Strong machine-learning ensembles capture much of the trend, seasonality, and known events in demand, yet they can still understate peaks, miss moving events, or mistake temporary changes for shifts in demand level. We address this gap by framing forecast refinement as a reward-driven decision problem. For each forecast period, we let an agent choose among values already produced by the ensemble: its base forecast, distribution-free bounds, quantiles, and a bias-corrected estimate. We evaluate these choices against context-matched historical periods, accounting for calendar position, recurring and moving events, trend, price, and promotions. We then apply a structural seasonal guard to constrain adjustments and record the model, agent, and planner contributions to every published value. We evaluate our approach across five retail, consumer-goods, and direct-to-consumer deployments over nine months. Our refined forecasts reduce mean WMAPE by 6.3% compared with the stacked machine-learning forecast for every customer, with gains reaching 16.7%, and achieve the lowest error in 35 of 45 customer-months. Compared with statistical forecasts, our approach reduces mean error by 18.7%. These results demonstrate that context-aware, auditable action selection can add forecast value beyond a strong ensemble.
Full-pipeline data analysis over real-world data lakes requires LLM agents to locate and integrate information across large collections of tables with different schemas. Existing data agents follow task-specific exploration, which exhausts the context window in massive lakes. As the context fills, the agent's reasoning degrades, leading it to locate and process incorrect data files. These errors propagate into the final answer. To overcome these limitations, we introduce Ontology-Grounded Data Agent (OGDA), which utilizes a compact ontology to reduce exploration overhead and constrain data processing. OGDA constructs the ontology through agentic hypothesis testing. At each step, a proposer agent induces schemas, concepts, and relations from raw sources, and a verifier checks each proposal against the underlying data. During task execution, the ontology serves as a representation of the data lake for the agent to explore. Every data-processing action is constrained by a task graph, which the agent can extend only after observing the semantic meaning of the data. Across KramaBench and DiscoveryBench with three LLM backbones, OGDA outperforms the strongest baselines by up to 10.82 points. These results suggest that explicitly modeling the data lake is a key step toward reliable end-to-end data agents.
During reinforcement learning, experience can be both internalized in LLM parameters and retained externally as reusable textual skills. Parameter updates alter the model's behavior and the interaction experience from which new skills are extracted. These skills, in turn, guide the model's subsequent interactions and parameter updates. As the model's capabilities evolve, previously beneficial skills may become redundant or detrimental. We present SkillDuet, a framework that coordinates parameter updates and skill evolution through model response feedback. Skills extracted from on-policy trajectories guide parameter updates through OPD at every step. A response-based score, Strength, ranks candidate skills by the magnitude of the shift they induce in the current model's responses, and the highest-scoring skills are periodically consolidated into a persistent global skill set (Global). The updated model and Global jointly guide subsequent interactions, generating experience for further parameter and skill updates. With Qwen3-4B, SkillDuet achieves success rates of 98.57% on ALFWorld and 92.97% on WebShop, exceeding the matched SEED-API baseline by 2.86 and 5.47 percentage points, respectively. Ablations show that response-based skill selection and Global consolidation improve task performance. Further experiments show that the benefits of retained skills depend on both their content and their compatibility with the model. SkillDuet provides a concrete mechanism for model-skill co-evolution and skill-level recursive self-improvement.
The rapid integration of artificial intelligence (AI), particularly generative and machine-learning systems, in climate services development and provision is transforming how climate information is produced, accessed, and interpreted. AI-enabled climate services offer significant opportunities for process automation, optimisation, tailoring, and the translation of complex climate data into actionable knowledge for diverse users. However, the growing reliance on privately controlled algorithms, data infrastructures, and computing facilities raises critical concerns related to governance, transparency, equity and economic sustainability, among others. While the technical architecture of climate services increasingly relies on advanced machine learning and large-scale climate models, their legitimacy as public services may be weakened in the absence of arrangements that prioritise public value and long-lasting societal benefit over profit maximisation. The tension between the public mandate of climate services and the private nature of AI infrastructure challenges traditional notions of openness. Using a consensus-based elicitation method coupled with the PESTLE framework, we analyse the political, economic, social, technological, legal, and environmental risks and opportunities shaping the co-production value chain of AI-enabled climate services. We draw policy-relevant conclusions for data, model and infrastructure supporting the next generation of upstream and downstream climate services as critical applications for society. Our work presents potential and plausible challenges and benefits for a pan-European, decentralised public climate service built on sovereign AI infrastructure and open governance principles.
