Extended Reality (XR) technologies, including Virtual Reality (VR) and Augmented Reality (AR), offer immersive experiences merging digital content with the real world. Achieving precise spatial tracking over large areas is a critical challenge in XR development. This paper addresses the drift issue, caused by small errors accumulating over time leading to a discrepancy between the real and virtual worlds. Tackling this issue is crucial for co-located XR experiences where virtual and physical elements interact seamlessly. Building upon the locally accurate spatial anchors, we propose a solution that extends this accuracy to larger areas by exploiting an external, drift-corrected tracking method as a ground truth. During the preparation stage, anchors are placed inside the headset and inside the external tracking method simultaneously, yielding 3D-3D correspondences. Both anchor clouds, and thus tracking methods, are aligned using a suitable cloud registration method during the operational stage. Our method enhances user comfort and mobility by leveraging the headset's built-in tracking capabilities during the operational stage, allowing standalone functionality. Additionally, this method can be used with any XR headset that supports spatial anchors and with any drift-free external tracking method. Empirical evaluation demonstrates the system's effectiveness in aligning virtual content with the real world and expanding the accurate tracking area. In addition, the alignment is evaluated by comparing the camera poses of both tracking methods. This approach may benefit a wide range of industries and applications, including manufacturing and construction, education, and entertainment.
The dominant metric for 2D human pose estimation, OKS-mAP, has three blind spots: it does not distinguish visible from occluded keypoints, does not measure joint-angle error, and is silent on the temporal stability of tracking-free per-frame inference. We propose a supplementary protocol with three modules: (M1) OKS-mAP stratified by ground-truth visibility, with Delta_occ = mAP(v=1) - mAP(full), quantifying how much the aggregate metric masks occlusion-specific differences; (M2) joint-angle error (RMSE/MAE/P95) plus knee-valgus detection consistency; (M3) frame-to-frame displacement, median-filter residual jitter, and jump rate, before and after OneEuro smoothing. We ran it on four datasets (COCO val2017, CrowdPose, OCHuman, PoseTrack18) and four models from two architecture families (YOLO11s-pose, an occlusion-domain fine-tuned variant, ViTPose-B/L), under a common detector-free oracle protocol and an open-source harness. Delta_occ is negative in every combination (-0.19 to -0.47): aggregate mAP overestimates occluded-scene performance in every tested combination. Joint-angle error is amplified in occluded layers; angle metrics reveal differences mAP cannot localize, e.g., the smaller occlusion-trained model significantly outperforming a larger transformer in occluded elbows. On PoseTrack18 val, raw jitter matched between the two YOLO models; pixel-domain OneEuro smoothing cut jump rates 9-10% at no matched-scope AP cost; residual jitter persists after detector removal; the transformer pair's jump rate is significantly lower; smoothing benefits replicate across architectures. Stratified mAP tracks full mAP closely (sign-consistency 1.0, 12/12), serving magnitude diagnosis rather than re-ranking; angle-metric difference vectors (especially knee) partially diverge, suggesting independent information. The protocol and harness are open-sourced, complementing rather than replacing OKS-mAP.
Seismic performance assessment relies on computationally expensive nonlinear simulations, limiting uncertainty quantification and inference of structural properties from observed responses. We develop a manifold-based generative framework that learns the joint distribution of uncertain structural parameters and seismic-response trajectories from limited high-fidelity simulations. Combining Diffusion Maps, score-based Diffusion Modeling, and Geometric Harmonics, the framework enables both forward probabilistic prediction and inverse inference. Applied to a nine-story steel moment-resisting frame, the results reveal which structural properties are identifiable from seismic performance and how uncertainty in these properties influences response. In particular, the learned representation captures post-capping degradation characteristics more effectively than individual strength parameters. Inverse inference identifies distinct structural configurations that produce similar seismic performance, with representative solutions independently validated using nonlinear simulations. These findings demonstrate how generative learning can support uncertainty-aware seismic assessment while revealing structural non-identifiability and alternative mechanisms underlying observed performance
This work presents a new benchmark that aims to address the limitations of current methods for evaluating LLMs in Brazilian Portuguese. We introduced Brasilidade, an assessment methodology comprised of standardized tasks distributed across four different domains, and we are making available the tasks from its first domain, which address nuances of the language, grammar, and variations of Brazilian Portuguese. The results show that model size, by itself, does not determine performance; several medium-sized models outperformed larger models in the protocols evaluated. Specific models for Portuguese, in general, did not perform above average, despite their targeted training. Different task levels revealed significant nuances: while some tasks proved easy for most models, others posed extreme challenges, and some yielded mixed results, with wide variations in performance. This clearly illustrates the need for benchmarks that address the limitations of the models in Brazilian Portuguese, allowing future training cycles to seek to overcome them. Project repository https://github.com/soberania-oficial/brasilidade
