In briefMythoSemantic: An Interactive Tool for Finding Meaning Links Between Myths and Religions
SummaryThis paper presents a tool for comparing mythological and religious texts using neural language models that map them into a shared semantic space. Users can explore semantic parallels across cultural traditions through an interactive interface. The authors report that the coarse structure of the embedded fragments aligns with well-documented historical relations between traditions, and they quantify how much of this structure is explained by lexical overlap and proper names.
Nikita Ganziuk, Innokentiy S. Humonen, Ilya Makarov et al.
In briefSharing Data Across Stores: Transfer Learning for Choice-Based Revenue Management
SummaryRetail chains need to tailor assortment and pricing to local customer preferences, but store-level data is often scarce, especially for short-selling-window products. The paper proposes a shrinkage estimator under a multinomial logit demand model that blends store-specific estimates with information pooled from other stores. It provides finite-sample guarantees on when transfer improves estimation and derives regret bounds for assortment and pricing decisions. Experiments and a study of several hundred stores of a European apparel retailer show shrinkage outperforms both local-only and fully pooled estimation.
Recep Yusuf Bekci, Stefanus Jasin, Joline Uichanco
In briefEvaOmni: A Benchmark for Evidence-Grounded Audio-Visual Reasoning
SummaryThis paper introduces EvaOmni, a benchmark of 392 videos and 1,027 human-annotated question-answer pairs across 15 capabilities and three difficulty levels. Models must return both the answer and the supporting audio and visual evidence intervals, scored by coverage and bottleneck metrics alongside correctness. Evaluation of 24 Omni LLMs shows the top model reaching 79.1% accuracy at the hardest level but only 12.3% on its bottleneck score, indicating room to improve grounded reasoning.
In briefA CNN classifier for subject-independent MBTI extraversion/introversion detection from short EEG segments
SummaryThis study tests whether short, lightly preprocessed 4-second EEG segments can distinguish MBTI extraverts from introverts. The authors compare EEGNet, EEGProgress, and a new compact network called SSCNet using subject-independent nested 5-fold cross-validation with 50 participants. SSCNet reports the highest mean accuracy, AUC, and F1-score, with the lowest variability across folds and good training efficiency.
In briefStratified Sparse Attention: Narrowing the Accuracy Gap of Sparse Attention for Long-Context LLMs
SummaryThis paper addresses the high cost of attending over long key-value caches in large language models. It proposes Stratified Sparse Attention (SSA), which deterministically attends to high-scoring tokens and samples the remaining tokens using proxy scores with inclusion-probability correction. Across several long-context benchmarks and five models from 3B to 27B parameters, SSA outperforms baselines at matched sparsity and achieves up to 2.96x decode speedup over FlashInfer at 1M context.
WhyUnifies deterministic retrieval with score-guided sampling and inclusion-probability correction, with consistent gains across several benchmarks and model sizes.
In briefDo neural models favor their submodels? A geometric study of whether loss minima lie in submodels.
SummaryThe paper studies the geometry of neural models, the sets of functions expressible by a given architecture, and asks whether local minima of the loss lie within their submodels, such as those with inactive neurons. It introduces a geometric framework and analyzes shallow and deep multilayer perceptrons as well as lightning self-attention. The authors find that, apart from a few exceptions, preference for submodels cannot be explained by model geometry and must arise from other sources such as training dynamics.
WhyGeometric framework proves shallow and some deep MLPs avoid submodels; novel and rigorous, with limited practical impact.
In briefEvery Model Leaves a Trace: Multi-Resolution LLM Fingerprinting (Paper)
SummaryThis paper studies whether the model-specific traces that large language models leave in their generated text can distinguish closely related models. The authors progressively suppress surface cues, paraphrase outputs, vary the amount of text, and test generalization across prompt sources and languages. They find that fingerprints remain detectable even in distilled models and LoRA adaptations of the same base model. Substantial model-specific signal persists as lexical cues weaken, and semantic representations are the most robust across languages.
Mohammad Reza Modarres, Armin Tourajmehr, Artin Tavasoli et al.
In briefSimHack: forecasting when detectable reward hacking starts, before RL training
SummaryThis paper studies when detectable reward hacking becomes dominant when training code-generating LLMs with reinforcement learning, so that lower-risk training recipes can be chosen in advance. The authors introduce SimHack, a simulator that learns training dynamics from a few smaller reference RL runs and simulates unseen target recipes without costly retraining. It predicts median onset with 11.3% relative error and 0.85 rank correlation across recipes, and screening curricula with it yielded runs 96.6% longer without hacking.