We present a real-time method for reconstructing fully integrated pixel radiance in rendered volumetric media under direct and ambient environmental illumination, using compact, low-cost scene features and statistics. Unlike Monte Carlo denoising pipelines for volumetric rendering, which depend on computationally expensive path-traced radiance estimates, our approach directly infers the final light-transport solution from inexpensive primary-ray and environment sample evaluations. We show that these features form a near-sufficient representation of the scene, enabling accurate reconstruction at substantially lower computational cost. Using an attention-augmented image-to-image network, we demonstrate that complete pixel radiance can be inferred directly from these auxiliary features, without requiring the original radiance signal. Operating in image space also provides a more stable computational cost as the camera zooms into or out of the volume, with substantially less dependence on volume density than approaches based on explicit multiple-scattering simulation. We demonstrate real-time performance and high reconstruction accuracy in challenging volumetric cloud scenarios.
Large-scale representation learning has produced public foundation models that provide strong general-purpose visual features. However, their pretraining objectives are not designed for robot representation learning, so the resulting embeddings tend to emphasize global semantics rather than the temporally sensitive representations required for robot control. Training robot models from scratch is also undesirable due to the limited scale and high collection cost of robotics data, as well as the substantial computational burden of pre-training. In this paper, we propose SPARK (Simple Post-training for Adapting pRetrained Knowledge), a post-training method that adapts foundation models to robot control. SPARK is built on two principles: temporal abstraction, which encourages the encoder to derive compact features that preserve information essential for understanding temporal changes and action-relevant scene structure, and knowledge preservation, which aligns patch-level representations with those of the original foundation model to retain useful pretrained semantics. These objectives yield compact visual state representations that remain semantically meaningful while preserving useful prior knowledge. Experiments across multiple robotics benchmarks show that SPARK consistently improves success rates and generalization over vanilla foundation models, and further demonstrate that these gains transfer to real-world robot manipulation.
AI systems increasingly appear in criminal investigations as tools, targets, and sources of evidence, yet AI forensics lacks a systematic account of the artifacts they produce: investigators often do not know which traces to look for, and for many traces they do find, no established analysis technique exists. In this paper we first developed the Forensic AI Life Cycle (FALC) that unlike other life cycle models shifts the focus onto the forensic artifacts created during the life cycle of an AI system. Through FALC we derived a reference list of artifacts that might serve an investigation. Second, we collected examples for crimes with AI involvement throughout FALC and questions likely to be asked when investigating these crimes. Third, we propose a taxonomy of forensic fields of expertise with a focus on the current possibility to analyze AI forensic artifacts using established techniques. For investigators, the model provides a systematic basis for identifying which artifacts to look for when an AI system is relevant to a case. For researchers, the classification shows where existing forensic methods already suffice and where new techniques are most urgently needed.
Late Chunking encodes an entire document before pooling at chunk boundaries, preserving document context in local retrieval units. Chunk-level post-training improves context-dependent retrieval, but shared context can make semantically similar same-document negatives difficult to distinguish. We study the local decision boundary between a gold chunk and its hardest sibling. Starting from InSeNT, we analyze InSeNT's within-document gradients and show that the entropy of the conditional negative allocation decreases monotonically with the scaling factor at fixed logits. We propose a two-part framework for calibrating document-local chunk boundaries. The first component adapts the discriminative scale from within-document margins, while the second exposes chunk partitions to attention during document encoding. Experiments on ConTEB and NanoBEIR show that the resulting model improves contextual chunk retrieval across the evaluated encoder settings while retaining competitive passage-level retrieval quality; ablations further characterize the distinct roles of adaptive scaling and chunk-aware routing.
