Michigan AI seminar: continuous methods for discrete data MichiganAI (@michigan_AI) announced Ruqi Zhang's seminar, Continuous Methods for Discrete Spaces: From Sampling to Diffusion Language Models, on November 10, 2026, at 4:00 p.m. Eastern Time. The University of Michigan event page lists the talk as free and open to the public, with a Zoom link. The abstract covers gradient-informed sampling and approaches that connect discrete and continuous diffusion. For readers following language-model research, the useful question is how techniques developed for continuous variables can help model discrete tokens without losing the structure of the problem. Event details: https://cse.engin.umich.edu/event/ai-seminar-ruqi-zhang
AI · News and comments
Humanoid skills from proprioception at CoRL 2026 Aditya Bhatt (@aditya_bhatt) announced workshop presentations of Blind Dexterity and The Blind Skateboarder, describing humanoid skills based on proprioception—the robot's internal sensing of its own state. He reports that the work uses no vision, motion capture, force/torque sensors or tactile sensors. The Everything Beneath the Policy workshop lists Blind Dexterity as an oral spotlight on November 12, 2026, in Austin. Bhatt also points to the afternoon workshop on contact-rich locomotion and manipulation. Our research question: which task assumptions allow internal sensing to support these skills, and how far do the results transfer when objects or conditions change? The demonstrations should be assessed within their tested settings. Morning workshop: https://beneath-the-policy.github.io Afternoon workshop: https://contact-rich-loco-manipulation.github.io
Always-on personal agents need a complete feedback loop Shuai Liu (@choiszt) shared the preprint Always-On Personal Intelligence: A Roadmap Towards Continuing, Self-Evolving Loops, posted on November 9, 2026. The authors organize personal assistants around sensing, memory and action, and introduce HALO to assess how these parts work together. Their review identifies particularly weak support for memory guiding sensing and for action outcomes returning to memory. Our takeaway for agent builders: retaining information is only part of the problem. A useful evaluation should also ask whether past outcomes change what an assistant notices and does next, within the user's permissions. Preprint: https://doi.org/10.20944/preprints202611.0793.v1
FlexGlue: learning when images should not match Philipp Lindenberger (@PhilippCSE) introduced FlexGlue, a model that finds image correspondences while also rejecting similar-looking images of different structures. Gabriele Berton (@gabriberton) highlighted its use of negative training examples. The paper describes joint training for matching and rejection, with early stopping on non-matching pairs. For 3D reconstruction, this addresses a practical problem: confident false matches can connect places that do not belong together. Our takeaway is that evaluation should test when a model refuses a match, alongside how accurately it finds one. Paper: https://arxiv.org/pdf/2610.32976 Code: https://github.com/cvg/FlexGlue