In briefA Plausible but Wrong Future: Steering World-Action Models Toward the Intended Outcome
SummaryThe paper proposes TaskWAM, a closed-loop, stage-aware framework that uses Mode Tokens to represent task requirements and State Tokens to track execution progress, steering world-action generation toward the intended future. It also introduces StageBench, a multi-stage manipulation benchmark with parameterized variations and automated evaluation. Experiments in simulation and real-world tasks report reduced Stage Confusion and performance competitive with state-of-the-art methods.
In briefRacer: Finding GPU Kernel Bugs via Symbolic Equivalence Checking
SummaryGPU kernel rewrites are usually verified by differential testing on random inputs, which can miss logical bugs whose errors fall within tolerance. This paper proposes symbolic equivalence checking, which compares a kernel and its reference as exact algebraic expressions over all inputs, and implements it as Racer, a PTX-level engine. Across 2,408 production kernels from 41 targets, Racer found 82 real bugs in 23 libraries, 76 of them previously unknown, with 41 confirmed so far.
WhyApplies symbolic equivalence checking to PTX-level GPU kernels, finding 82 real bugs across 2,408 production kernels with strong empirical evidence.
In briefBrain-region EEG classifier for hypnotic susceptibility, tested on unseen subjects
SummaryThis study trains a 2D convolutional neural network on EEG recordings from 62 male participants to classify hypnotic susceptibility across seven brain regions. Random-split tests favoured the occipital region, but leave-one-subject-out evaluation dropped its accuracy to near chance, indicating that the earlier results reflected individual EEG patterns. The frontal and prefrontal regions stayed above chance under subject-independent testing, and the authors identify them as more reliable markers.
In briefWorldSR: Grounded image super-resolution using external visual world knowledge search
SummaryThis paper proposes WorldSR, a framework that splits an input image into semantic entities and retrieves structurally and semantically aligned reference images for each from a large external corpus. A memory mechanism and multi-step filtering improve efficiency and reliability, and the selected references guide a multi-reference super-resolution model. Experiments on real-world benchmarks report better structural fidelity and visual quality than existing generative super-resolution methods.
In briefDISTRACT: Testing whether text guides video trackers when the image is degraded
SummaryThis paper introduces DISTRACT, a benchmark that compares four text descriptions against four image degradations while keeping the video, target, and distractor fixed. Five vision-language trackers were tested, and Molmo2 and JointNLT showed directed use of text, while UVLTrack, MMTrack, and DUTrack responded only weakly. Internal analyses found that text effects appear at earlier intermediate states, but matched-correct text did not significantly outperform unrelated text, so target-specific semantic rescue was not established.
In briefCode2Video-Compass: A Benchmark for Generating Videos Through Code Across Text, Image, and Video Inputs
SummaryThis paper introduces Code2Video-Compass, a benchmark for evaluating how well multimodal large language models generate videos by writing code. It covers text-to-video, image-to-video, and video-to-video settings, with 2,104 tasks across 10 categories and 24 subtypes. The authors evaluate 15 recent models using twelve objective metrics and five subjective metrics judged by an MLLM. They find that even leading models struggle to reliably produce high-quality videos that match requirements, and performance generally drops as task complexity increases.
In briefPhaseBreak: Proactive protection against unauthorized image-to-video animation
SummaryThe paper addresses the risk that image-to-video diffusion models can animate a shared photo of a person without authorization. The authors propose PhaseBreak, which lets an image owner add a bounded perturbation before release that keeps the image shareable but disrupts motion generation. It retimes intermediate denoiser features along the true temporal axis and uses a subject-aware mask to confine changes to the depicted subject. Tests on Stable Video Diffusion and DynamiCrafter show protected images produce larger short-term motion variation and lower temporal regularity, hindering stable motion customization.
In briefBrain2Vec: a self-supervised 3D foundation model for brain MRI scans
SummaryThe paper introduces Brain2Vec, a foundation model pre-trained on 74,425 T1-weighted brain MRI scans pooled from multiple studies. It can be used as a weight initialization for fine-tuning or as a frozen feature extractor. In brain age prediction, pretrained initialization improved final accuracy and convergence speed compared with random initialization. Its frozen embeddings outperformed anatomically defined volumetric features on four of five downstream classification tasks.
