In briefWaveletLoRA: Frequency-aware separation of content and style for personalized image generation
SummaryThis paper addresses separating a subject's content from its visual style using a single image, a key problem in personalized image generation. It proposes WaveletLoRA, which applies a multi-level discrete wavelet transform to split the reference image into low- and high-frequency subbands, each handled by its own LoRA branch: low frequencies capture style, high frequencies preserve content. A dense Mixture-of-Experts module and a frequency-domain regularization objective are added to reduce style leakage and improve content fidelity. The authors also introduce WaveBench and a VLM-based evaluation protocol, and report that the method consistently outperforms existing approaches and better matches human preference.
In briefLearning-based filling of detector gaps in photon-counting X-ray imaging
SummaryThis paper addresses gaps between tiled detector modules in X-ray images, which must be filled before phase retrieval or tomographic reconstruction. The authors adapt WGAN-GP and a conditional Pix2Pix model, restricting changes to gap pixels and adding losses that preserve mean intensity and intensity distribution. Both models achieved the best LPIPS and histogram distance scores, and Pix2Pix needed about five times fewer training GPU-hours than WGAN-GP.
Johannes B. Thalhammer, Patrick Ilg, Tina Dorosti et al.
In briefCombining contrastive and bootstrap self-supervised learning for classifying farm images with few labels
SummaryThis paper proposes CBRL, a method that jointly applies contrastive learning (SimCLR-style) and bidirectional BYOL-style target prediction within one representation path, aimed at agricultural image classification with scarce labels. Using a ViT-L/16 backbone and 10% of downstream labels, it reports single-seed point estimates on four primary agricultural benchmarks and an exploratory 33-image evaluation. CBRL achieves the highest macro-F1 point estimates on the four primary benchmarks, with notable gains on DeepWeeds and PlantVillage, while all methods fail to reach practical performance on PlantDoc. The authors note that the results do not establish statistical superiority or robustness across stochastic runs.
In briefGLCR: A Routing Method for Multi-Task Electromagnetic Signal Processing
SummaryThis paper proposes GLCR, a two-stage method that routes unknown electromagnetic signals to suitable expert models within a heterogeneous pool. A lightweight 1D CNN first performs global dispatch, and a gradient-boosted-tree specialist then re-evaluates confusable task pairs. The authors also introduce HeteroSignal-7, a benchmark of 140,000 samples across six task types. GLCR reports 96.16% routing accuracy and a 76.30% system score, with expected system FLOPs reduced 5.9 times compared with executing all experts.
In briefFADO: Learning Causal Explanations with Prior-Fitted Networks
SummaryThis paper addresses explaining machine learning models on tabular data, noting that common methods such as SHAP and permutation importance measure predictive correlation rather than causation. The authors define a feature intervention effect and propose FADO, an amortized method using prior-fitted networks trained on synthetic tabular data from structural causal models to predict this effect from observational data. Experiments reported in the paper indicate better explanations than standard predictive attribution in settings with proxy variables, confounders, colliders and placebos.
In briefDo Scaling Laws for Universal Machine-Learned Interatomic Potentials Hold Generally?
SummaryThis paper studies how the accuracy of three machine-learned interatomic potentials (M3GNet, MACE, PET) scales with training data size under fixed model configurations. It compares random sampling, several chemistry- and structure-based stratification policies, iterative DIRECT, and furthest-point sampling. Errors improve with a power law, but the fitted exponent, error floor, and sample efficiency depend on model, target quantity, sampling policy, and test distribution. The authors conclude that structure count alone is necessary but not sufficient for scaling.
Theo Keane, Anthony Onwuli, Richard Tomsett et al.
In briefCoCode-World: A Shared 3D Proxy for Consistent Multiplayer Video Worlds
SummaryThis paper presents CoCode-World, a framework for generating videos of two players in a shared game world. Programmable 3D proxies control scene layout, motion, and occlusion, while RGB references set character and scene appearance. The proxy videos, appearance references, and target views are arranged together and generated jointly by a pretrained video model. Experiments on a two-player GTA V dataset show appearance transfer to unseen characters and styles, with gains in video quality and cross-view player identity over the evaluated baseline.
In briefCat Royale Comprehensive Dataset for pose detection, segmentation, animal action recognition and scene understanding
SummaryThis paper presents the Cat Royale Comprehensive Dataset (CRCD), a large multi-modal dataset for feline understanding built from over 500 hours of video captured during the Cat Royale artwork and research project. It includes 20,000 images with instance segmentation masks, 5,000 images with multi-subject pose coordinates, over 800 video clips with action classes, and over 800 clips with detailed text descriptions. The authors provide benchmarks for instance segmentation, pose detection, temporal action localisation, and video-text retrieval.
Zane K. J. Hartley, Joseph Ho, Alan Chamberlain et al.
In briefMemCredit: choosing agent memories by decision utility, not just relevance
SummaryThis paper introduces MemCredit, a memory framework for long-horizon web agents that selects memories by their decision utility rather than semantic relevance. It estimates each memory's marginal contribution through controlled counterfactual interventions and distills these into a lightweight utility estimator used under a fixed token budget. On ALFWorld, MemCredit reports higher SR@50 than BM25 retrieval under a 500-token historical-memory budget. The authors argue that decision utility is a more effective principle than relevance alone for agent memory management.
