Online groups and communities often need to make decisions around social concepts like what content is appropriate. But how do we ensure these decisions are aligned across human decision-makers or even AI systems? We explore this in our work (CI '25): 📜 Case Law Grounding ⚖️

Should you RL your compound AI system or optimize its prompts? We think both! 🤯 A short preview of work co-led with Noah Ziems and Lakshya A Agrawal!👇

We introduce MoVer, a Motion Verification DSL that automatically checks if AI-generated motion graphics animations match your text prompts! We make it easy for designers to specify and verify complex animations with LLM-powered iterative refinement. Catch our #SIGGRAPH2025 talk:

Self-presentation is multifaceted, but the expression is often limited to physical accessories. How could Audio AR transform social interaction? We introduce Audio Personas, body-anchored sounds to dynamically shape social impression. Upcoming in TOCHI: arxiv.org/pdf/2505.00956

"GPT-5 shows scaling are coming to an end"

The Echoes in AI paper showed quite the opposite with also a story continuation setup. Additionally, we present evidence that both *syntactic* and *discourse* diversity measures show strong homogenization that lexical and cosine used in this paper do not capture.

Omar Shaikh Dilara Soylu RISE and GRIND Omar Shaikh

Soon, AI agents will act for us—collaborating, negotiating, and sharing data. But can they truly protect our privacy? We simulate privacy-critical scenarios, using alternating search to evolve attacks and defenses, uncovering severe vulnerabilities and building protections.

I’m excited to share our new nature paper 📝, which provides strong evidence that the walkability of our built environment matters a great deal to our physical activity and health. Details in thread.🧵 nature.com/articles/s4158…

New Paper Day! For EMNLP findings—in LM red-teaming, we show you have to optimize for **both** perplexity and toxicity for high-probability, hard to filter, and natural attacks!

📢 Belated update, but I'm thrilled to share that I've joined The University of Chicago Chicago Booth as an Assistant Professor in the newly created Applied AI group! I'll continue to work on behavior-bound machine learning: understanding how AI shapes, is shaped, and should be shaped by the

New paper! We explore a radical paradigm for AI evals: assessing LLMs on *unsolved* questions. Instead of contrived exams where progress ≠ value, we eval LLMs on organic, unsolved problems via reference-free LLM validation & community verification. LLMs solved ~10/500 so far:

Introducing Generative Interfaces - a new paradigm beyond chatbots. We generate interfaces on the fly to better facilitate LLM interaction, so no more passive reading of long text blocks. Adaptive and Interactive: creates the form that best adapts to your goals and needs!

🔍 How do we teach an LLM to 𝘮𝘢𝘴𝘵𝘦𝘳 a body of knowledge? In new work with AI at Meta, we propose Active Reading 📙: a way for models to teach themselves new things by self-studying their training data. Results: * 𝟔𝟔% on SimpleQA w/ an 8B model by studying the wikipedia

Maybe this is part of why I find GPT-5 on ChatGPT so annoying -- apparently its system prompt is explicitly set to *not* ask clarifying questions?!? I find it really annoying the way it just goes off and tries to solve the world in one shot. I really want to iterate!

I led a session at @mila_quebec HCAI reading group on Omar Shaikh et al.'s General User Model paper! So excited to find and present this paper to our group. We discussed both the paper's novelty and people's concerns, as well as technical approches to address them.