Spiros Maggioros, Guray Erus, Gareth Harman et al.
In briefA multi-agent system that automatically trains medical image models using evidence-based planning
SummaryThe paper introduces MedicalMLOmni, a multi-agent AutoML framework that trains models for diverse single- and multi-modal medical imaging tasks through a natural language interface. It retrieves published results to guide model selection and configuration, and runs parallel hyperparameter optimization where several agents explore adjustment strategies concurrently. It achieves the best score on 19 of 20 benchmarks, with PMC-VQA as the exception, where a vision-language model with broader pre-training performs better.
In briefReward Designs for Budget-Aware Agents That Adapt to Any Test-Time Limit
SummaryThis paper studies whether large language model agents can learn to allocate a limited inference budget (tokens, API calls, or environment steps) across multi-turn interactions. The authors compare eleven reward objectives and train Qwen3.5 4B and 9B models on AppWorld and BabyAI. Rewards that penalise overspending improve within-budget success and natural compliance, and these gains extend to unseen budgets.
Tom A. Lamb, Adam Golinski, Eleonora Gualdoni et al.
In briefInstruction tuning can weaken factual recall in language models
SummaryThis paper describes a solution to the AKBC LM-KBC 2026 Shared Task, scoring 0.6836 on the official leaderboard against a baseline of 0.2964. The task asks for all possible objects of subject-relation pairs using a language model of up to 32 billion parameters, without retrieval or fine-tuning. The base model prompted with text completion reached 0.6836, while the instruction-tuned version using a chat interface reached 0.5681, which the authors attribute to instruction tuning itself. The final method prompts the base model with training examples and keeps only repeated outputs from 64 to 256 generations, with two additional relation-specific methods..
In briefDIVE: A Benchmark for Real-Time Full-Duplex Interaction with Visual Engagement
SummaryThis paper introduces DIVE, a benchmark that tests whether multimodal AI models can interact in real time with continuous audio and video streams. It covers six tasks grouped into three scenarios: proactive intervention, continuous engagement, and interaction coordination, using 525 annotated audio/video samples. Experiments show that leading real-time models respond in under a second on average, but their window response rates stay at or below 53.77%. Adapting ongoing responses to newly arriving information remains especially challenging.
In briefStreamEval: A Benchmark for Evaluating Interactive Streaming Video Generation Models
SummaryThis paper introduces StreamEval, an open-source benchmark for interactive streaming video generation, with public data, evaluation code, protocols, and scoring rules. It covers five tasks with 300 scripted cases lasting 15 to 300 seconds, assessed by ten metrics for task execution, state consistency, visual fidelity, and audiovisual coordination. Results from 37 models show that high visual quality and identity retention do not necessarily align with reliable request execution or sustained state consistency.
In briefDecentralized multi-view, multi-zone tracking for large camera networks, explained
SummaryThis paper presents a decentralized system for tracking people across many cameras. Single-view detections are mapped to bird's-eye-view coordinates and fused by independent trackers, each handling a dynamic zone. A locking protocol and periodic track-merging, called global collate, unify trajectories across zones, with a delayed-output buffer that retroactively corrects history. In the evaluated configurations, the system keeps global ID consistency with minimal accuracy loss compared to a single-instance baseline.
In briefWhen Copying Is Hard: Copy-Constrained Decoding for Exact Span Reproduction
SummaryThe paper examines why LLMs often fail to reproduce a target span verbatim from long context, drifting within the span or stopping incorrectly. The authors build a benchmark across contracts, scientific literature, code, and random sequences, and propose CopyGen, a lightweight plugin for frozen LLMs that splits copying into deciding when to copy, how far, and when to stop. According to the paper, CopyGen improves average exact match by 35% relative without finetuning across three backbones and four tasks, adds 4% with SFT, and speeds up generation.
In briefSpecEvo: Low-cost scientific discovery by coordinating several LLMs in evolutionary search
SummarySpecEvo is a framework that uses lightweight language models to generate and refine candidate programs in parallel, while an advisor model distills successes, failures, and saturated directions into lasting guidance. At sparse checkpoints, a frontier model intervenes to resolve search stagnation. Across 176 tasks in four scientific and algorithmic discovery domains, the authors report matching or outperforming frontier-model baselines while reducing API cost by up to 80%.
