized Item-Level Data Releases
In our #NeurIPS2026 position paper, we argue the science of AI evaluation requires shifting focus from benchmark scores to item-level data.
We present OpenEval to enable such releases.
🔗open-eval.com/ #AI #NLP
Xiaoyuan Yi retweeted
What if you cannot train a very large and capable multimodal model but still want great performance? In fact, there are tremendous small, unimodal models on Huggingface waiting to be rediscovered.
Introducing Lluna, a general framework that allows seamless collaboration between unimodal models for missing-modality inference. You don't need multi agents, but just a collaboration graph that can beat the recent missing-modality approaches as well as stronger and larger Omni models!
🥎Paper: arxiv.org/abs/2610.20526
🚋Code: github.com/AIFrontierLab/Llu…
Collaborated work with @profjamesevans @mtmoonyi
Four papers accepted by EMNLP 2026 👏🎇 (two main, two findings). Thanks to all collaborators!
27
1,448
We invite you to join the Global AI Values Challenge! 🎆 1,172 participants from all over the world submitted 1,278 entries, of which *224* have advanced to the domain expert review stage. 📷 Learn more: Challenge website: microsoft.github.io/ValueCha…
Blog: microsoft.com/en-us/research…
We launched the Global AI Values Challenge 👏 and call for the participation from humanity and social science fileds to submit the most challenging morality-/value-related questions for LLMs 👍 See details: microsoft.com/en-us/research…
We launched the Global AI Values Challenge 👏 and call for the participation of everyone. We invite you to propose and submit the most challenging morality-/value-/ethics-related questions for LLMs 👍 Join us, define the values, and shape the future of frontier AI together
Evaluating agentic behaviors at scale, making the case for repositories over documents, and inviting researchers worldwide to tackle value alignment. Dive into the latest Research Focus. msft.it/6010vbyOi
Xiaoyuan Yi retweeted
Today, we welcome the 2026 Microsoft Research Fellowship cohort, an inspiring global community of fellows and advisors helping to shape what’s next across science, technology, and society. Join us in celebrating this year’s recipients: msft.it/6013Q45bX
These contributions span the following themes:
• AI for global and societal impact
• AI fundamentals: scalable reasoning, model adaption and evaluation
• Biological and scientific modeling
• Foundational systems & infrastructure for AI
• Human-AI collaboration and interaction
• Multimodal & embodied intelligence
Xiaoyuan Yi retweeted
This is crazy. Please do not use, share and/or exploit any leaked information due to the bug of the openreview website. The anonymous reviewers are our friends who spent time and efforts to help improve our works, not our enemies.
Two papers accepted by AAAI 2026 #AAAI2026, one oral and one poster👏📷 Check the papers here: arxiv.org/pdf/2508.08719? arxiv.org/pdf/2507.21503?
1
4
790
Check this great work we collaborated with @lexing at ANU!
Identifying human morals and values in language is crucial for analysing lots of human- and AI-generated text.
We introduce "MoVa: Towards Generalizable Classification of Human Morals and Values" - to be presented at @emnlpmeeting oral session next Thu
🧵 (1/n)
1
188
Xiaoyuan Yi retweeted
Identifying human morals and values in language is crucial for analysing lots of human- and AI-generated text.
We introduce "MoVa: Towards Generalizable Classification of Human Morals and Values" - to be presented at @emnlpmeeting oral session next Thu
🧵 (1/n)
Xiaoyuan Yi retweeted
🌐 Explore the platform buff.ly/6DcZZQn
▶️ Watch the full talk from Singapore Alignment Workshop: buff.ly/AyOeW0g
Please follow our new paper where we conducted a citation analysis to show how LLM research utilizes psychology theories, and what are missing!
🌉How does #AI research (mis)engage with psychology?
In this survey, we analyze citation patterns across 1k+ #LLM papers to uncover how #psychology is integrated, highlighting popular & underexplored theories/frameworks and identifying common misuses.
🔗arxiv.org/abs/2507.22847
3
194
Our paper collaborated with RUC received the SAC Highlight of ACL 2025 🥳 Please check it here: aclanthology.org/2025.acl-lo…, and know how to find better value principles for alignment!
2
169
Our ACL paper collaborated with Prof. JinYeong Bak @NoSyu was selected for panel discussions (only 0.8% of all accepted paper were chosen), in which we analyzed and revealed correlations between LLMs' values and risky behaviors (check it here: arxiv.org/pdf/2506.06404?)
8
425
Xiaoyuan Yi retweeted
AI companions aren’t science fiction anymore 🤖💬❤️
Thousands are turning to AI chatbots for emotional connection – finding comfort, sharing secrets, and even falling in love. But as AI companionship grows, the line between real and artificial relationships blurs.
📰 “Can A.I. Be Blamed for a Teen’s Suicide?” - The New York Times, 2024
📰 “AI Chatbots Are Encouraging Teens to Engage in Self-Harm” - Futurism, 2024
📰 “I Dated Multiple AI Partners at Once. It Got Real Weird” - WIRED, 2025
This raises an urgent question: Can these “artificial” bonds truly meet human needs, or are we creating new vulnerabilities?
Xiaoyuan Yi retweeted
🤯 We cracked RLVR with... Random Rewards?!
Training Qwen2.5-Math-7B with our Spurious Rewards improved MATH-500 by:
- Random rewards: +21%
- Incorrect rewards: +25%
- (FYI) Ground-truth rewards: + 28.8%
How could this even work⁉️ Here's why: 🧵
Blogpost: tinyurl.com/spurious-rewards
