**Purpose:** Ultrasound biometry is widely used across various clinical applications, including prenatal and cardiac examinations, to obtain quantitative measurements by placing calipers at anatomical landmarks. However, these measurements remain subject to interoperator variability as they rely largely on manual caliper placement. Conventional methods often use specialized prediction heads for different ultrasound tasks, whereas vision-language models offer flexibility through text instructions. However, direct coordinate prediction does not explicitly model spatial likelihood distributions, limiting the representation of localization ambiguity. **Methods:** We propose Kev-Bio, a vision-language model that uses clinical measurement questions to guide ultrasound landmark detection. Inspired by System One decision models, Kev-Bio uses decision representations of the language model to generate queries for each caliper endpoint. These queries are matched against shared dense ultrasound features to produce spatial probability maps, from which landmark coordinates and clinical measurements are derived. The model is trained end-to-end using both landmark localization and clinical measurement losses. **Results:** On the nine-view FU/GU Biometry 2026 dataset, Kev-Bio achieves a mean radial error (MRE) of 22.50 pixels and a clinical measurement mean absolute error (MAE) of 15.38, outperforming the compared baselines in overall performance. **Conclusion:** These results demonstrate the effectiveness of Kev-Bio for ultrasound landmark detection and clinical measurement across different anatomical views. Our code and models are available at https://github.com/moeinheidari7829/Kev-Bio.
Suicide prevention remains a critical public health challenge, and electronic health records (EHRs) provide an opportunity to characterize suicide-related risk indicators using documented clinical, behavioral, medical, and social-contextual information. This paper presents a multidimensional machine learning framework to model and estimate suicide-related outcomes using MIMIC-IV EHR data. We construct clinically motivated diagnosis categories and define a composite outcome from documented codes corresponding to suicidal ideation, suicide attempt, prior suicidal behavior, and intentional self-harm. To address severe class imbalance and reduce age-related confounding, we apply age-matched resampling that retains all positive patients and samples comparison-cohort patients from the same age group. Logistic regression, random forest, and XGBoost classifiers are evaluated before and after resampling with isotonic calibration. The calibrated XGBoost model achieves the strongest matched-cohort performance, outperforming the other evaluated classifiers. SHAP-based interpretation shows that diagnostic burden, depression-related features, age, personality-disorder history, social determinants related to violence, and medical comorbidity contribute to elevated predicted probabilities. Our results demonstrate that combining broad EHR-derived features, age-matched cohort construction, calibration, and feature-level interpretation provides a practical framework for retrospective suicide-related risk modeling.
Robust machine learning for graph-structured data has made significant progress against test-time attacks, yet certified robustness to poisoning - where adversaries manipulate the training data - remains largely underexplored. For image data, state-of-the-art poisoning certificates rely on partitioning-and-aggregation schemes. However, we show that these methods fail when applied in the graph domain due to the inherent label and structure sparsity found in common graph datasets, making effective graph-partitioning difficult. To address this challenge, we propose a novel semi-supervised learning framework called deep Self-Training Graph Partition Aggregation (ST-GPA), which enriches each graph partition with informative pseudo-labels and synthetic edges, enabling effective certification against node-label and graph-structure poisoning under sparse conditions. Our method is architecture-agnostic, scales to large numbers of partitions, and consistently and significantly improves robustness guarantees against both label and structure poisoning across multiple benchmarks, while maintaining strong clean accuracy. Overall, our results establish a promising direction for certifiably robust learning on graph-structured data against poisoning under sparse conditions.
