In briefDecentralized predictive control for swarms of tiny vibration robots in a confined arena
SummaryThe paper presents a distributed predictive flocking controller in which each robot solves its own nonlinear model predictive control problem and shares only predicted trajectories with nearby robots. Six barrier and control-barrier-function safety formulations were compared on a digital twin of an open-source vibration-driven bristlebot in a 0.64 square meter arena. A look-ahead discrete-time control barrier function performed best overall, and the controller was tested on phototaxis, phototactic orbiting, and cohesion tasks from ten robots up to the geometric packing limit. Collisions stayed low in small swarms and rose steeply near the packing limit, while coordination and cohesion degraded gradually.
Prajol Shrestha, Siddhant Mohapatra, Filip Novkoski et al.
In briefWindow likelihood training for yielding and holding turns in full-duplex speech models
SummaryFull-duplex spoken dialogue models must yield the floor when a user interrupts and keep talking when a user only backchannels. The paper finds that whole-rollout GRPO with a timing reward fails in two ways: a binary window reward gives zero advantage on key windows, and a single per-rollout score is spread over decisions the reward does not describe. The authors propose window likelihood (WL), which integrates the stop decisions in closed form and maximises the log-probability of correct action. On HumDial-FDBench, WL gives the best yield/hold F1 on all six structurally different open models tested and adds no re-opening of the floor on any model.
In briefUnmixing hyperspectral images with spectral variability using a new additive mixing model
SummaryThis paper addresses spectral variability in hyperspectral images, caused by atmospheric effects, material differences, and illumination or topography changes, which standard linear mixing models ignore. The authors propose the Structured Additive Variability LMM (SAV-LMM), which adds two terms to the standard model to handle variability additively. They develop two Nonnegative Matrix Factorization-based unmixing algorithms and a new initialization scheme that accounts for variability. Experiments on synthetic and real datasets show the method outperforms tested state-of-the-art approaches on images with spectral variability.
Salah Eddine Brezini, Yannick Deville, Moussa Sofiane Karoui et al.
In briefInfinity Machine: A Self-Improving AI Agent for Research
SummaryThis paper presents the Infinity Machine, a self-improving AI agent used for research on problems whose answers can be checked by a program or data. Over six days on fourteen problems, the system reports four results beyond the published state of the art on open quantum information problems, and cut SABRE qubit-routing added gates from 68,559 to 12,087. The authors also document failures, including rejected harness changes and proposers gaming lenient scorers.
In briefExCal: Optimization-Free Multivariate Time Series Anomaly Detection via Excess Calibration
SummaryThis paper presents ExCal, a multivariate time series anomaly detector trained only on normal data. It combines raw state retrieval, signal and derivative subspaces, and robust statistics across temporal scales, without gradient optimization or anomaly labels. On 180 series from TSB-AD-M, it ranks second of 24 methods on AUC-PR and VUS-PR, with no significant difference from the two leading non-pretrained methods but significant gains over the other 21 baselines. Inference is batch and transductive, and online suitability remains untested.
Kadir-Kaan \"Ozer, Ren\'e Ebeling, Markus Enzweiler
In briefCopying as Defense: Stopping Image Edits by Making Models Copy
SummaryThis paper proposes Copying-as-Defense (CaD), a method that protects photos from being manipulated by diffusion transformer image editors such as Flux Kontext. It adds an imperceptible perturbation that pushes the model's attention toward the protected reference image in early blocks, causing it to reproduce the input instead of performing the requested edit. Across editors and benchmarks, edit success drops from 81.4% on unprotected images to 24.0%, and the method runs about 1.8 to 2.0 times faster per optimization step than comparable methods.
WhyProposes a novel copy-inducing defense via early attention interventions, substantially cutting edit success across multiple editors and benchmarks.
Salma Abdel Magid, Sanghyuk Chun, Eric Slyman et al.
In briefHigh-resolution touch sensing reveals hidden badminton racket control, making it learnable
SummaryThe study collected over 14,000 badminton strokes from more than 70 participants across five skills, using an 864-node tactile-inertial glove sampled at 100 Hz. A two-layer mechanics-based framework inferred object-driving and grip-stabilizing control from tactile data and recovered its spatiotemporal force realization. Higher proficiency was linked to selective rather than extensive control, and the authors used these findings to provide personalized guidance that improved learners' expert-rated skill levels more than level-only feedback in a randomized cohort.
In briefHow multi-fidelity training reshapes uncertainty in atomistic models
SummaryThis paper studies how multi-fidelity training, which pretrains on abundant low-fidelity data and fine-tunes on scarce high-fidelity labels, affects predictive uncertainty in atomistic models. The authors use deep ensembles of mean-variance MACE models trained on ANI-1x, comparing high-fidelity training from scratch with pretraining followed by fine-tuning. Pretraining reduces prediction error and improves uncertainty-based ranking, but yields poorly calibrated predictive variances, with calibration worse under energy extrapolation, and most degradation arises during high-fidelity optimization.
