Humans cannot always intuit what scenarios are most challenging to LLMs. Hoping to capture challenging edge cases, developers either design problems to be difficult for humans or curate extensive benchmarks. What if we could instead anticipate which scenarios a model will fail on? In this paper, we use an LLM's representational geometry to predict which concept combinations it will fail on. We attribute this compositional failure to interference between salient features. In tasks that require systematic composition - toy programmatic settings, multihop reasoning, multilingual factual recall - we find that when a pair of concepts is encoded near-orthogonally, the model reliably composes them. When their linear encodings are close, producing interference, the model fails to compose them. Our method reliably anticipates failure modes across different compositional tasks, without evaluating specific inputs. These results lay the groundwork to use representational geometry to identify high-risk examples, construct targeted stress tests, and provide a scalable foundation for active learning in real-world deployment.

Lawyer-client consultation is a critical starting point for legal services. Effective legal assistance hinges on eliciting sufficient and truthful information from clients in order to devise strategies that best protect their interests. This task requires Large Language Models (LLMs) not only to perform robust legal reasoning, but also to strategically elicit material facts through multi-turn interactions and effectively guide clients with diverse personalities. Yet existing legal benchmarks overlook this interactive capability. To fill this gap, we introduce DLawBench, a diagnostic benchmark for real-world legal consultation. Drawing on realistic client behavior, we characterize lawyer-client interactions into four types: Cooperative, Dependent, Withdrawn, and Adversarial. Using dialogues grounded in real cases, DLawBench evaluates whether LLMs can effectively conduct legal consultation under realistic conditions. DLawBench comprises 461 cases from Chinese and U.S. law, 5,532 paired fact entries, 3,411 inquiry rubrics, and 3,348 issue-resolution rubrics, and evaluates 26 representative LLMs. Systematic experiments show substantial headroom: the best-performing model, GPT-5.5, achieves only 0.562 on consultation-grounded legal reasoning. More importantly, DLawBench exposes both sycophancy in legal consultation and a paradox: models perform worse when clients need guidance most.

Vision-language models (VLMs) are typically trained as passive answerers, while their ability to actively ask diverse, non-trivial, visual-centric and grounded questions remains underexplored. Existing visual questioners' performance is bottlenecked by the availability of high-quality training data or the cost of curating them. We show that a VLM can continuously improve itself as a visual questioner without any external supervision. We propose a self-evolving framework that uses a VLM itself as both a proposer and a filter to produce harder, more informative, and visual-centric questions, while maintaining their exploration diversity to avoid training collapse. These questions are then used to train the VLM in both questioner and answerer modes. To evaluate the questioner, we introduce an agentic protocol that assesses questions along perception, reasoning, and diversity dimensions. Experiments across various backbone VLMs show that our method substantially enhances the quality and substantially expands the difficulty boundary of autonomous question generation. Under the same budget, our self-supervision is more effective than training on the static source data. Moreover, the self-evolving questioner remains a competitive or even better answerer.

Large language models can often close proof gaps in interactive theorem provers, but a verified theorem is not the same thing as a reusable library contribution. We study this distinction through a detailed case study: a semi-autonomous formalization of Grothendieck's vanishing theorem. The initial version compiles with no sorries, but an expert review found serious problems in definitions, theorem generality, file organization, and the API. We then ran a review-driven refactor and compression process and obtained a second expert review. The before-and-after comparison shows a sharp split: agents adapted well to local, mechanically checkable feedback, but remained weak at choosing definitions and designing APIs. We argue that autoformalization should be evaluated not only by closed sorries, but by whether the resulting formalization survives expert review.

Each year, college admissions offices face an overwhelming challenge: processing millions of high school transcripts, each with unique formats, grading systems, and layouts. This manual process creates operational bottlenecks that delay admissions decisions and consume valuable resources. We present a transformative solution through a multi-agent AI system where specialized agents collaborate to automatically process diverse transcript formats through intelligent coordination and communication. Our multi-agent architecture consists of three specialized agents-a Pattern Recognition Agent for format-specific parsing, a Semantic Analysis Agent for natural language understanding, and a Vision Intelligence Agent for multimodal document analysis-coordinated by an Orchestration Agent that manages agent communication and result reconciliation. Our key innovation lies in agent-based quality control using GPA extraction as a coordination signal, ensuring reliable agent collaboration and preventing critical information loss. When evaluated on 40 real world transcripts from high schools across 13 U.S. states, our agent system successfully processed every document, achieving 96.7% accuracy compared to expert manual review while maintaining practical processing speeds of 45 seconds per transcript. This work demonstrates how multi-agent coordination can solve complex document processing challenges, offering institutions a scalable, collaborative AI solution that preserves accuracy while dramatically reducing processing time.