While structural and functional magnetic resonance imaging (sMRI, fMRI) offer profound potential for researching autism spectrum disorder (ASD) and understanding individual neurodivergence, current classifiers lack the explanatory power needed for scientific study, clinical use and personal-level insights. Most models learn from imaging features that also vary with age, sex, scanner site and head motion. They fuse the two modalities without a common reference and explain decisions against arbitrary reference inputs rather than a typical brain. We present NormMoX, a normative mixture of experts that makes the typical brain the shared reference for prediction, fusion and explanation. Normative models trained on large healthy cohorts give the value each brain measure is expected to take for a person's age and sex. NormMoX learns from each person's deviations from these expected values. For each modality, a neural network expert turns the deviations into a score of evidence for ASD. A fusion head, MoX, compares each score with those of typical controls before combining them. It therefore needs no imputation when a modality is missing and shows in which modality each person is atypical. On 1,035 participants of ABIDE I, NormMoX reached an AUROC of $0.728 \pm 0.004$ in repeated cross-validation, above linear, single-modality and deep fusion baselines. NormMoX thus shows not only whether a person is likely to have ASD, but where, and in which modality, their brain departs from typical. The implementation is available at https://github.com/Novelex/NormMox.
Attention remains a substantial computational cost in Transformer training, particularly at long context lengths. NVIDIA Blackwell GPUs provide high-throughput low-precision matrix multiplication, but exploiting it in attention requires addressing quantization overhead, causal information leakage through shared scales, and numerical inconsistencies between the forward and backward passes. We present FLAMIA, a family of attention kernels for Blackwell with MXFP8 and mixed NVFP4/FP8 training paths. Our design combines token-block scaling that reuses quantized operands across matrix orientations, targeted safeguards against scale-induced leakage, and consistent forward and backward computations. Fused quantization and architecture-specific kernel optimizations reduce overhead, while document-local scaling and workload-aware scheduling support variable-length documents directly in packed layouts without repacking. Across various causal packed-document workloads, the MXFP8 path accelerates the combined forward and backward attention call, including quantization, by factors of $1.10$--$1.40\times$ on B200 and $1.06$--$1.42\times$ on B300 relative to the fastest evaluated baseline for each workload. We validate training stability and quality preservation on pretraining Llama-style models. These results show how jointly designing quantization, training numerics, and document scheduling translates low-bit hardware throughput into practical attention acceleration.
Geospatial text spotting involves locating and transcribing geographic labels, such as road names, district names, and landmarks, in raster maps. These labels are often small, densely arranged, and arbitrarily oriented. Although Arabic text recognition has been studied extensively in documents and natural scenes, Arabic geospatial text spotting remains underexplored. We introduce ARABGEOTEXT, an Arabic-centric benchmark covering 970 geographic locations in Riyadh and Jeddah, with 17,875 manually annotated text instances per view type. For each location, we pair an OpenStreetMap (OSM) rendering with a composite created by overlaying that rendering on the corresponding remote sensing (RS) imagery. The paired views preserve label content and pixel geometry, allowing evaluation across different backgrounds. Geographic partitions support in-city and city-disjoint evaluation of detection, recognition, and end-to-end spotting. Experiments with direct Qwen3.5-4B spotting show difficulties in recovering all labels within a tile. We therefore propose a tool-augmented vision-language framework that combines YOLO26 localization and crop rectification with Qwen3.5-4B recognition adapted through low-rank adaptation (LoRA). The framework outperforms direct LoRA-adapted Qwen3.5-4B spotting across both image views and test cities. With localization held fixed, reader adaptation further improves recognition over the pretrained zero-shot reader. The dataset and code will be made publicly available at https://github.com/yakoubbazi/ArabGeoText.
Generative AI is billed per token, yet tokens are not equally expensive. Each token enters the KV-cache and raises the cost of those that follow, so energy grows non-linearly along a sequence whose tokens look identical. No analytical model states what a token costs, and the most accurate SOTA model is a fit, not a law: it absorbs architecture into fitted coefficients, requiring per-LLM fitting. We derive closed-form expressions of the computational and memory-access complexity of LLM inference in terms of architectural hyperparameters, covering MHA and GQA, and compare a family of analytical models of increasing granularity. From these, we obtain the SIMPLE Inference Model to Predict LLMs Energy (SIMPLE), a four-parameter law that separates prefill from decode, and decode into KV-cache cost and next-token computation. Over traces of 13 dense LLMs from six families on NVIDIA H100 GPUs, it attains R2 = 0.969 and 11.97% MAPE, 3.1-3.5x below the SOTA, without per-LLM calibration. The SIMPLE law explains as well as predicts: prefill energy grows quadratically with the hidden dimension, so larger LLMs suffer most from long inputs; and more aggressive GQA KV grouping reduces the per-token energy by up to 3x relative to standard MHA. Increasing the reasoning token budget pushes output length beyond the analytically derived energy minimum, driving super-linear energy growth; although accuracy grows with the budget, energy grows faster, so Intelligence per Joule peaks at an intermediate budget and falls by up to 3.14x across a 4.0x reasoning-budget range. Agentic serving raises energy per request by 13.0x relative to non-agentic ones, driven by the growth of both the prefill and the KV-cache terms with input length; while prefix caching removes most of the prefill cost (-93%), the KV-cache term is unaffected, so even a cached agentic request costs 4.8x more.