In briefFilling gaps in near-surface air temperature records with a graph neural network that accounts for distance
SummaryThis paper presents a graph attention network that reconstructs near-surface air temperature fields from incomplete observations, treating each grid cell as a node linked to its nearest neighbors. The model is trained on CMIP6 CESM2 historical simulations masked with HadCRUT5 observational coverage from 1850 to 2014. It reports a spatial RMSE of 0.997 K, lower than LaMa, ordinary kriging, and GraphEM. An ablation study suggests the performance comes mainly from the geometric graph construction rather than auxiliary physical covariates.
Emma C. Weber, Dominique Guillot, Daniyar Abykhanov et al.
In briefWhen Local Definitions Are Missing: Errors That Language Models Repeat
SummaryThis paper studies how language models go wrong when a familiar word is given a special meaning locally in a text, but the model reads it with the usual meaning. The authors build multiple-choice questions from a synthetic vocabulary, commercial contracts, Reddit community glossaries, and decentralized exchange documentation, and compare answers with and without the definition. Redefined familiar words cause the most errors, and models often repeat the same wrong answer, while their repeated answers and stated confidence do not reliably reveal these errors.
In briefFinding Software Bugs with an AI-Driven Symbolic Execution Framework
SummaryThis paper presents an agentic framework that combines the CrossHair symbolic execution tool with an AI coding agent to find bugs in Python code without human intervention. The system adds assertions to a repository, synthesizes a constraint-aware symbolic environment for the database, and builds a concrete test for each reported bug. Evaluated on 11 open-source Python web applications totaling about 1.16 million lines, it found 1,050 bugs, and the authors report that symbolic execution and LLM-based code inspection find different kinds of bugs and complement each other.
George Granberry, Wolfgang Ahrendt, Andy Gordon et al.
In briefSmall output heads, strong decisions: a controlled comparison of typed-decision models
SummaryThis paper compares small output-head designs for models that make structured judgments about text, such as yes/no checks, choices among labelled options, and graded scores. The authors re-implemented five readouts and trained them under one shared backbone, training data and recipe. They found that head complexity has little effect on accuracy, while backbone size matters more, and they recommend investing in the backbone and keeping the head minimal.
In briefEMS-Nav: A multi-scale navigation framework for embodied robots
SummaryThis paper presents EMS-Nav, a biologically inspired multi-scale navigation framework designed to run online on a robot's onboard device. It organizes perception, spatial understanding, and navigation around regions, grouping objects and their relationships into functional spaces within a hierarchy from local areas to the wider environment. In experiments across four target types (objects, rooms, regions, and specific instances), EMS-Nav achieved the highest success rates, improving first-attempt success by 9.39 percentage points over the strongest baseline while reducing mapping overhead and language-model input size.
In briefVWEN: A Curated Small Language Model for Vietnamese Paraphrasing
SummaryThis paper presents VWEN, a 0.8-billion-parameter model that paraphrases Vietnamese text at a requested level of change and handles related tasks such as generating multiple paraphrases, simplification, gist extraction, register transfer and paraphrase judgement. It is built by fine-tuning Qwen3.5-0.8B on about 7k samples chosen through a four-stage pipeline. On high-divergence paraphrasing, VWEN reaches a success rate of 0.61, compared with 0.46 for the ViSP model at matched compute.
In briefCoupled noise method for consistent long-horizon diffusion planning
SummaryThis paper addresses a problem in compositional diffusion planners, where separately generated trajectory chunks can be individually feasible but incoherent when stitched together in multimodal environments. The proposed training-free method, MCDC, correlates initial noise between neighboring chunks while keeping each chunk's marginal distribution exactly standard Gaussian. On OGBench long-horizon stitch tasks, the authors report consistent gains over independent sampling with nearly unchanged inference time, and a better success-cost trade-off than search-based composition methods.
In briefAI4Plasma II: Geometry-aware surrogate models for plasma with variable electrode gaps
SummaryThis paper compares six geometry-aware surrogate models for predicting plasma fields and ion density near the wafer when electrode spacing varies. The models use prescribed or learned geometry mappings, extended-domain representations, continuous coordinates, or geometry-native discretizations. Each reference simulation averages 5.32 CPU hours. The authors report that prescribed-mapping Geo-FNO and Transolver reach 2% mean error with 100 training cases, and that adding charge-separation weighting reduces Geo-FNO's requirement for 5% profile error from 1000 to 100 cases.
In briefFoveaDet: Hardware-Software Co-Design of a Foveated CNN-Transformer for Object Detection
SummaryThe paper proposes FoveaDet, a framework inspired by biological foveation that predicts fixation points from visual saliency and prunes irrelevant background patches in Vision Transformers. On standard GPUs it reduces floating-point operations by 65.4% to 92.2% and speeds up inference by 4.76x with a negligible mAP drop. The authors also present a custom RISC-V core with specialized instructions to accelerate the foveated processing pipeline.
In briefAuditing EHR Foundation Models for Patient Memorisation
SummaryThis paper tests whether autoregressive electronic health record foundation models trained on MIMIC-IV disclose training patients' information at inference. It evaluates token reconstruction, sensitive-attribute disclosure, and random patient-level canaries across several checkpoints. The authors find weak cohort-level memorisation but no reliable patient-level disclosure under the attacks tested, and show the audit detects strong memorisation in an overfit positive control.
Inigo Stibbe, Joshua Placidi, Padmanabhan Ramnarayan et al.