In briefSelf-Credit: Credit Assignment for Long-Horizon Agent Reinforcement Learning
SummaryThis paper addresses the difficulty of credit assignment in long-horizon tasks where reinforcement learning rewards are sparse. Self-Credit has the policy model generate summaries that cite earlier turns containing key evidence, then rescales these self-generated credits by comparing cited actions across successful and failed rollouts for the same prompt. The rescaled weights are used to reweight advantages of the cited turns. The authors report gains over base models and strong agentic RL baselines on BrowseComp-Plus, GAIA, and xBench-DeepSearch.
In briefVisualEvolve: A Framework for Vision-Language Models to Self-Improve Visual Reasoning Skills
SummaryThis paper proposes VisualEvolve, a framework that lets vision-language models improve reusable external skills (visual operations and tool-use strategies) through self-evolution without updating model parameters. It uses a Select-Execute-Evolve-Evaluate cycle, selecting a historical skill-library snapshot by performance on a fixed validation set to limit regressions, and a coarse-to-fine evidence access hierarchy to diagnose capability gaps. Across four model backbones and six evaluation sets, it improves over the initial skill library in 23 of 24 settings and needs far fewer task trajectories than representative RL-based methods.
In briefRACE: A Method That Shrinks Attention Memory by Keeping the Most Important Rotary Position Dimensions
SummaryThis paper presents RACE, a post-training method that compresses the attention heads of a pretrained language model to reduce key-value cache memory and attention computation. RACE selects rotary position embedding (RoPE) pairs using their joint query-key activation energy, and folds the result into the model weights, so no runtime reconstruction or specialized kernels are needed. According to the paper, at 50% KV-cache retention and 256K context, RACE reduces total inference FLOPs by 47.8% and yields 1.79x higher throughput compared with an uncompressed model.
Irene Tenison, Abdulrahman Alabdulkareem, Daniel Karl I. Weidele et al.
In briefRACE: Training Image-Based Search Agents with Referential Abstraction and Compositional Evidence
SummaryThis paper introduces RACE, a framework that synthesizes training tasks requiring visually grounded search, reducing shortcuts in synthetic questions. It replaces identity-revealing text with visual or relational references and links facts about multiple entities into joint questions. Fine-tuning Qwen3.5-27B on 12K RACE-generated trajectories raised average accuracy across ten multimodal benchmarks from 53.5 to 56.0.
In briefCAPSafe: A Benchmark Testing Whether Embodied AI Agents Reason Safely About Actions and Plans
SummaryThis paper introduces CAPSafe, a benchmark for evaluating safety reasoning in embodied agents powered by large language models. It tests whether single actions are safe in their environment and whether agents can produce safe long-horizon plans for hazardous household tasks. Evaluating nine LLMs, the authors find that hazard recognition is accurate (over 90% at best), but safe long-horizon planning is much harder, with Gemini-3-pro reaching the highest safe success rate of 78.19% and other models below 65%.
Youmna Farag, Svetlana Stoyanchev, Simon Keizer et al.
In briefObject-Oriented Harness: Evolving Agent Harnesses via Object-Oriented Design
SummaryThis paper introduces OO-Harness, a framework that optimizes agent harnesses, the executable code defining prompts, tools, and control flow around a language model, using object-oriented design. During training, an evolution agent builds specialized harness subclasses from execution records, and at test time a router selects a suitable harness per task using lessons learned in training. Across three long-horizon benchmarks, the method achieves the best performance among compared approaches, with an absolute gain of up to 15% over the base, and ablation studies examine its core components.
Gaotang Li, Bhavana Dalvi Mishra, Yifei Ming et al.
In briefDon't Skip the Baseline: A Fair Test of Dynamic Expert Skipping in MoE Models
SummaryThis paper compares dynamic expert-skipping methods for Mixture-of-Experts models, which reduce the experts activated per token during inference. The authors group four prior methods into one taxonomy and test them against a simple baseline that statically lowers the number of experts k at matched compute, on Gemma4-26B-A4B, Qwen3-30B-A3B, and Qwen3-VL. The simple baseline was hard to beat consistently, and the tested methods often underperformed it, with gains being minor when present. The authors recommend calibration-free, per-modality static k reduction as the baseline that dynamic methods must surpass.