Biomedical ontologies are structured knowledge systems that must evolve continuously to represent newly emerging knowledge, but their manual maintenance remains time-consuming and laborintensive. Textual definitions are especially costly because they must be readable to humans and compatible with controlled entities, relations, and logical constraints. Large language models (LLMs) offer a promising way to assist ontology engineering tasks, including definition generation. However, textual plausibility does not guarantee knowledge-system compatibility. A generated definition may be fluent and close to a reference definition, yet still fail to ground its implied entities or violate ontology-level constraints. Directly inserting such definitions into ontology libraries may therefore introduce entity-level hallucinations and logical inconsistencies. To address this problem, we propose two structured-knowledge incompatibility metrics, Entity Error Rate (EER) and Logical Inconsistency Rate (LIR), to quantify entity grounding and logical consistency for LLM-generated ontology definitions. We further develop an agentic quality-control (QC) workflow that treats definition generation as an iterative process rather than a direct decoding step: it generates definitions, extracts axioms, checks entity existence and logical consistency, and uses the resulting diagnostics to drive repair loops. Comprehensive experiments on Disease Ontology, Uberon Ontology, and Uberon-ConstraintStress demonstrate that EER and LIR complement conventional reference-based metrics by evaluating whether generated content can be grounded and checked inside the target ontology. The workflow effectively decreases EER while maintaining strong conventional text-generation performance. LIR further exposes sparse ontology-level logical failures that are invisible to text similarity metrics, although their repair remains limited.
One of the fundamental needs for reaching resilient production infrastructure is flexible manufacturing, which requires translating changing production requirements into implementable system configurations. This paper introduces \textbf{CRAFT-AI}, an agentic AI approach combining natural language (NL) requirement interpretation, Asset Administration Shell (AAS)-based capability descriptions, and reusable engineering knowledge from existing manufacturing configurations. A tool-using AI agent matches requirements with available capabilities and reference implementations to generate manufacturing scenes and corresponding AAS instances. The approach is demonstrated using a Factory I/O prototype with twelve reference scenes and five illustrative manufacturing tasks of varying complexity, providing an initial feasibility assessment of the proposed AAS-enabled, system-level generative engineering approach. For the four valid user task descriptions, proper requirements, correct references to the relevant engineering knowledge, correct components and lastly a valid Factory I/O scene are generated. For the task with missing engineering knowledge and a wrong sensor selection, the user is informed about which requirements of the scene generation lack domain knowledge and also which sensor is not compatible with the selected raw material by the user with link to the related documentation. After fixing the task description, the correct scene is seamlessly generated. \textbf{CRAFT-AI} highlights the importance of safeguards for risk perception and communication in foundation model-driven solutions.
Wearable Inertial Measurement Units (IMUs) enable pervasive Human Activity Recognition (HAR), but traditional Deep Neural Networks (DNNs) suffer from an "arithmetic-latency deadlock" due to dense floating-point MAC operations. While Spiking Neural Networks (SNNs) offer ultra-low power, existing ones face spatial blind spots, rigid neurons, and gradient degradation. To address these, we propose TSFFormer, a purely spiking Fourier Transformer for Green HAR. It introduces two key innovations: (1) a MultiHeadFourier attention that performs frequency-domain global mixing, reducing complexity from \(O(V^2)\) to \(O(V\log V)\), eliminating spatial blind spots, and saving \(\sim\)50% GPU memory; and (2) a dual-compartment Temporal Spiking LIF (TSLIFNode) that utilizes dynamic thresholds to adapt to non-stationary rhythms and mitigate gradient degradation. Extensive experiments on 8 public IMU benchmarks demonstrate that TSFFormer surpasses state-of-the-art SNNs and even the classic DeepConvLSTM, achieving a \(\sim\)3\(\times\) faster training speed with an energy footprint of 0.1 pJ. Ablation studies confirm the synergistic effect of our modules, providing a highly accurate, ultra-low-power event-driven paradigm for wearable sensing.