Large language models (LLMs) are increasingly being deployed in clinical documentation workflows, where clinicians review and edit their outputs before use. In principle, these clinician edits represent a scalable source of supervision for model training. In practice, not all edits represent actionable feedback: for instance, on data drawn from real-world deployment of an AI clinical documentation product, our review suggests that over 45% of edited outputs contain user-added information that is not directly supported by the inputs to the model. Naively fine-tuning on such edits risks teaching the model to hallucinate. In this paper, we develop a method to address this challenge. Given that real clinician edits on deployed documentation tools are not publicly available, we first develop a taxonomy of clinician-editing behaviors, informed by data from a deployed AI clinical documentation product. Then, we develop a method for generating semi-synthetic clinician-edit data, informed by our taxonomy, on a discharge-summary generation task based on MIMIC-III, a public dataset. Finally, we introduce simple filtering methods for identifying actionable edits, with accuracy evaluated by expert review. We show empirically, on our semi-synthetic dataset, that naive fine-tuning on raw edited output increases ungrounded content, while our approach of progressively filtering toward grounded edits enables improvement on clinician-alignment metrics while retaining faithfulness to the input context.
Context compression is now routine for language models with long inputs, but a compressed context is useful only if it keeps the evidence for the answer. Existing checks compare the reader model's answer or its most likely words, and thus miss a frequent failure, silent evidence loss: key evidence is removed while the answer stays the same, leaving the answer unsupported and risky for other models that reuse the context. To address this challenge, we propose SELD (Silent Evidence Loss Detector), which needs no training and no labels. Its principle is that when evidence is lost, a reader model's preferences among the answers it leaves unsaid shift even if its answer does not, so comparing these preferences before and after compression reveals losses that answer checks cannot see. We also build SEL-Pairs, a benchmark of length-matched evidence and control deletions from three question-answering datasets. Quantitative and qualitative experiments with five reader models on it show that SELD consistently outperforms strong and recent baselines across compressors, compression rates and models, identifying the evidence-removing edit in 83% of matched pairs with unchanged answers, against 70% for the strongest baseline. These results show that evidence loss in compressed contexts can be detected even when the answer is unchanged.
Time series from real-world systems are often accompanied by text, such as a report on an event that affects them. Two common tasks on these series are understanding, e.g., answering questions about a series, and forecasting its future values. Both tasks require reasoning over the series and the text together. Moreover, understanding outputs text while forecasting outputs a time series, so serving both requires a model that reads and generates both modalities. Large language models (LLMs) read time series as digit tokens and handle long series poorly. Time series foundation models (TSFMs) forecast well from large-scale pretraining but typically neither read nor generate text. Time series language models add a time series modality to the LLM but forecast on the LLM side, leaving the TSFM's pretrained forecasting ability unused. We present TimeOmni-o1, which couples a pretrained TSFM and an LLM bidirectionally. In one direction, the TSFM encodes the series in compressed form for the LLM, which reasons over it together with the text. In the other, the LLM's reasoning is fed back into the TSFM, which produces the forecast conditioned on it. One model architecture thus reads and generates both series and text, reasons in text, and draws on the TSFM's pretrained forecasting ability. To train it, we build MMTR, a multimodal time series reasoning dataset with annotated reasoning traces, covering both tasks across 12 domains with 500k+ samples. On MMTR, TimeOmni-o1 outperforms existing models on both understanding and forecasting across most domains, as well as the LLM and the TSFM inside it when used individually. The gains carry over to three open benchmarks, where it outperforms the dedicated time series models.
Recent advancements in multimodal diffusion language models (MDLMs) have exhibited strong global modeling capabilities through iterative refinement. We observe that mask uncertainty follows an overall downward trend during iterative denoising, while a few denoising steps exhibit localized entropy spikes closely associated with hallucinations. We argue that anomalous steps can be identified by measuring the entropy deviation of each denoising step from its neighboring temporal window. We hypothesize that discrete token commitment discards the distributional information at masked positions, limiting the continued interaction between candidate semantics and the context during entropy spikes. With this goal, we present StepRefine, a plug-and-play decoding strategy that leverages semantic context to achieve reliable denoising. The core of our method lies in a step-aware switch between discrete denoising and latent refinement. The model employs probability-weighted embeddings at entropy spike steps to preserve diverse reasoning hypotheses and performs multiple rounds of continuous refinement. Meanwhile, a refinement controller monitors distributional stability to reduce unnecessary loops and switch the model back to standard discrete denoising. Through extensive experiments, StepRefine demonstrates significant hallucination mitigation across different MDLMs on multimodal benchmarks.