Jakob Steimle, Thomas Elsken, Daniel Ehrhardt et al.
In briefRelative representations allow a monitoring probe to move across language models
SummaryThis paper examines whether independently trained language models organize behaviorally relevant information in compatible internal representations. The authors express each example through angular distances to shared anchors and build a global anchored probe that summarizes all transformer layers, so a frozen probe can be applied to another model without labeled target data. Across trivia, safety monitoring, and math reasoning tasks, transferred probes preserve on average at least 85% of same-model AUROC above chance. Transfer depends on the task and target and can fail even when the property is decodable in the target model.
In briefFALCA: Falsification-aware reinforcement learning for long-horizon deep search agents
SummaryThis paper addresses how LLM agents in long-horizon deep search can amplify early mistakes, such as anchoring on wrong candidates and overlooking contradictions. It proposes FALCA, which frames search as a propose-falsify process using a Searcher that suggests answers and independent Falsifiers that seek counterevidence and report conflicts. Training combines SFT warmup on falsification trajectories with agentic reinforcement learning using parallel rollouts and hierarchical credit assignment. The authors report improved performance over backbone and harness ablations on five search benchmarks, with reduced confirmation bias and error accumulation.
In briefMasked MDPs: Learning Unknown Action Constraints While Finding an Optimal Policy
SummaryThe paper studies reinforcement learning when the set of permissible actions, or action mask, is unknown in advance, under a linear MDP where each attempted action reveals an exact linear validity score. The authors first give a deterministic optimistic optimization algorithm that is minimax optimal and computationally efficient, but show it is suboptimal on non-degenerate instances. They then propose a computationally efficient alternative that plans its exploration, interpreted as an analogue of information-directed sampling for deterministic feedback. The framework is finally applied to transfer learning where the invalid action set changes over time.
Sebastian Reboul, Lorenzo Croissant, H\'el\`ene Halconruy
In briefCountVideoBench: A benchmark testing prediction biases in video-language models' object counting
SummaryExisting video question-answering benchmarks have counting questions whose ground-truth counts skew toward small numbers, which hides model biases on rarer counts. The authors introduce CountVideoBench, with 2,600 synthetic videos covering 26 cardinalities from 5 to 50, each containing a single class of static objects. Evaluating sixteen video-language models zero-shot, no model exceeds 27.6% Exact Match, and models concentrate predictions on preferred integers, peaking there and collapsing at adjacent counts.
Marc Chapus, Carlos Crispim-Junior, V\'eronique Eglin et al.
In briefBeyond the Point Mass: Acceptance-Derived Weights and Distributional Losses for Speculative Decoding
SummaryThis paper studies how to train small draft models for speculative decoding, where proposed tokens are verified by a larger target model. The authors use the overlap between draft and target distributions as the acceptance signal and derive position weights and a loss (D-PAL) from the expected accepted length. Replacing cross-entropy with Rényi divergence of order 1/2 raises mean acceptance length and throughput over D-PACE on six benchmarks with a Qwen3-4B target.
In briefS-Agent 2.0: Spatial Reasoning with a General Coding Agent
SummaryThis paper presents S-Agent 2.0, which equips an off-the-shelf general coding agent with a command-line 3D perception toolkit. The design helps a vision-language model gather scene evidence, store it in a persistent scene memory, and interpret each reading in plain spatial language. On MMSI-Bench and ReVSI, the toolkit improves over the same agent without tools and outperforms the specialized system SpatialClaw on the same backbone.
In briefRayCo4D: A Ray-Conditioned Continuous Motion Model for Reconstructing Moving 3D Scenes from Single-Camera Video
SummaryRayCo4D is a method for reconstructing dynamic 3D scenes from monocular video by separating visual measurements from latent 3D motion states. It links fragmented long-range tracks to shared continuous motion nodes and transfers the inferred motion to Gaussian primitives, with local residuals capturing fine non-rigid deformation. Experiments on diverse monocular dynamic scenes report competitive novel-view synthesis quality, better structural and temporal consistency under rapid motion and incomplete tracks, and support for continuous-time rendering.
In briefGated prototype transfer for medical vision-language models without labeled support
SummaryThe paper addresses few-shot adaptation of medical vision-language models, where some low-prevalence classes have no labeled support examples. It introduces GProtoT, which transfers visual residuals from supported classes to missing classes using VLM similarity and an ontology-based clinical distance, with an evidence gate limiting unsupported transfer. In 50 paired one-shot trials on MESSIDOR, gated transfer improved missing-class recall and average class accuracy over SS-Text+ and greatly reduced negative-transfer frequency, at a small cost to supported-class recall and top-1 accuracy.