Slay the Spire 2 (STS2) is a deck-building roguelike in which a winning run at the highest difficulty takes several hundred decisions, and most losses happen in late boss fights that a learner seldom reaches. We train agents that choose every action with one forward pass of a 6.6M-parameter entity transformer, without lookahead. Training and evaluation run the game's own compiled assembly (v0.107.1) in a headless .NET host. The host replays a run exactly from its seed and actions, so training episodes can restart from states the learner has visited, and a failure-driven curriculum allocates these restarts in proportion to where complete runs are lost. On 12,800 held-out games at Ascension 10 (A10), the released policies win 8.05% averaged over the five characters (Ironclad 15.6%, Regent 12.5%), and AlphaSpire's released networks used without search win 0.55% on a common panel. Longer A10 training raised Ironclad from 12.4% to 15.6% and Defect from 3.0% to 8.1% on these games. At Ascension 0 the released policies win 80.2%, level with the 80.3% that the tree-search agent Spire Mentor reports, while Mentor leads by 5 to 8 points from A4 upward. Run in our host on the same 2,000 A10 games, Mentor's search wins 8.8% against 8.2% for our policies, which need 21-27 ms per decision on one CPU core. GPT-6 Astra at maximum reasoning won none of 11 A10 games through our text interface. In a controlled study of 18 training runs from six random initializations, failure-driven allocation exceeds uniform restart allocation on each of three Regent initializations, by 3.6 percentage points on average, while the Silent effect is small and changes sign. Code, trained policies, an in-game mod and the records behind every table are at https://github.com/HalfVulpes/sts2ai.
Embodied navigation requires agents to achieve persistent spatial awareness and long-horizon execution stability in open-world environments. However, existing navigation architectures struggle to sustain both capabilities, remaining effective only over a handful of recent frames. We present Compass 1, the first embodied navigation foundation model to unify language understanding and video generation while enabling memory backtracking through nearly 2mins of visual history. Without altering the underlying architecture, we introduce History Representation, which integrates sparse visual keyframes with an explicit History Representation Map to preserve long-term spatiotemporal context. To effectively leverage this history, we further introduce History Utilization, a structured VQA supervision protocol comprising three categories: task-stage grounding, exploration consistency, and spatiotemporal target retrieval, enabling map-grounded reasoning and stable persistent long-horizon navigation. To fuel this system, we establish a complete annotation pipeline for monocular RGB videos and open-source a massive dataset featuring over 500 hours of high-quality real-world interactions and over 2,000 hours of annotated web videos. Trained mainly on the distribution of 1m height camera, Compass 1 generalizes seamlessly across diverse embodiments and scenarios, ranging from indoor environments to pedestrian-dense streets, spanning wheeled robot at 0.35m, quadruped dog at 0.5m, humanoid at 0.85m, and aerial drone at 1.5m. Compass 1achieves a dominant overall success rate of 79.4% across open-world evaluations, compared to 17.8% for the strongest baseline. It demonstrates exceptional zero-shot execution exceeding 90% on foundational tasks, a 250% boost in exploration, and an unparalleled leap from 0% to 58% in challenging persistent history backtracking.