Electromagnetic Navigation Systems (eMNS) have gained considerable attention for minimally invasive surgery and targeted drug delivery. While most of the literature relies on quasi-static control of these systems, recent work has demonstrated the benefits of dynamic approaches. However, trajectory tracking far from equilibrium states remains largely unaddressed. We close this gap by demonstrating the first swing-up of a magnetically actuated inverted pendulum using the clinically-ready Navion eMNS. Although the inverted pendulum is not clinically relevant in itself, the proposed method utilizes torques and forces as control objectives, making it applicable to other magnetically actuated devices such as catheters and guidewires. Our approach combines trajectory optimization that accounts for internal eMNS dynamics with time-varying Linear Quadratic Regulator (LQR) state feedback and Iterative Learning Control (ILC), which leverages previous trial data and the system's dynamic model to progressively refine the feedforward command. While LQR alone fails due to the complex phenomena of magnetic actuation, ILC enables successful swing-up within six iterations. Furthermore, post-experimental analysis reveals that the learned ILC correction closely matches the torque discrepancy predicted by high-fidelity magnetic field model calibration, suggesting learning and adaptation as a promising tool to deal with uncertainties in electromagnetic actuation arising, e.g., from patient-specific physiological motion patterns and field model calibration inaccuracies.

Automated aerial wildlife surveys increasingly rely on deep learning, yet standard object detectors require bounding-box annotations, reported to be up to seven times slower and three times more expensive to produce than point-level labels. To address this bottleneck, we introduce the Overhead Wildlife Locator (OWL), a weakly supervised density-estimation framework with three variants: OWL-C, a fully convolutional model for high-throughput screening; OWL-T, a Swin-augmented hybrid for heterogeneous, cluttered scenes; and OWL-D, built on a frozen DINOv3 ViT-H+/16 encoder with a DPT-style fusion decoder. We benchmark all three against POLO, YOLOv11n, and YOLOv11l across five public aerial datasets, from sparse fixed-wing savanna surveys to dense UAV paddock imagery, and against the published HerdNet baseline on its native Delplanque split. OWL-D sets a new state of the art on Delplanque (0.934 AP vs. HerdNet's 0.840) and records the highest AP on four of the five datasets. Performance is regime-dependent: on the extreme-density SheepCounter UAV dataset the hybrid OWL-T leads (0.978 AP) and the convolutional variants attain the lowest counting error, whereas the foundation-based OWL-D degrades, indicating which variant suits which survey type. We further validate operational readiness on the Alaska Department of Fish and Game's 2022 Central Arctic Caribou census: under cross-herd and cross-temporal transfer, OWL-C fine-tuned on the 2017 Porcupine Caribou Herd split attains F1 = 0.965 on a held-out patch test set, with a signed count error of +3.1% aggregated across the released test patches. We release the OWL code, model weights, and the annotated Porcupine Caribou Herd 2017 (PCH) and Central Arctic Herd 2022 (CAH) patches, the first open patch-level datasets for large-scale caribou aerial surveys, at https://github.com/microsoft/MegaDetector-Overhead.

Autonomous staff-free public transport requires reliable in-vehicle passenger monitoring. However, perception inside moving vehicles is challenged by confined spaces, variable illumination, motion-induced background variation, occlusion, and limited viewpoints. To mitigate these spatial constraints, ceiling-mounted fisheye cameras provide full-scene coverage from a single viewpoint. Yet existing public overhead fisheye datasets are recorded in static environments and do not capture the domain shift introduced by vehicle motion. To fill this gap, we introduce PMOF, Passenger Monitoring using Overhead Fisheye cameras, the first public dataset of top-view fisheye imagery captured inside a moving vehicle, comprising over 19k manually annotated frames. PMOF provides rotated bounding boxes, tracking identifiers, and action labels, supporting object detection, tracking, and action recognition. We benchmark PMOF using YOLO26m-obb models fine-tuned under multiple dataset configurations that combine PMOF with existing overhead fisheye datasets. Cross-domain fine-tuning with custom rotation-aware augmentation achieves 94.8% AP50 on PMOF and 96.5% AP50 on an unseen overhead fisheye dataset from a different domain. Our results highlight the domain gap between static and moving environments and show that incorporating PMOF improves detection performance and advances generalization beyond passenger monitoring to broader fisheye-based person detection tasks. The dataset and code are available at https://swermuth.github.io/pmof/.