Artificial intelligence (AI) technologies and their capabilities have advanced rapidly over the past few years, leading to widespread interest and adoption in the biological and ecological sciences. However, the data infrastructure and computational resources that provide the foundation for AI workflows in ecology and biodiversity science remain fragmented. Moreover, the flow of value remains largely unidirectional: data move from collectors, through repositories and aggregators into models, and finally to research products or management actions. The derived value, such as corrected taxonomies, annotations, quality flags, and insights, rarely returns to source communities. We propose an extension of the original FAIR principles to address the needs of AI-ready infrastructure: the AI-FAIR metadata criteria, which place AI-enabled science in the context of a bidirectional Question-Data-Model triangle. This framework is designed to facilitate large-scale content-based analysis across diverse collections and ensure that the scientists generating the data benefit from the AI-enabled advances built on their data. Alongside these criteria, we introduce an agentic evaluation workflow that both rates existing datasets on the AI-FAIR and traditional FAIR criteria and provides actionable feedback where critical metadata is missing or incomplete. This modular dataset evaluation tool, paired with a domain-appropriate checklist, translates these principles into a reusable assessment of machine-learning readiness, AI-readiness for specified tasks, traceability, and ethical and CARE-aligned stewardship. Through this approach we aim to create a community-informed system to close the model data loop, so that AI also enriches data to further enable scientific discovery.
Activation steering offers a lightweight mechanism for controlling model behavior, but existing methods depend on sentence-level contrastive data that is unavailable for most of the world's languages. We show that multilingual sentiment lexicons are a viable substitute across three model families: aggregated multilingual lexicons outperform English-derived steering vectors for mid- and low-resource languages, while monolingual lexicons remain stronger for languages a model has specialized in during pre-training. The steering efficacy is mechanistically dictated by both representational density and pre-training priors. Large-magnitude English vectors overwhelm the compact representations of low-resource languages, causing "linguistic collapse." This fragility is mitigated by the norm-dampening of multilingual aggregation. Architectural biases, including bilingual specialization penalty and latent numerical rigidity, also heavily constrain cross-lingual steering responsiveness. Ultimately, our work establishes lexicons as a practical steering signal and offers a comprehensive mechanistic account of cross-lingual steering dynamics.
Previous research shows that vision-language (VL) models struggle to understand fine-grained image details and tend to focus on the overall layout. Moreover, any weakness in English tends to be further amplified in the multilingual setting. To study that, we extend InpaintCOCO, an English dataset focused on fine-grained visual details, to 13 typologically diverse languages across five language families, creating MultiInpaintCOCO - a new multilingual benchmark of 953 minimal-pair examples. We further test several state-of-the-art models in multilingual multimodal retrieval, considering both light VL CLIP-like encoders and large language models. Our results show that performance degrades substantially for lower-resource languages, with the strongest predictor being the data resources available for the languages rather than typological similarity to English. In cross-lingual retrieval, most models exhibit a systematic bias toward English: when given a correct caption in a non-English language and an incorrect one in English, models preferentially select the English caption. This bias varies considerably across architectures. Among the tested models, Gemma 3 shows the strongest English preference while Qwen 3 is the least biased, despite both reaching similar monolingual accuracy across languages. The code is available at https://github.com/ufal/multi-inpaint-coco.
