Accepted as NeurIPS2025 Spotlight! Existing large multimodal models (LMMs) have very poor visual understanding and reasoning over part-level attributes and affordances (only 5.9% gIoU). We developed novel part-centric LMMs to address these challenges arxiv.org/pdf/2505.20759

🚀 Introducing UserRL: a new framework to train agents that truly assist users through proactive interaction, not just chase static benchmarking scores. 📄 Paper: arxiv.org/pdf/2509.19736 💻 Code: github.com/SalesforceAIRe…

*Human-Like* Creativity is perhaps the most out-of-reach task for modern LLMs I'm super excited to share our new work evaluating LLMs with a creativity framework! We develop a synthetic creativity task to measure LLMs' capabilities in generating novel, creative, combinations,

Please note that #EMNLP2025 volunteer notifications have been sent. If you haven’t received yours, please check your spam folder or contact the chairs at [email protected] as some email addresses were entered incorrectly in the form

Very excited to see Tinker released by Thinking Machines! Even more thrilled that Search-R1 is featured as the tool-use application in Tinker’s recipe 👇 🔗 github.com/thinking-machi… When we first built Search-R1, we opened up everything—data, recipes, models, code, logs—and kept

Ever felt like your GUI agents are dragging their feet? 🧐The culprit? Crunching through endless streams of screenshots, especially in those marathon long-horizon tasks. Thrilled to unveil ⭐️ GUI-KV ⭐️— our plug-and-play powerhouse that taps into the spatial saliency within

🚨 New paper alert at COLM 2025! 🚨 An interesting open problem for those into Sparse Autoencoders (SAEs): "Top-K activation constrains L0 (the number of non-zeros), but how do we obtain E[L0]?" This was even the very first limitation noted in Leo Gao’s recent paper. (1/8)

🚨 New preprint out! 🚨 In "BAP v2: An Enhanced Task Framework for Instruction Following in Minecraft Dialogues", we work towards a core AI challenge: how can agents follow complex, conversational instructions in a dynamic 3D world? To this end, we introduce an enhanced task

Thrilled to announce that I'll be joining UIUC CS Siebel School of Computing and Data Science as an Assistant Professor in Spring 2026! 📢 I’m looking for Fall '26 PhD students who are interested in the intersection of Software Engineering and AI, especially in LLM4Code and Code Agents. Please drop me an

🚀 Introducing BroRL: Scaling Reinforcement Learning via Broadened Exploration When step-scaling hits a plateau, scale rollouts, not steps. BroRL takes reinforcement learning beyond saturation—reviving stalled models by expanding exploration with large-N rollouts. 👇 (1/n)

Multimodal conversational agents struggle to follow complex policies, which also impose a fixed computational cost. We ask: 👉 How can we achieve stronger policy-following behavior without having to include policies in-context? 🌐: mikewangwzhl.github.io/TriMPI/ 🧵1/3

World Model Reasoning for VLM Agents (NeurIPS 2025, Score 5544) We release VAGEN to teach VLMs to build internal world models via visual state reasoning: - StateEstimation: what is the current state? - TransitionModeling: what is next? MDP → POMDP shift to handle the partial

Today, we’re overjoyed to have a 25th Anniversary Reunion of Stanford NLP Group. So happy to see so many of our former students back at Stanford University. And thanks to Stanford HAI for the venue!

Zora Wang 's research compares AI agents vs humans across real work tasks (data analysis, engineering, design, writing). Key findings: 👉Agents are 88% faster & 90-96% cheaper 👉BUT produce lower quality work, often fabricate data to mask limitations 👉Agents code everything,

🤖💬AI agents can be easily persuaded (like Anthropic’s Claudius often giving discounts). 🤔Previous study on persuasion has been exclusively on text-only modality. We wonder: are AI agents more susceptible when presented with multimodal content? Introducing MMPersuade, a

Many of my students cannot attend EMNLP in person due to visa problems, but the super rising star Cheng Qian Cheng Qian @ EMNLP2025 will be there presenting multiple papers. Please drop by our posters and talk to him!

What if your policy could reason and think dynamically, especially about uncertainty, enabling better real-world behavior? ⚡️Introducing EBT-Policy, an instantiation of Energy-Based Transformers for Policies! TLDR: - EBT-Policy broadly outperforms Diffusion Policy in both

🏠🤖 Are our household robotic agents actually safe? BEAT introduces the first visual backdoor attack on MLLM-based embodied agents, where a single object (e.g., 🔪 or 🏺) can silently flip a home robot from normal behavior into harmful multi-step actions. 🚨Check out our work!

Had a great time at the UIUC NLP large group on “Sparse Autoencoders: Discoveries, Limitations, and the Bridge between Experiment and Theory” Thanks to Prof. Heng Ji and Jeonghwan Kim for kindly hosting, and to everyone for the engaging discussion! 🔗arxiv.org/pdf/2503.24277