Longitudinal social-media histories offer an opportunity to assess depression, potentially supporting the inference of standardized depression questionnaire responses. However, this task requires evidence that is semantically relevant to a symptom and attributable to the user, while being sufficiently specific to support an individual questionnaire item. Existing approaches rely on semantic similarity to retrieve posts, which can introduce misleading or non-attributable evidence, particularly in multi-symptom posts and discussions of others' experiences. We propose SREE (Symptom Retrieval and Evidence-Grounded InfErence), an evidence-grounded pipeline that extracts first-person and symptom-specific text spans, selects a representative subset using a Facility Location-inspired objective, and performs Evidence-Gated Inference over the selected evidence. We evaluate evidence retrieval on BDI-Sen and evidence-conditioned questionnaire inference on eRisk 2019-2021. Compared with state-of-the-art retrieval approaches, SREE achieves relative improvements of 70% in post-level precision and 59% in F1, while retaining at least 30% fewer posts. On multi-symptom sentences, it also outperforms dense retrieval in precision and F1 even with twice the retrieval budget. Across all three eRisk collections, SREE achieves the highest exact item-level agreement while remaining competitive on ordinal and aggregate metrics. These results highlight the importance of localizing attributable evidence and controlling its composition before inferring depression questionnaire responses.
Neural operators (NOs) have been introduced as deep learning-powered solutions to the computational bottleneck of solving families of problems. Once trained, an NO can rapidly output approximate solutions, avoiding the need to rerun classical solvers. However, standard NO architectures cannot guarantee static no-arbitrage when predicting European call-price surfaces from market or model features and, as we verify experimentally, can yield negative PnL in downstream trading tasks. We introduce a certifiably arbitrage-free (AF) and universal encoder-decoder NO framework, continuously mapping inputs from a reproducing kernel Hilbert space (RKHS) $\mathcal{H}$ over any fixed subset of $[0,\infty)^2$ to a convex combination of arbitrage-free European call-price surfaces by continuously solving an approximate nearest-neighbour problem. We show that a nearest arbitrage-free correction of an unconstrained predicted surface need not exist, implying that no-arbitrage cannot, in general, be restored through post-processing. Further, we show that standard NOs, such as the FNO and certain encoder-decoder NOs, admit parameter configurations whose outputs violate static no-arbitrage. Thus, out-of-the-box NOs generally cannot be certifiably restricted to arbitrage-free configurations, but may only be guided towards them through PINN-type regularization during training.
Satellite altimetry samples ice sheets densely along tracks but sparsely across tracks, with frequent gaps in spatial coverage. We study the impact of this irregular support on surface reconstruction and introduce a new method, D-TIN, that applies Delaunay triangulation and barycentric interpolation directly to observed locations. A spatial holdout removes contiguous track intervals and nearby same-pass observations. We compare D-TIN with two common workflows, kriging and tile-based polynomial interpolation, to evaluate its accuracy and computational efficiency. Across seven ICESat-1 campaigns, D-TIN reduces RMSE by 12.1 m to 23.6 m relative to independent 5 km tile and polynomial based reconstruction. Its aggregate accuracy is close to the ordinary kriging workflow using a 1 km grid on the commonly used GStatSim platform while its median runtime is 252 times lower. A fixed-weight multiscale extension of D-TIN, called MS-D-TIN, gives the lowest aggregate RMSE, 67.5 m, while retaining the low computational cost of the D-TIN family.