A frontier language model was recently reported to solve 1D NMR structural elucidation without tools. We ask whether this capability transfers to tandem mass spectrometry (MS/MS) in the de novo setting, where the model must propose a structure directly from the spectrum without a candidate list or reference library. To separate reasoning from recall, we build a 100-analyte benchmark from MassSpecGym, split by each analyte's textual footprint into in-distribution (well documented in literature) and out-of-distribution (essentially undocumented) arms. We hold the model, harness, and output format fixed and vary only the LLM's access to domain-specific tools. Without tools, a frontier LLM reaches 52% Top-1 accuracy in-distribution but only 2% out-of-distribution. Only access to de novo structure generators improves out-of-distribution accuracy, raising it to as much as 38% Top-1, and the resulting agent outperforms the standalone de novo models both in and out of distribution. These results suggest that LLMs are not effective at de novo MS/MS structural elucidation on their own, and that further gains depend on the specialist tools available to the agent rather than on the LLM itself.
Calibrated multiaccuracy asks that a predictor be globally calibrated and multiaccurate with respect to a bounded test family $G$; that is, its residuals should be uncorrelated with every test $g(x)\in G$. This weakening of multicalibration suffices for many applications. Its sample complexity was not known to improve on that of full multicalibration, which is $\widetilde\Theta\left(\frac{\log|G|}{\varepsilon^2} + \frac{1}{\varepsilon^{3}}\right)$ when $|G|$ is at least a sufficiently large power of $\log(1/\varepsilon)$. We resolve the sample complexity of $\varepsilon$-calibrated multiaccuracy, showing that it scales as $$ \widetilde\Theta\left(\frac{\log|G|}{\varepsilon^2} + \frac{1}{\varepsilon^{5/2}}\right), $$ separating it from the complexity of multicalibration. The upper bound is obtained by an algorithm that outputs a randomized predictor. For deterministic predictors, we establish the tight rate $$ \widetilde\Theta\left( \frac{\log|G|}{\varepsilon^2} + \frac{(\log|G|)^{1/3}}{\varepsilon^{8/3}} \right) $$ whenever $\log|G|=\Omega(\log(1/\varepsilon))$. When $\log|G|\le\varepsilon^{-1+\kappa}$ for any fixed $\kappa\in(0,1)$, this rate is polynomially larger than the rate for randomized predictors. Thus, unlike for multicalibration, calibrated multiaccuracy exhibits a genuine sample-complexity separation between deterministic and randomized predictors. The matching upper and lower bounds show that the unusual cube-root dependence on $\log|G|$ is intrinsic. At $\log|G|\asymp\varepsilon^{-1}$, the minimax rates for randomized and deterministic calibrated multiaccuracy both coincide with the minimax rate for multicalibration.
A model pretrained to predict a customer's next event can estimate the probability of future events without training on their labels and provide a reusable representation for supervised downstream banking tasks. We pretrain EventGPT, a hierarchical causal transformer, by next-token maximum likelihood estimation on the public Multimodal Banking Dataset (MBD-mini: 100,224 clients, 38.8 million transactions and 66.3 million geolocation pings). Keeping one state per event, EventGPT trains in about two hours on one L4 GPU at windows up to 8,192 tokens. From its next-event distribution, with no labels, EventGPT scores five targets of the form "the client makes a transaction of a given kind for the first time next month" at ROC-AUC 0.79 to 0.92. Gradient boosting on 114 engineered features needs 10,000 to 300,000 labeled client-months to match it on four targets and stays below it on the fifth. For product prediction, frozen EventGPT embeddings score below the engineered features (mean ROC-AUC 0.787 against 0.808 over five test folds), and a rank-averaged ensemble of EventGPT, profile fields and token counts with the engineered features adds 0.014 to 0.017, of which 0.007 to 0.016 is EventGPT's own. The code and pretrained weights are released.
