research scientist, gemini post-training @googleresearch, phd @columbia, ba @cornell
Washington, DC
Joined July 2021
- Tweets156
- Following296
- Followers467
- Likes1.2K
Pinned Tweet
PhD Candidate ➡️ PhDone!
Excited to share that after 3.5 years of internship experience and earning 4 degrees, I have successfully defended my PhD thesis @ColumbiaCompSci and started working @GoogleAI!
Huge thank you to @Zhou_Yu_AI for her unwavering support throughout my PhD!
Max Chen retweeted
as we grow increasingly accustomed to relying on llm-generated reports to understand agent-written experiments and code, it's worth asking how much trust we should be placing in these llm-generated reports.
hear about our findings in "language models are 'insecure' reporters" covered in the latest episode of @Last_Week_in_AI hosted by @jeremiecharris and @andrey_kurenkov 🎙️
(segment starts 1:01:30) open.spotify.com/episode/23D…
📣 Come check out our work on quantifiably steering model behavior! 📣
We’ll be presenting a poster at 11am on Wednesday.
Our team at Google is also hiring! If you are working on long-horizon, personalization, or user models, please DM/visit the Google booth/our poster
#COLM
Is your LLM 𝒊𝒏𝒔𝒆𝒄𝒖𝒓𝒆?
Find out in this new paper led by our amazing student researcher @jennyihuang !!
Max Chen retweeted
ThoughtTrace just hit 5K+ downloads within two weeks of release!
TT is a simple, ready-to-use dataset with surprisingly diverse applications for understanding users and training AI models. Check it out!
What are users thinking during their interactions with LLMs?
We introduce ThoughtTrace — the first large-scale dataset that captures what users think during real-world human–AI conversations, not just what they type.
→ 10,174 thought annotations
→ 2,155 multi-turn conversations, 17,058 turns
→ 1,058 users
→ 20 LLMs
These thoughts improve user behavior prediction (+41.7%) and model alignment (+25.6%).
This opens a new paradigm of user-centric LLM research. Full information in the thread 🧶
Read our paper: arxiv.org/abs/2605.20087
Check our project website: thoughttrace-project.github.…
📣 Check out our new work led by the amazing @chuanyang_jin ! 📣
We built a novel framework to reveal what users are thinking in diverse real world Human-AI interactions. Deepening our understanding of user thoughts is crucial to improving Theory of Mind and model alignment!
What are users thinking during their interactions with LLMs?
We introduce ThoughtTrace — the first large-scale dataset that captures what users think during real-world human–AI conversations, not just what they type.
→ 10,174 thought annotations
→ 2,155 multi-turn conversations, 17,058 turns
→ 1,058 users
→ 20 LLMs
These thoughts improve user behavior prediction (+41.7%) and model alignment (+25.6%).
This opens a new paradigm of user-centric LLM research. Full information in the thread 🧶
Read our paper: arxiv.org/abs/2605.20087
Check our project website: thoughttrace-project.github.…
📣Excited to finally share our latest work on quantifiably adapting model behavior based on unique preferences 📣
We teach language models to adjust their clarification behavior using scalar coefficients and find they can generalize to unseen coefficients at inference time!
Max Chen retweeted
Our most anticipated launch of the year is here.
- Gemini 3, our most intelligent model
- Generative interfaces, for perfectly designed responses
- Gemini Agent, made to complete complex tasks on your behalf
See how Gemini 3 can help you learn, build & plan anything 🧵
Max Chen retweeted
Recruiting PhDs & postdocs for:
🤖 agents "taking over" science (hypogenic.ai and 📌)
🧪 Real scientists ➡️AI (e.g., materials, chem, physics)
📜 Theory + incentives for H-AI collab & credit (e.g., formalizing tacit knowledge)
new adventures for me, 🔄 if you can! 🙌
Max Chen retweeted
Colab Pro is now available for free for verified US students and faculty for one year.
Big news for data science in higher ed! 🚀Colab now offers 1-year Pro subscriptions free of charge for verified US students/faculty, interactive Slideshow Mode for lectures, & an AI toggle per notebook. Enhance teaching & learning in the upcoming academic year! Read all about it here: goo.gle/40pxnl1. #GoogleColab #EdTech.
Really excited to share our #ICLR2025 paper!
We propose ACT, an approach for improving LLMs’ ability to navigate multi-turn conversations via implicit action optimization.
Our LLMs are better equipped to achieve multi-turn goals and recognize when to ask clarifying questions!
Today on the blog we propose Action-Based Contrastive Self-Training, a data-efficient #ReinforcementLearning tuning approach for improving multi-turn conversation modeling in mixed-initiative LLM interaction. Read all about it →goo.gle/3Sxas2T
ALT A conversational agent capable of disambiguation could recognize when there is ambiguity and ask a clarifying question towards a more accurate final answer.
Excited to be attending ICLR in person!
I’ll be presenting joint work between @GoogleAI and @columbianlp on improving LLMs’ ability to navigate multi-turn conversations via clarifying questions in the Friday morning poster session: Hall 3 + Hall 2B #399
This work wouldn’t have been possible without my amazing collaborators: @ruoxi_cc @tomaspfister @sercanarik
Check out our paper here: openreview.net/pdf?id=SIE6VF…
Sad to be missing EMNLP this week! Definitely reach out to @qbetterk if you want to chat about dynamic LLM benchmarking: aclanthology.org/2024.findin…
We find that frontier LLMs' performance on math and language reasoning benchmarks degrades even from simple variable perturbations.
I am attending EMNLP 2024 in person! Come and talk if you are interested in our recent work in robust LLM evaluation: VarBench(arxiv.org/abs/2406.17681). #EMNLP2024