Sliced inference such as SAHI is the standard remedy for small objects in high-resolution imagery, but its gains are usually measured against low-resolution baselines. We revisit sliced inference with scale as the explicit variable and ask when tiling actually pays off. We compare seven dense detectors from four families on VisDrone and DOTA, under pre-specified hypotheses, calibration-only tuning and a single evaluation on held-out test data. Two findings emerge. First, much of the benefit attributed to slicing is a resolution effect: on VisDrone, a single pass trained at a higher resolution outperforms tiling with box fusion (merging decoded boxes) for every detector that could be trained at that resolution, and at equal compute a higher resolution helps more than a larger model. Second, tiles carry complementary information that box fusion wastes. We introduce SCAHI (Slicing with Coverage-Aware Hyper Inference), which accumulates the dense outputs of all passes with weights that give every image position the same total coverage, and reads objects off a peak grid fine enough to keep crowded neighbours apart. On identical passes, SCAHI outperforms box fusion for every YOLO detector on both datasets and matches it for FCOS. On DOTA, whose images far exceed the resolution a single pass can be trained at, tiles with SCAHI beat the best single pass for every detector, by 4.3 to 5.6 mAP50 at 2048 px, and replacing only SAHI's merge step by SCAHI adds 2.3 to 4.0 mAP50. An ablation attributes the gain to the border weighting and the resolution of the peak grid. Source code is available at https://github.com/beckerio/scahi.
This technical report presents MNN-Pocket-0.3B, a compact 300M-parameter dense language model designed for efficient on-device deployment and supporting context lengths of up to 32K tokens. With an INT4 deployment footprint of only 173 MB and decoding throughput of up to 466 tokens/s on iPad Pro M5 (Metal), it can be seamlessly integrated into applications on edge devices such as smartphones and smart glasses. MNN-Pocket-0.3B is developed through the co-design of its architecture, data, training pipeline, and inference system. Hardware-aware architecture selection and dependency-aware vocabulary pruning balance model capability with on-device memory and compute constraints. A four-stage pre-training pipeline covering approximately 4.1 trillion high-quality Chinese and English tokens equips the model with broad knowledge and reasoning capabilities. This is followed by a three-stage post-training pipeline comprising instruction fine-tuning, multi-teacher on-policy distillation, and quantization-aware training. Finally, hardware-specific optimizations in the MNN inference engine enable fast, memory-efficient inference and long-context processing. Extensive evaluations demonstrate that MNN-Pocket-0.3B achieves frontier-level performance among models of comparable size, with particularly strong results in Chinese and English knowledge and reasoning, mathematics, coding, and instruction following. It consistently outperforms similarly sized open-source models and even surpasses several 0.5B-parameter models on instruction-following benchmarks. We release MNN-Pocket-0.3B with open weights and ready-to-use MNN deployment packages. We hope it will provide a practical foundation for capable, responsive, and memory-efficient edge applications while advancing research and development in on-device language models.
Flow-based and diffusion-based models have become the principal means of learning and sampling high-dimensional probability distributions, from images and video to proteins, molecules and physical fields, and increasingly language. Each carries a simple law to the law of the data along a curve of probability measures, and behind the variety of their names lie three geometries, those of the state space on which the data live, of the space of probability measures through which the curve passes, and of the space of paths traced by the generating process. A single analytic idea runs through the subject. Every score identity used to train a diffusion model is an integration by parts, on the noise, on the data or on Wiener space, and Malliavin calculus extends this to nonlinear and infinite-dimensional diffusions. We develop these foundations with full definitions and follow them through normalising flows, flow matching and few-step maps, diffusion models and their convergence theory, conditioning, Schr\"odinger bridges and stochastic control, and their extension to manifolds, Lie groups, discrete and function spaces and jump processes. We close with statistical physics, renormalisation and, more speculatively, holography, in which the radial direction of anti-de Sitter space can play the role of flow time. A table of the main methods and a collection of open problems complete the survey.
Machine learning (ML) services increasingly expose explainability outputs alongside model predictions. While such explanations are intended to improve transparency, they may also reveal information about the underlying training data. In this work, we show that the local surrogate geometry employed by Local Interpretable Model-agnostic Explanations (LIME) can be repurposed as a proxy for a sample's distance to the model's decision boundary. We extend this boundary-distance view to tabular counterfactual explanations and, for the first time, to visual counterfactual explanations for image classifiers. We present XposedMIA, a one-shot, explanation-based black-box membership inference attack (MIA) framework that converts the returned explanation into a proxy for the queried sample's distance to the model decision boundary. After a one-time calibration step on a small labeled calibration set of as few as 30 samples, our attack requires only a single query for each target sample. Unlike prior label-only attacks requiring iterative queries or shadow training, our method relies solely on the predicted label and its accompanying explanation. Our evaluation indicates that XposedMIA achieves high effectiveness, with success rates of up to 0.95 using only a single query per target, revealing that releasing explanations introduces an additional channel for membership leakage.