Inverse Reinforcement Learning (IRL) aims to recover expert rewards that transfer across environments with differing dynamics. Prior work has shown that such transfer is possible when gathering expert demonstrations across sufficiently different dynamics. In practice, however, expert demonstrations are costly, making it critical to select the right dynamics to reduce the quantity of expert data required for learning transferable rewards. We propose a two-phase approach: we first cycle through the available dynamics to obtain a reliable reward estimate, then actively select dynamics to reduce our uncertainty about the expert reward. Crucially, this yields novel sample complexity guarantees on transfer that hold for any dynamics, with a rate independent of the number of available dynamics. Empirically, our method significantly reduces the number of expert demonstrations needed to learn a transferable reward.
Zeroth-order (ZO) optimization enables memory-efficient large language model (LLM) fine-tuning, and recent methods incorporate curvature-aware strategies. However, in the high-dimensional parameter space of LLMs, these methods inherently suffer from exploding gradient variance and prohibitive costs. Conversely, confining updates to a low-dimensional subspace reduces variance but discards critical high-dimensional information, degrading performance. This creates a fundamental dilemma: full-space optimization is informative but has high variance, while subspace optimization has low variance but loses information. To break this trade-off, we propose ENGINE (Endogenous Variational MultiScale). Inspired by computational mechanics, we adapt static Variational MultiScale (VMS) theory to dynamic optimization, treating low- and high-dimensional spaces as resolved coarse and unresolved fine scales. Through a multigrid-style V-cycle, ENGINE alternates between spaces, dynamically probing fine-grained information in the high-dimensional space and injecting it to correct the efficient low-dimensional sampling process. Under appropriate assumptions, statistical analysis characterizes variance reduction and correction of curvature-induced bias, and convergence analysis establishes a stationarity bound for the corrected coarse updates. Evaluations on standard LLM benchmarks show that ENGINE achieves higher accuracy and faster convergence than existing ZO methods, achieving up to a $2.61\times$ speed-up in forward passes with zero additional peak memory overhead. Code is available at https://github.com/AI4Science-WestlakeU/ENGINE.
A memory-augmented agent must decide which memories to keep, merge and retrieve, and its only evidence of their value is its retrieval log: the memories it retrieved in each episode and whether it succeeded. Existing methods assign every memory a value from this log, even when the log cannot separate one memory's contribution from another's. We treat the log as an experiment that nobody designed and ask which memory values it determines. Two counts answer this before any outcome is observed: how often each memory was retrieved, and how often each pair was retrieved together. From these counts, we prove that top-$k$ retrieval leaves every memory value undetermined, however many episodes are logged, and that a question with its own memories, asked once, determines at most one of their values. On fourteen standard memory benchmarks, no memory value is determined in any of 55 configurations. Both causes are choices about retrieval. We introduce LEDGER, which estimates each determined Shapley value in closed form with a confidence interval and reports every other value as undetermined. It merges memories with identical histories without losing a determined value, and it retrieves so that the next log determines more. On pooled HotpotQA, LEDGER raises the fraction of memories determined from 0.001 to 0.797, at a cost of 2.6 accuracy points on HotpotQA and with a gain of 2.7 on LoCoMo.
