In briefA SegFormer-based model for predicting daylight factor distribution in BIM
SummaryThis paper presents a machine learning method that predicts the spatial Daylight Factor (DF) of residential rooms directly within BIM software, returning results in seconds without model export. It was trained on about 20,000 synthetic rooms and validated on a held-out synthetic set and 102 rooms from 18 real buildings. The default-materials version showed low error for Swedish BBR and Norwegian TEK17 compliance metrics, while the custom-materials variant had higher error. The authors recommend the default version for preliminary daylight checks and state the custom version needs further development.
Stanislava Fedorova, Alejandro Pacheco Di\'eguez, Christoph Berkmiller et al.
In briefHow geometric structure affects evaluating generative models on curved spaces
SummaryThis paper studies how curvature and dimension of Riemannian manifolds affect distances and evaluation results when assessing generative models. It systematically compares geodesic, ambient, and tangent-space versions of Wasserstein and MMD using controlled-noise perturbations and Riemannian Flow Matching samples. The authors find that no single metric is uniformly reliable across geometries and offer recommendations on reporting evaluation distances, comparing metric types, and calibrating metrics to sample geometry.
Misha Oberski, Valentijn Oldenburg, Xanthos Kekkos et al.
In briefLearning compact sets of representative ports for efficient fluid antenna systems
SummaryFluid antenna systems offer many candidate ports, but checking all of them for every channel instance is costly online. This paper proposes an offline method that compresses the candidate-port space once and reuses the result for future channel realizations, collecting optimal ports from sampled training realizations. On a DeepMIMO 7-by-7 system with 49 candidate ports, 120 training realizations yielded about 27 representative ports on average. These retained 99.995% of the channel power gain achieved by exhaustive port selection.
In briefStrategic Calibration in LLM Multi-Agent Systems: Truthful Confidence Reporting
SummaryIn LLM-based multi-agent systems, reported confidence determines how scarce downstream attention and computation are allocated, so agents may strategically overstate their confidence. The paper proposes the Audited Pivot Scoring Mechanism (APSM), which combines allocation-benefit offsetting, inverse-propensity-weighted sparse audits, a Clarke pivot externality charge, and a strictly proper score. Under the stated assumptions, truthful reporting is the unique expected-utility best response. Experiments show LLM agents inflate confidence under Top-K incentives, while APSM reduces distortion, enlarges the utility margin against allocation-changing deviations, and improves robustness under approximate execution.
In briefHuman-feedback-guided multi-agent inverse reinforcement learning for modeling interactive driving behavior
SummaryThe paper proposes HFG-MAIRL, a framework that uses multi-agent inverse reinforcement learning to infer the latent objectives behind driver behavior in mixed traffic. Each vehicle and its directly interacting neighbors are modeled as a localized Markov game, with finite marginal-response iterations approximating coupled responses without enumerating the full joint action space. Human feedback is incorporated as ordinal preference supervision via a Bradley-Terry likelihood, with driver emotion labels as one instantiation. Experiments on the AIDE dataset report that HFG-MAIRL outperforms representative IRL baselines in decision modeling.
In briefReconstructing fine-scale PDE trajectories from coarse data via latent dynamics
SummaryThis paper introduces Structured Latent Dynamics (SLD), a framework that reconstructs high-resolution trajectories from spatially and temporally downsampled PDE simulations. It combines temporal interpolation and learned corrections with a latent state updated by gated responses, spatial differences, and a Laplacian term, then decodes via Haar wavelet synthesis. Tested on Gray–Scott, complex Ginzburg–Landau, and viscous Burgers systems, SLD reconstructs 100 frames from 25 inputs and reports lower error than unstructured latent dynamics and ConvLSTM baselines.
In briefRoboJev: An Open Vision-Language Model for Judging Robot Task Success
SummaryThe paper presents RoboJev-27B, an open-weight vision-language model that takes a task instruction and frames from a robot manipulation attempt to predict whether the task succeeded. It is trained from JEV-27B-VL with a LoRA adapter on 52,598 episodes and reaches 95.2% balanced accuracy and 0.993 AUROC on 268 held-out episodes from the same robot. The authors also report generalization to episodes from unseen labs and robots, and release the inference code and prediction results.
