Sabitlenmiş Tweet
steven
185 posts
steven
@stevenlu0
cs & ai policy @berkeley_ai, phd’ing soon at @SCSatCMU
Berkeley, CA Katılım Haziran 2025
131 Takip Edilen155 Takipçiler
Would you trust an AI agent to negotiate on your country's behalf at the G20?
Real coordination is long-horizon, asymmetric, and non-binding; current multi-agent evaluations miss this.
We build Cooperate to Compete (C2C): a testbed for LM agents coordinating with rivals. 🤝🔪🎭
English
i'm co-organizing a workshop on AI governance!
we'll have student presentations in the morning, then various presentations in the afternoon ft. CA State Sen. Jerry McNerney, Prof. Suresh Venkatasubramanian, speakers from DeepMind, CCST, Mila & more! register for free food 😋
English
call me a nerd but pbs newshour is literally my favorite show on tv. very well deserved!
Lisa Desjardins@LisaDNews
Peabody! Incredible honor for @NewsHour and our team coverage of immigration. Could not be prouder of the work we - and mostly those below - have done. Among the congratulations to: @WmBrangham, @lbarronlopez, @ElizLanders, @TheStephSy, @IAmAmnaNawaz, @GeoffRBennett, @mattloff, Elizabeth Summers, @ecarpeaux, @KyleMidura, @shraipopat, @mikewfritz, Jonah Anderson, @DougAAdams, @newshourfred, @sarajust among many. pbs.org/newshour/press…
English
steven retweetledi
🎉 Thrilled to have two papers accepted to ACL 2026 main!
1. Graph-based models match LLMs on close-ended human simulation tasks with far less compute & greater transparency
2. (oral) How to allocate human samples towards fine-tuning vs post-hoc rectification in simulation
English
@chrisalbon wouldn’t this be because of the I-80 closures this weekend?
English
steven retweetledi
🎉 Excited to share that GEMS is accepted to ACL 2026 main!
We show that a lightweight GNN can match or outperform LLMs at simulating human behavior in discrete-choice settings — with multiple advantages, including efficiency and transparency.
Paper: arxiv.org/abs/2511.02135
Joseph Jeesung Suh@JosephJSSuh
LLMs have dominated recent work on simulating human behaviors. But do you really need them? In discrete‑choice settings, our answer is: not necessarily. A lightweight graph neural network (GNN) can match or beat strong LLM-based methods. Paper: arxiv.org/abs/2511.02135 🧵👇
English
Last year when working with WildChat we needed to do painstaking cleansing to get interaction transcripts with real epistemic content.
Now there's this new dataset for human-AI epistemic interaction in the wild, from @grok interactions on X. Lovely work from Matteo's team!
Tianyi Alex Qiu@Tianyi_Alex_Qiu
As wonderful a dataset as WildChat is, people around me have complained how hard it is to do meaningful analysis on it. Which I totally agree with. As such, we are open-sourcing WildChat-curated, where we filter away bots, extract 5.4M concepts (0.2M of which are juicy politics/philosophy) into a hierarchy, and create user/time/concept-wise dataframes containing stats and diversity measures. We hope this ready-for-analysis version of WildChat can make the lives of HCI and AI safety researchers a bit easier. Check it out and let us know if it helps! As always, thanks a ton to the WildChat team, whose amazing work made this possible in the first place.
English
How should we talk about LLMs? Does it matter if we frame them as a machines 📠, tools ⚒️, or companions 👥? In our #CHI2026 paper, that these framings can alter what people believe about LLMs and how they use them. See 🧵for more!
English
@shanli_xing @CarnegieMellon @tqchenml @ericxing @uwcse @UWSyFi congrats, and hope to see u in the fall! 😁
English
Super excited to share that I'll be joining @CarnegieMellon as a PhD student, working with @tqchenml and @ericxing!
It has been a wonderful journey at @uwcse @UWSyFi learning and building systems that power frontier AI in production. I want to express my sincerest gratitude to @ye_combinator @tqchenml @luisceze for all the opportunities and guidance along the way, and to many others at UW and CMU who have been hugely encouraging, supporting, and intellectually inspiring me. I wouldn't have made it this far without all of you.
Looking ahead, I'm eager to explore how AI-system co-design can advance the capabilities of both sides. On the system side, I believe better abstractions and verification signal design can enable the AI-driven cycle for system improvements. On the model side, I'm interested in how to enable models to perform well in long-horizon, sparse-goal tasks that require periodic knowledge consolidation, like doing system research itself.
Always happy to chat and collab!
Keep building 🤟
English
In Barcelona for #chi2026! Presenting our work on eliciting LLMs' assumptions about users, and how this mismatches with user expectations, in the Tues poster session! (Spoiler: users assume that LLMs give objective info much more than they actually do --> sycophancy 😢)
English
We're launching an international, cross-sector Delphi study to establish consensus on conducting and reporting AI evaluations.
All critical infrastructure—from bridges and aircraft to pharmaceuticals—has agreed-upon, rigorous evaluation standards. AI systems will be at least as consequential, yet current practices are uneven, siloed, and hard to compare across organizations and contexts.
We need voices from frontier labs, auditors, academia, policymakers, civil society, and industry practitioners to create a shared reference.
English