Exploration in deep reinforcement learning (RL) for continuous control is commonly driven by temporally uncorrelated Gaussian (white) action noise, which leads to jerky motions with locally constrained exploration. In contrast, temporally correlated coloured noise, particularly pink noise, has been shown to support more efficient exploration in RL. Motivated by these findings, we inquire whether scheduling changes to the noise colour over the course of training can outperform white noise for continuous control robotics. We perform a systematic study of 49 linear schedules between white and red noise, together with generalised state-dependent exploration (gSDE), on four sparse-reward manipulation tasks with a Fetch arm and a Shadow Dexterous Hand. We find that the noise colour at the end of training does not affect learning efficiency or end performance, whereas fading towards red noise halves the jerk of exploration in every task. Balancing learning efficiency, end performance and smoothness, red noise emerges as the preferred strategy: it learns as efficiently as pink noise with almost 2-fold less movement jerk. We therefore recommend exploration schedulers that fade towards red noise as the default for off-policy robot learning.
Sensory delays create temporal misalignments between the state of the environment and the observations available to an agent, turning Markovian control problems into partially observable ones. In this work, we inquire whether such misalignments can be mitigated during training by using hindsight to recover aligned state-action transitions. We introduce Hindsight Critics, an asymmetric extension of the Soft Actor-Critic for reinforcement learning under partial observability. The critics are trained on hindsight-aligned transitions whereas the actor learns from the resulting privileged value estimates despite only having access to delayed observations. Comparisons against delay-aware and unaware baselines on the DelayBench suite show that Hindsight Critics can partially compensate for observation delays without requiring modifications to the agent's observations during deployment.
When a vision-language model encounters a new class or domain, how much should it trust its learned prompts? Because source adaptation can correct pretrained errors on some target images while disturbing valid predictions on others, we introduce Contextual Adaptive Fusion (CAF), which retains four prediction branches by pairing frozen and prompted visual and textual representations. To arbitrate when branches disagree on an image, CAF combines each branch's predictive confidence with how frequently the branches favor its proposed class across the unlabeled target collection, placing greater trust in predictions that are locally confident yet less dominated by collection-wide frequency bias. The fusion coefficients are learned once on source data and remain fixed at inference. Trained on ImageNet, CAF achieves 69.96% average accuracy across ten cross-dataset benchmarks (1.92 percentage points above ManiPT) and remains highly competitive across four ImageNet variants, setting a new best of 52.21% on ImageNet-S. In base-to-novel transfer across eleven datasets, CAF further reaches 79.77% novel-class accuracy and an 82.81% harmonic mean. Code is available at https://github.com/zhenyuanchenai/CAF.
Alignment of text-to-image diffusion models with reward functions is a key challenge in controllable generation. Existing approaches typically rely on optimizing model inputs or fine-tuning denoising networks, which requires computationally expensive backpropagation through large denoising models. In this work, we propose Direct Trajectory Optimization (DTO), a training-free framework that performs alignment by directly optimizing the sampling trajectory of diffusion models. Specifically, DTO treats intermediate denoising scores and sampled noise as the optimization variables, eliminating the need for gradients through the denoising network. To mitigate the out-of-distribution problem, we introduce Reward-Optimized Score Guidance, which interpolates between the optimized scores and the pretrained model predictions to stabilize sampling, while providing a provable theoretical guarantee of reward improvement under constrained distributions. Empirically, DTO improves alignment performance across three diffusion models on two benchmarks, outperforming existing inference-time alignment methods while maintaining high efficiency. These results suggest that trajectory-level optimization is a promising direction for efficient and effective diffusion model alignment.
In financial backtesting and other point-in-time tasks, external inputs can be restricted by date, but model parameters may retain financial facts disclosed later, introducing parametric look-ahead bias. We introduce FinCutBench, a continually updated monthly financial knowledge benchmark for measuring how far a fixed model's financial knowledge extends over calendar time. The current dataset contains 9,000 Chinese and English four-option multiple-choice questions from January 2023 to September 2026, covering Chinese and U.S. markets and four categories of financial information. We construct monthly knowledge profiles using the approximate months in which the tested facts became public. After a profile passes the decline test and knot checks, we define the FinCut Boundary as the first month when the fitted profile reaches its lower level, and report the stability of the estimated date under resampling. Boundary dates and localization precision vary across models. Using the independently measured Transition Start and Boundary as fixed references for downstream sample grouping, we find significant relative declines in U.S. asset-ranking performance and in earnings recall for Sol and Luna. Downstream effects vary across markets and tasks. The FinCut Boundary thus provides an empirical reference for the temporal scope of parametric knowledge in point-in-time financial evaluation, supporting evaluation-window selection and result interpretation.