We study how much the interaction strategy around a fixed, locally hosted 7B-parameter code model (Qwen2.5-Coder-7B-Instruct) affects its ability to replace PyTorch operators with custom CUDA kernels. On 30 KernelBench tasks (levels 1-3, first 10 each, evaluated on a V100), we compare four regimes of increasing structure: zero-shot prompting, profiling-guided prompting, iterative refinement on execution feedback, and a multi-agent loop (analysis, retrieval, generation, evaluation, feedback triage) in which the same model fills four roles around a non-LLM evaluator. Three findings stand out. (1) Correctness, not speed, is the binding constraint: even correct generated kernels average 0.97-0.99x the speed of eager PyTorch, because a 7B model cannot out-write cuBLAS/cuDNN on dense GEMM and convolution; the only real wins are fused memory-bound epilogues. (2) The best strategy depends on difficulty: profiling-guided prompting is strongest overall (43% correct, 27% faster than baseline), but the agent loop is the only method that solves any level-2 fusion problem (3/10 vs. 0/10 for all others). (3) Naive agentic scaffolding is actively harmful: our first agent scored 3% correct, below zero-shot's 30%, and the decisive repairs were matching the agent's ambition to the model's competence and spending compute on best-of-n sampling under a real execution oracle (3% to 33% correct from n=1 to n=8), not on more reasoning steps.
We present SadaqahStories, a parallel English-Arabic corpus of 41 illustrated four-page stories about charitable giving in Islam (Sadaqah, Zakat and Waqf), written for readers aged 7 to 18. The stories were first drafted with large language model (LLM) assistance, then audited against hadith collections, Qur'an references and historical sources, and finally rewritten for children and signed off by an Islamic scholar. Every story carries an evidence label and links to its sources. We release all three versions of each story (AI-assisted draft, audited text, final child-oriented text) together with the correction log, so the corpus can be used to study how AI drafts of religious and historical content go wrong. The audit found substantive errors in 39 of the 41 drafts. The most frequent were unsupported specific details (dates, figures, prices, names), invented quotations, institutions or anecdotes, and conflation of separate events or people; six drafts contained errors in the wording, grading or ruling of a religious text. Rewriting cut the estimated Flesch-Kincaid grade of the English text from 14.0 to 6.1. We also release an 81-entry charity lexicon covering Modern Standard Arabic and Qatari Arabic, and describe the free, account-free website that serves the corpus (https://sadaqahjourney.com).
Deep learning models, particularly convolutional neural networks (CNNs), are widely adopted in computer vision, yet their opaque decision-making raises concerns in high-risk applications like healthcare and autonomous systems, where trust and accountability are critical. Therefore, robust explainable AI (XAI) methods are desirable to provide transparent insights into model decisions. To ensure these methods are reliable and effective across diverse scenarios, it is important to understand how model performance and architectural complexity influence explanation quality. Our study takes an initial step toward systematically investigating the impact of these two factors on key XAI evaluation metrics. We constructed a pipeline using 10 subsets of ImageNet-1K, each containing 80 unique classes, on each of which we trained three CNN architectures of increasing structural complexity: VGG16, ResNet50, and ConvNeXt-Tiny. For each individual model, we analyzed 10 training checkpoints representing increasing levels of model accuracy using two XAI methods: Grad-CAM (feature-attribution-based) and CRAFT (concept-based). We assessed these methods across four evaluation categories: Faithfulness, Correctness, Sensitivity, and Simplicity. Our findings reveal that as training progresses, the complex architectures (ResNet50 and ConvNeXt-Tiny) exhibit improved Faithfulness and lower Sensitivity, indicating more robust explanations. In contrast, VGG16's simpler architecture yields simpler but less faithful and correct explanations.
Remote face authentication requires protection against both presentation attacks and biometric data injection attacks prior to face matching. Methods based on active screen illumination appear promising for both by introducing randomized visual stimuli through the user's device screen and evaluating the corresponding facial responses. Although such methods have been more studied for presentation attack detection (PAD), their systematic evaluation for injection attack detection (IAD) remains limited. To address this gap, we first classify injected content into static images, prerecorded videos, and real-time deepfakes based on their expected responses to the challenge. We then adapt an active screen-illumination approach into an implementation we term color wave and develop a systematic, injection-specific evaluation protocol. The protocol examines detection effectiveness, generalization across operating conditions, robustness to signal degradation, and susceptibility to selected evasion strategies. The findings demonstrate a strong discriminative signal under controlled conditions but also reveal performance limitations under overexposure, low screen brightness, and subject motion, as well as vulnerabilities to selected evasion strategies. By relating detection outcomes to attack types and operating conditions, this study provides an interpretable assessment of when and why active screen illumination succeeds or fails as an IAD mechanism.
