In briefFault-tolerant fast sliding control for drones with a disturbance observer that uses tracking error
SummaryThis paper studies fault-tolerant tracking control for quadrotor drones under external disturbances, parameter uncertainties, and actuator faults. The method splits the system into position and attitude loops and combines an appointed-time prescribed performance function, a tracking error-based fixed-time composite disturbance observer, and a fast second-order terminal sliding mode controller. Simulation results show better tracking accuracy, faster fault recovery, and stronger robustness than conventional methods.
In briefTemperature, Not Knowledge: How Self-Training Changes Model Confidence
SummaryThis paper examines how a language model's confidence changes over five rounds of self-training, using Qwen3-1.7B with three seeds per method. Accuracy and overall discrimination stay nearly unchanged, but confidence becomes sharper and the most confident answers become less reliable. The authors attribute most of this drift to fitting each generated answer twice, and compare several remedies that affect temperature and discrimination differently.
In briefEfficient conformal calibration for paired recommendation decisions, a method to restore useful changes to recommendation rules
SummaryThe paper addresses risk control that suppresses useful changes to a frozen recommendation rule. It proposes Averaged Conformal Risk Calibration (ACRC), which averages losses before clipping to recover valid activation from the same observed fractional losses. Across six datasets, changing only the calibration improves discounted net utility over the raw and RB baselines, and a separate Jester confirmation and comparison against SCoRE also show gains.
In briefLand Units Built Adaptively from Earth Satellite Embeddings
SummaryThe paper clusters AlphaEarth Earth embeddings with parallel Bayesian hierarchical clustering to form adaptive land units that are large where land is uniform and small where it is heterogeneous, without using labels. Applied to Denmark, it reduces 402 million pixels to 1.27 million units, about 320 times fewer. Land-use models trained on these units stay within a few accuracy points of pixel-based models at a small fraction of the cost, and enable whole-country estimates in under a minute.
Albert Alonso, Jamal Jokar Arsanjani, Andres R. Masegosa
In briefInfinite Worlds from One Example: Sharing Randomness in Local Generators
SummaryThe paper studies local generators that create unbounded game worlds on demand, with every small window required to come from a single example, as in Wave Function Collapse. It shows that a generator applying one position-independent rule to independent randomness at each cell cannot be both exactly valid and faithful to the example's statistics when constraints vary across the example. It gives a finitely checkable condition for exact validity using seed-based row types and position modulo a period, and tests it on 52 public Wave Function Collapse examples and 27 from other implementations.
WhyThe authors show local generators need shared randomness, prove independent randomness cannot be both exactly valid and faithful, and give a matching 1/r lower bound.
In briefHow Close to the Theme? Controlling Similarity in Game Music Generation
SummaryThis paper trains a pretrained MIDI model on pairs of sections from Super Mario pieces, with a label giving each pair's similarity and a context vector summarising the given section. For 157 test contexts, the model generates six outputs per label, scored by twelve measures. Outputs requested as similar are closer to their context than those requested as dissimilar, especially in melody.
In briefEntropy-regularized reinforcement learning for a regime-switching consumption-investment problem
SummaryThis paper develops an entropy-regularized reinforcement learning framework for continuous-time consumption and investment decisions in a regime-switching market with general utility functions. It derives closed-form optimal exploratory strategies, showing the investment strategy is multivariate Gaussian and the consumption strategy is a Gibbs measure, and that classical deterministic decisions correspond to the mode of these distributions. The paper also proves convergence of the exploratory value function to the classical one as the temperature parameter vanishes, and proposes a continuous-time q-learning algorithm supported by numerical experiments.
In briefA paper arguing agents need both exact code rules and learned fuzzy rules to stay safe
SummaryThe paper argues that dependable AI agents need both hard rules written as code, for limits that can be stated exactly, and learned fuzzy rules, for judgments that cannot. Tested on 1,340 episodes across purchasing, delegation of authority and data access, models acting alone executed unauthorized actions, while a hard guard placed before actions executed none. A small model taught judgment rules with Concept Programming flagged most rule-breaking requests that exact rules cannot detect.
