🚨new method day! 🚨 Trust[ing results] in ML conferences is utterly broken. Let's fix it with an algorithm! We are excited to introduce 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗪𝗶𝘁𝗻𝗲𝘀𝘀𝗲𝘀, a system for “honesty” certificates of data and evaluations which a verifier can cheaply check and trust.

📚 Saturday Robotics @saturdayrobotic & World Models Reading Club #33: Perceiving, Predicting, and Planning for Physical Interaction + AdaJEPA San Francisco · October 10 · 2–5 PM 👉 RSVP: luma.com/y9omqzz0 What does a robot actually need to understand before it can reliably act in the physical world? For Reading Club #33, we’re bringing together two perspectives on robot world models: grounding intelligence in physical interaction and building latent spaces that make planning and adaptation work. 🎤 @Hongyu_Lii — Robotics Researcher, @NVIDIARobotics, @nvidia Perceiving, Predicting, and Planning for Physical Interaction Robots manipulating the real world need to do more than recognize objects. They need to perceive contact, predict physical change, and plan under uncertainty. Hongyu will present a unified line of work spanning: • NovaFlow — zero-shot manipulation by distilling actionable 3D flow from generated video • NovaPlan — closing the loop on long-horizon tasks through video-language planning • Deform360 — large-scale multi-view visuotactile data for deformable world models • Hydra-0 — a generalist world model conditioned on action flow, representing robot actions as pixel motion The bigger question: can a shared visual interface connect touch, vision, actions, world models, policy evaluation, and control across different robots, tasks, and environments? 🎤 @yingwww_ — @NYUDataScience AdaJEPA: Adaptive World Models for Planning Under Distribution Shift JEPA-style world models predict the consequences of actions in a learned representation space—but what makes that latent space actually useful for planning? Ying will discuss temporal straightening, which encourages locally straight latent trajectories so Euclidean distance becomes a better proxy for geodesic distance and makes gradient-based planning more stable. She will also introduce AdaJEPA, an adaptive world model that plans, acts, and adapts in a closed loop: every action produces a new observation, which updates the latent representation and prediction when reality diverges from the model. Two complementary questions: 🧠 What should a robot represent about the physical world? 🔄 What should it do when its world model is wrong? 📍 San Francisco 🗓 Saturday, October 10, 2026 | 2–5 PM 2:00–2:30 — Doors Open & Social 🍓🧋 2:30–3:30 — Keynote 3:30–5:00 — Q&A + Open-Floor Technical Roundtable Hosted by @junfanzhu98 & @aurorafeng_01 & Audrey Pe (@sievedata). Come discuss robot world models, visuotactile perception, deformable objects, action-conditioned video models, latent planning, distribution shift, test-time adaptation, and what it actually takes to build robots that can perceive → predict → plan → act → adapt. 👉 RSVP: luma.com/y9omqzz0 #Robotics #WorldModels #EmbodiedAI #PhysicalAI #RobotLearning #JEPA #RobotManipulation #NVIDIA #NYU

Congratulations to CDS alum Qihan Wang (@wnfurfur), former CDS Faculty Fellow @NickATomlin, CDS PhD student @michahu8, UMass Amherst's Brian Dillon, and CDS Associate Professor @tallinzen: their paper on getting AI to forget the way people do has been accepted to #NeurIPS2026.

To make AI better at imitating humans, limit its memory. Recent CDS Faculty Fellow @NickATomlin, CDS PhD student @michahu8, and CDS Assoc Prof @tallinzen show that capping a model's memory at 4 slots makes it a far better stand-in for a real person. nyudatascience.medium.com/fo…