Video diffusion transformers (DiTs) generate realistic videos, but their heavy computational cost limits broader real-world applications. Mixed-resolution generation reduces this cost by generating the visually important region at high resolution (HR) and the rest at low resolution (LR), yet existing methods suffer a severe quality drop on videos and require knowing the HR region in advance. We show that these failures are not inherent to mixed resolution, and propose $\textbf{MPEG-Diffusion}$ to address their three root causes: (1) dual-resolution attention to mitigate the curse of mixed resolution, where each query reads the K/V of its own resolution, (2) a mixed-resolution decoder to blend the HR and LR latents before decoding and resolve distortion, and (3) a self-selected HR region using the DiT's knowledge to find the visually important region without requiring it in advance. Processing only $30\%$ of full-resolution tokens, MPEG-Diffusion speeds up denoising by $4.8\times$ on Wan2.1 14B ($3.8\times$ on 1.3B), outperforming prior methods on nearly every metric. In the user study, we demonstrate the high quality of MPEG-Diffusion, with raters preferring it over prior mixed-resolution methods in over $90\%$ of pairs and rating it close to full resolution.
Long-term memory lets LLM agents personalize, stay consistent, and reuse what worked before. Parametric memories need training, consolidation can lose details for good, and keeping every record leaves a growing store to search. We present MIRIX-2.0, a multi-agent memory system designed to be practical, accurate, and efficient at once. It is the most accurate of the compared systems on four of five QA benchmarks. It reaches 91.4% on LoCoMo and 72.6% on LongMemEval-S, where it leads the strongest baseline by 14.3 points, and with AutoDream disabled it gains 31.0 and 22.0 points over HyperMem on single-hop and multi-hop document QA. As procedural memory, it yields the highest mean success among the compared memory methods on ALFWorld and WebShop and reaches 49.2% on MetaClaw-Bench, against 44.0% for MetaClaw's skill library. Six typed memory agents keep complete records as the system of record, and a provenance hypergraph indexes episodic and semantic records with n-ary frames that cite their rows. Retrieval runs under explicit budgets, so the evidence a question inspects does not grow with the history. AutoDream merges duplicate anchors outside the write path without rewriting any record, and it distills task sessions into skills. With AutoDream disabled, MIRIX-2.0 answers faster than HyperMem on all five timed benchmarks. It needs no fine-tuning, runs with closed and open models, and its code is public.
Can a large language model (LLM) write the parameters of a medical image classifier from medical knowledge alone, without any training data? Such classifiers normally require labelled images, which are costly and often unavailable. We investigate this question through weight-space programming (WSP), in which an LLM writes or adapts the weights and biases of a neural network directly, without optimisation on training data. Claude Opus 5.5 and GPT-6 Astra each generate the parameters of a small LeNet-style CNN in a single reply, without tool use and without access to any image. We compare these networks against zero-shot MedSigLIP and against the same architecture trained on labels using the Adam optimiser, across 7 classification tasks in four modalities: chest X-ray, breast ultrasound and abdominal CT (all from MedMNIST), and pelvic MRI (SynthUterus). The LLM-written networks perform above chance on all six MedMNIST tasks, with a mean test AUC of 0.719 for Opus and 0.711 for Astra. Both models outperform MedSigLIP on 5 of 6 tasks while using more than 10,000 times fewer parameters. However, the same network trained with Adam on only 256 labelled images achieves a higher AUC on every MedMNIST task (mean AUC 0.828). On SynthUterus, Opus attains an AUC of 0.764, whereas Astra falls below chance (AUC 0.136) owing to a systematic error. These results show that LLMs can write functional medical classifiers, but that they do not yet replace training data. We release the protocol, logs and all generated weights at https://github.com/MaximilianLindholz/weight-space-programming-medical.