Sumaiya Fazal, Rizwan Qureshi, Laleh Behjat et al.
In briefCharacterizing Cross-Relay Generalization Failure in Machine Learning-Based Power System Protection
SummaryThis paper studies whether machine learning fault detectors trained at one relay location still work at other relay locations in a power network. Using simulated 110 kV transmission data, the authors test a multilayer perceptron across eight relays and all source-target combinations. Mean F1-scores stay nearly unchanged, but the false-positive rate rises markedly, and this increase follows a directional pattern tied to bus location.
Nithin Pradeep Nadayil Kizhakkethil, Julian Oelhaf, Christian Bergler et al.
In briefAuto-Director: Multi-agent AI that automatically plans and edits videos
SummaryThis paper presents Auto-Director, a framework of cooperating AI agents that mimics professional video editing: planning a script, indexing footage, assembling clips, and iteratively reviewing the result. The authors also introduce VEdit-Bench, a benchmark of 100 expert-curated topics evaluated on five dimensions. The paper reports that Auto-Director significantly outperforms baselines in both automated metrics and human evaluation.
In briefDeclineCraft: A benchmark testing whether RAG models abstain when retrieval and internal knowledge both fail
SummaryThe paper introduces DeclineCraft, a dynamic benchmark of unanswerable questions about Minecraft features added after a corpus cut-off, with audits for possible parametric exposure. Across 27 LLMs, it measures how often models correctly abstain, how often they attempt answerable questions, and answer accuracy. Single-shot retrieval significantly raises correct abstention in most configurations, while iterative RAG generally lowers it with little change in accuracy. Model size and release date show no significant association with correct abstention.
In briefHistoLoop: a looped Transformer that iteratively infers spatial gene expression from tissue histology images
SummaryThis paper introduces HistoLoop, a looped Transformer that predicts spatial gene expression from H&E histology images as a scalable complement to experimental spatial transcriptomics. At each refinement step, it builds context from current spot states, processes it with a shared spatial Transformer, and updates persistent states through a memory mechanism, moving to finer resolutions. Under the HEST-Bench HVG50 protocol, it achieved the highest macro-averaged gene-wise Pearson correlation among compared methods across six cancer tasks in five organs from HEST-1K. Ablation studies support the importance of recurrent refinement and progressively finer context reconstruction.
In briefBeyond One Sentiment Score: How Topic and Emotion Shape Emoji Meaning on Twitter
SummaryThe paper analyzes 16,575 English tweets and introduces a "stability delta" metric measuring how topically clustered each emoji's contexts are in sentence-embedding space. It finds that emojis used in tightly clustered topics show more polarized crowd-sourced sentiment (Spearman ρ = +0.361), validated against a sentiment lexicon and Spanish 13-emotion ratings. The authors conclude that topical coherence and affective dispersion are distinct dimensions of emoji meaning.
Rohan James, Ioannis Christofilogiannis, Yuuv Jauhari et al.
In briefRecovering Dependent Sources from Multi-Domain Mixed Signals
SummaryThis paper studies a linear mixing model where hidden components are dependent and the data come from multiple domains that differ only by domain-specific scalings of the components. The authors show that, given sufficient domain variability, the latent variables and mixing functions can be identified from second-order statistics alone. They propose the MuDo-CoM algorithm, which generalizes joint diagonalization, and validate it on simulated data and a real fMRI dataset.
In briefObservation Assimilation Transformer: A neural model that refines weather forecasts using sparse station data
SummaryThe paper introduces the Observation Assimilation Transformer (OAT), a geometry-aware neural process that corrects a numerical weather prediction field using sparse station observations to produce a high-resolution probabilistic analysis. It relies on relative positions between locations, predicts at arbitrary continuous locations, and is trained with a CRPS loss. OAT outperforms several baselines on regional and continental datasets with well-calibrated predictions. In a case study downscaling ERA5 (about 31 km) to COSMO (1 km), OAT achieves lower MAE and CRPS for temperature and precipitation at test stations than COSMO.
Christian Donner, Michele Cattaneo, Maxim Samarin et al.