In briefRetrieving Constraint Dependency Graphs to Detect Compliance in Long SKILL Documents
SummaryLLM agents follow long, cross-referenced SKILL documents, which are standard operating procedures, but these exceed the context budget of small auditors. SkillCDG converts each document into a constraint dependency graph and selects rules by dependency closure rather than similarity. Experiments on three enterprise datasets and two public benchmarks report higher accuracy and F1 than full-document and retrieval-based baselines.
In briefParaBlock++: A parallel federated learning method combining communication and computation via satellites.
SummaryThis paper proposes ParaBlock++, a hierarchical federated learning framework assisted by low Earth orbit satellites, letting resource-limited ground clients train models collaboratively over ground-satellite links. Satellites act as edge servers, computing clients, and inter-satellite relays. It uses asynchronous BCD-type updates instead of synchronous Jacobian-type ones and provides convergence guarantees for convex, strongly convex, and non-convex objectives. Experiments on ResNet-18 report faster initial convergence, comparable final performance, and fewer global rounds and communication costs.
In briefpico-WAM: Testing Whether Prediction or Trajectory Scoring Limits a Compact Robot Control Model
SummaryThis paper introduces pico-WAM, a compact World Action Model with supplied physical state, and studies why it fails at control. Using 206 Push-T demonstrations and 15 training seeds, the authors compare data regimes and search settings. Counterfactual branch training improves short-horizon cost ranking and widens the coverage advantage of policy-restarted search by 4.8 percentage points. Replacing learned trajectory scores with simulator costs improves coverage, while substituting simulator states in proposal generation alone shows no detected gain.
In briefSelf-normalized inference for constrained parameters using Riemannian stochastic gradient paths
SummaryThe paper studies building confidence regions for parameters estimated by Riemannian stochastic gradient paths on constrained manifolds. It represents retained iterates in a common random tangent space, so an unknown covariance factor cancels in a Brownian bridge pivot. Under regularity conditions, a joint expansion of the averaged estimator and integrated path controls nonlinear coordinate errors. Numerical studies report coverage competitive with plug-in Wald inference, without estimating the Hessian or residual density.
In briefPredicting Visitor Numbers Using Open AI Camera Counts and Weather Forecasts
SummaryThis paper examines whether small heritage and nature sites can forecast visitor numbers using only one AI camera counter per site, a public holiday calendar and a free weather forecast. Using about 22 months of open camera counts from three sites in Fukui, Japan, the authors measure how rain affects visits and predict daily counts up to a week ahead. They find that more exposed sites are more sensitive to weather, and that weather forecasts improve three-day-ahead predictions at a summit park.
In briefSatellite-assisted asynchronous federated learning across space, air, and ground networks
SummaryThis paper proposes HAFL, a hierarchical asynchronous federated learning framework that uses Low Earth Orbit satellites to let resource-limited ground clients collaborate over ground-satellite links. The satellites act as edge servers, training accelerators, and inter-satellite relays. HAFL uses asynchronous Gauss-Seidel-type updates and provides convergence analysis for convex, strongly convex, and nonconvex objectives. Evaluations show faster initial convergence, comparable final performance, and fewer global rounds and communication costs.
In briefJet-QSparse: A training-free method that speeds up AI agent inference using 4-bit sparse attention.
SummaryThe paper presents Jet-QSparse, a training-free method that stores the attention key-value cache in compact NVFP4 format and computes query-key products natively in FP4 on Blackwell tensor cores. It scans the full key history and selects 128-token blocks for sparse FP16 value aggregation with FP32 accumulation. Across four hybrid LLMs, using 128 blocks per KV head showed no detected accuracy loss on the evaluated math reasoning, coding, and agentic benchmarks.
In briefRobust pixel-wise digital image correlation using a global-to-local hybrid neural field under large motion and lighting changes
SummaryThis paper presents a neural framework for digital image correlation (DIC), a method for measuring full-field displacement without contact. A first stage estimates the dominant rigid-body motion through rotation search, phase correlation, and masked refinement to provide a reliable starting point. A second stage models the remaining displacement with a hybrid Kolmogorov-Arnold network and multilayer perceptron, optimized with an illumination-robust structural loss. Simulated and real experiments report stable displacement recovery under noise, non-uniform deformation, large rotation, and illumination variation.
