User simulators are only as faithful as the user representations they are conditioned on. Missing details force simulators to fill gaps with assumptions, leading to generic behavior⚠️ Introducing UserRepGen-4B, which generates representations optimized for simulation fidelity.

Why are LLM parameter updates so sparse during post-training? 🔍 In our new paper, we study this phenomenon through the lens of Fisher information, connecting theory and empirical findings across RL, SFT, and on-policy distillation. A thread on what we found 🧵👇

Grateful to Microsoft Research for featuring our work, StudentSim! Since its release, we've been amazed by the attention it has received, not only from the education community, where professors have called it "the very first foundational work in this area" and shared that they'd envisioned this direction long before data and training made it possible, but also from researchers and project leads working on user simulation in recommender systems, healthcare, and financial services. It's exciting to see a shared recognition of the importance and potential of user simulation for personalized AI services. Hope StudentSim can be a useful stepping stone for the community!

StudentSim is a research system for evaluating and improving AI tutoring tools before they are tested with learners. It models how individuals may engage with tutors, giving feedback to improve tutor performance across subjects like chess, writing, and mathematics: msft.it/6013aS3JF

PSI-Bench won the best paper award at Social Simulation with LLMs workshop at @COLM_conf ! We introduce a framework for interpretable and clinically meaningful evaluation of depression patient simulators. Read more: arxiv.org/abs/2604.25840

In 1995, Wooldridge and Jennings listed proactiveness among the basic properties of intelligent agents. 31 years later, our agents can reason, code, and run complex workflows. Yet someone still has to notice the work and tell them to start. We think this is a major gap. 1/7 It's easy to miss the problem when agents are judged on a task they've already been given. A long open-ended project is different. Instructions are incomplete, plans fail, priorities change. Humans still decide what to investigate and when to change course. That won't scale forever. 2/7 There's also a product shift underway. Meta's Muse and OpenAI's dots both point toward proactive agents that keep track of goals, anticipate needs, and work beyond a single chat session. We expect users to ask less often for individual tasks and more often for continuing progress. 3/7 Proactiveness isn't a new idea. Relevant work spans agent theory, planning, HCI, robotics, ubiquitous computing, and LLM agents. Communities have developed different terms and benchmarks for related problems. It's surprisingly hard to get a clear picture of the field. 4/7 We felt this was a good moment to put those threads together. 23 of us from UIUC, Google, Stanford, UC Berkeley, and CMU worked on a survey, with backgrounds spanning agents, robotics, retrieval, computational biology, and finance. 5/7 We tried to write it differently from a bibliography dump. Chapters were drafted by people closest to the topic, then discussed and revised across backgrounds. When searching for papers is becoming cheap, the harder work is deciding which connections are meaningful. 6/7 The resulting survey organizes proactive agents around 6 capacities: awareness, anticipation, agenda formation, arbitration, action, and adaptation. We also review evaluation at three levels and compare initiative with action authority across 13 domains. 7/7 We don't expect this to be the final map of the area. We'd be happy if it helps people find a question worth working on, or shows them where to disagree with us. Proactive AI for Superintelligent Agents empathyang.github.io/Proacti… research.google/pubs/proacti… #ai #agent #survey #uiuc #google #stanford #ucb #cmu

🤔 AI agents can do a lot once you give them a task. Can they also figure out what’s worth doing and get started before being asked? 🏫 That’s the question behind Proactive AI for Superintelligent Agents, our joint survey by researchers from UIUC, Google, Stanford, UC Berkeley, and CMU. Our backgrounds span AI agents, embodied AI, information retrieval, computational biology, economics, and finance. 🕵 As finding papers gets cheaper, we think a survey’s value increasingly comes from hard-earned research intuition: knowing what matters and what to try next. Contributors drafted the sections closest to their own research, then refined them through several rounds of discussion and revision across research backgrounds. 📃 Paper: empathyang.github.io/Proacti… 🌎 Website: empathyang.github.io/Proacti… 💼 Google Research: research.google/pubs/proacti… More in the video below 👇

✨ Creativity is not just recognizing what an object is — it is imagining what it could become. 🔧 A key edge can cut tape. 🛡️ A rubber pad can protect a wall. 🪮 A comb guard can clear a sink slot. But can multimodal AI agents discover these hidden physical affordances from images? 🚀We introduce MM-CreativityBench, a benchmark designed to test whether LMMs can creatively repurpose everyday objects by interactively inspecting scenes, entities, and object parts. 🔍 Our findings show that today’s LMMs often identify the right object, but fail to ground their reasoning in the right part. They hallucinate properties, overlook physical constraints, or propose solutions that are not mechanically valid. 🧠 To move beyond plausible guesses, we propose affordance-grounded alignment: training models to explore visual evidence, reject hallucinated affordances, and reason from geometry, material, and mechanics. 📄 Paper: arxiv.org/pdf/2605.26396 🌐 Project: creativitybench.github.io/mm… 💻 Code: github.com/CreativityBench/M… 🤗 Hugging Face: huggingface.co/papers/2605.2…

What happens when you compare the distributions of real and simulated user behaviors? 🔍 The gap is large. We introduce a method to measure this gap and evaluate 24 LLM-based user simulators across coding and writing tasks. @convai_uiuc @MSFTResearch @berkeley_ai 🧵 1/N

Excited to share that our paper, “Learning to Persuade Exposes How Easily LLMs Abandon Correct Beliefs,” has been accepted to NeurIPS 2026! 🎉