Monocular head pose estimation is usually formulated as absolute regression from a single face crop, yet many applications observe the same person over time and can supply an explicit anchor image. Recently, relative head pose estimation was proposed, which casts the problem as estimation of the relative geometric change between an anchor and a query image. However, existing methods still require the anchor image pose to be known at inference, significantly limiting their practical use. We first extend relative head pose estimation to full 6DoF by additionally predicting a depth-normalized relative translation, which composes with the anchor to recover metric query pose in the camera frame. To eliminate the need for a ground-truth anchor pose, we reconstruct a personalized avatar from a single unlabeled enrollment image and render a bank of virtual anchors with known poses. The bank provides robust rotation hypotheses through multi-anchor consensus, while the real enrollment image serves as the real anchor in the translation branch via landmark PnP. We also introduce an efficient variant that amortizes the bank cost once at enrollment, after which each query requires a single real-to-real relative prediction, with limited accuracy sacrifice. The relative pose estimator is trained exclusively on synthetic face renderings, without real pose supervision, pseudo-labels, domain adaptation, or fine-tuning on the evaluation sets. On real 6DoF benchmarks, ReVA-HPE achieves state-of-the-art rotation accuracy on BIWI and, given a near-frontal enrollment image, on FaMoS, while remaining competitive in metric translation despite relying solely on synthetic pose supervision. Project page and code: https://vasilikivas.github.io/ReVA-HPE/
The widespread adoption of Large Language Models (LLMs) demands reliable detection of human-authored versus synthetic text. However, current Transformer-based detectors often experience severe performance degradation during domain shifts and operate as opaque classifiers. This paper investigates the robustness and interpretability of fine-tuned Transformer models (DistilBERT, RoBERTa) by evaluating them against a custom out-of-distribution (OOD) dataset of hyper-local news synthesized via open-source LLMs (Llama 3, Mistral, Gemma). To analyze model decision boundaries, we apply a tripartite Explainable AI (xAI) framework using SHAP. Our analysis reveals three key phenomena: First, detectors often rely on a "Call-to-Action" syndrome inherent to instruction-tuned LLMs rather than generic machine text. Second, we observe a "Specificity Bias," where LLMs successfully bypass detection by hallucinating specific named entities and mimicking conversational imperfection. Finally, we illustrate the "Wikipedia Paradox," wherein formal, objective human writing is frequently misclassified as AI due to its low statistical burstiness. These findings demonstrate that purely statistical detection paradigms face systematic limitations in highly structured human domains, emphasizing the critical need for semantically-aware, interpretable detection architectures.
In node selection tasks, the goal is to identify a single unknown target node in a graph. We consider probabilistic supervised node selection, where the estimator outputs a probability distribution over the nodes of the graph and the ground-truth node is available for evaluation. Classical evaluation metrics for such tasks are either statistically ill-conceived or blind to the graph structure. To fill this gap, we introduce two strictly proper scoring rules that are explicitly graph-aware: the Resistance score, which is parameter-free, and the Diffusion score, which operates at an adjustable coarse-graining resolution. We show that the Diffusion score interpolates between the multi-category Brier score and a cluster-level weighted Brier score. The Resistance score generalises two classical scoring rules: the multi-category Brier score and the Ranked probability score. Applied to rooted trees, it also serves as the strictly proper version of the improper Hierarchical error distance used in hierarchical classification. In the same setting, we furthermore establish strict propriety of the Hierarchical cross-entropy family, a recently introduced separate class of hierarchy-aware scoring rules. Experiments on epidemic source detection and on ImageNet classification confirm that the proposed scores detect meaningful differences in estimation quality that classical metrics miss.
Large-scale annotation of environmental sound recordings presents significant challenges under time and resource constraints, particularly when sound categories are imbalanced and only part of a dataset has been labeled. This paper describes the development and evaluation of a human-AI workflow used during annotation of the Kenya Forest Sounds Dataset (KFSD), using a project-stage working collection of 7,531 audio recordings across 21 categories. We investigated zero-shot classification, a feature-based ensemble, and a fine-tuned Audio Spectrogram Transformer (AST). The fine-tuned AST achieved 89.36% accuracy, 85.98% precision, 89.24% recall, and 87.58% F1 on held-out test data for binary classification of Person talking. Because this result remained below the predefined 95% threshold for autonomous annotation, model predictions were used as suggestions and subjected to human verification. Of the 5,026 initially unlabeled recordings, 2,337 were routed to a speech-verification queue and 2,689 remained for ordinary manual annotation. The verifier reviewed all 2,337 machine-labeled files and detected zero false positives, yielding 100% verification accuracy. The observed average verification time was approximately 4.2 seconds per file. At an average rate of 35 seconds per file for manual annotation, the hybrid workflow reduced work on the remaining queue from approximately 48.9 to 28.9 person-hours, saving about 20.0 hours (41%). The study documents the model-selection process, provenance-aware file organization, verification procedure, observed errors, practical limitations, and opportunities for extending AI assistance beyond one category.