Exploratory question answering (EQA) over data lakes requires an LLM agent to discover relevant sources, analyze retrieved data, and adapt its actions based on intermediate results. End-to-end accuracy alone cannot distinguish failures in search, planning, data analysis, or the agent's Action Policy: its decisions about what to do next and when to submit an answer. We present SANA (Search Agent Navigation Ablation framework), a diagnostic ablation framework that transforms EQA tasks into runtime profiles containing gold source sequence, sanitized subquestions, and execution records. SANA uses these profiles to construct idealized search, planning, and data-analysis tools, allowing each component to be ablated; the residual gap is diagnostic evidence for policy failures. To illustrate SANA as a reusable evaluation framework, we adapted two recent EQA benchmarks, LakeQA and KramaBench, and evaluated lightweight and mid-sized agents under fixed prompts, budgets, data lakes, and runtimes. Across both benchmarks, data analysis is a consistent bottleneck while planning is less so. Search is a major limitation in LakeQA's large data-lake setting, but less so for the smaller-scale KramaBench. SANA thus deconstructs end-to-end task accuracies into a diagnosis of where data-lake agents fail, and allows for systematic comparisons of progress in search, planning, data analysis, and agent design.

Spiking Neural Networks (SNNs) have emerged as an energy-efficient alternative to conventional neural networks, demonstrating strong performance in computer vision and robotics. More recently, SNNs have been applied to time series forecasting (TSF), with methods exploring spiking temporal backbones, spike-compatible positional encodings, Fourier-domain processing, and redesigned neuron dynamics. However, existing SNN forecasting approaches process variables independently, lacking explicit mechanisms for modeling inter-variable dependencies. This is a critical limitation in multivariate settings, where cross-variable correlations carry substantial predictive information. We propose Spiking Fourier Graph Operators (SpikF-GO), which addresses this gap by combining a hypervariate graph formulation in which every scalar observation becomes a graph node with spike-driven spectral processing. SpikF-GO introduces a Hard Concrete frequency gate for learnable sparse frequency selection and a Complex LIF gate that applies independent spiking neurons to real and imaginary Fourier components, preserving binary, event-driven computation throughout the spectral domain. We further present a variant incorporating Central Pattern Generator-based positional encodings for stronger long-range temporal modeling. Evaluated on eight benchmarks under a unified experimental protocol, SpikF-GO achieves the best average rank among all SNN methods and outperforms its ANN counterpart, FourierGNN, at reduced energy cost. SpikF-GO maintains competitive accuracy even at substantially smaller embedding dimensions, thereby achieving significant energy reductions. To our knowledge, this is among the first works to bring graph-based multivariate modeling into the spiking domain for TSF and the first to provide a unified comparison across SNN forecasting architectures under a common experimental protocol.

Creative image editing tools, such as Photoshop's Remove or Generative Fill buttons, are central to everyday customer use and account for a major share of traffic in Photoshop and Lightroom. However, current generative AI models face significant latency challenges, which become even more pronounced when transitioning from convolution-based U-Nets to Diffusion Transformers (DiTs). In our evaluation on hundreds of representative image editing samples spanning a wide range of mask ratios, the DiT module alone accounts for an average of 73% of the total model latency, even after being distilled from 50 timesteps down to 8 timesteps. To tackle this challenge, we propose $\textbf{HiLo-Token}$, an input-adaptive token compression framework that allocates more token budget to high-frequency, rich-context regions while assigning fewer tokens to low-frequency areas. Specifically, for the editing region specified by the user mask, we retain all tokens within a dilated mask to preserve strong locality and contextual relevance. Outside the editing region, we introduce a simple yet effective high-frequency token selection strategy based on spatial frequency to capture important local details, while using tokens from a 16x downsampled image to represent low-frequency components and preserve the blurry but global structure. Extensive experiments on production-level evaluation data validate the effectiveness of the proposed method, achieving 3.13x, 2.59x, and 1.67x DiT speedups on A100-80GB for image editing tasks across small, medium, and large mask ratio categories with average ratios of 6.38%, 15.92%, and 35.36%, respectively, without any regression in generation quality.

Self-supervised geospatial foundation models (GeoFMs) learn transferable representations from remote sensing data, but their downstream behavior is difficult to characterize. We study six representative GeoFMs spanning joint-embedding, reconstruction, and multimodal pretraining families, and evaluate transfer across classification, regression, and segmentation benchmarks under different label availability and downstream pipelines. We find that model rankings change across tasks and adaptation settings. Layerwise probing shows that, in most cases, task-relevant information is more accessible in intermediate transformer blocks compared to final-layer embeddings, and that GeoFMs exhibit distinct depthwise profiles. In segmentation case studies on PASTIS and Sen1Floods11, downstream adaptation settings such as decoder design and fine-tuning can be as impactful as the choice of GeoFM, and standard dense-prediction heads may be poorly aligned with how GeoFMs organize information over depth. Finally, CKA analysis on case studies shows that fine-tuning does not rewrite GeoFMs uniformly across depth, and the strongest changes are localized to the first linear layer of the MLP in ViT blocks. These results help explain why GeoFM rankings shift across benchmarks and motivate more representation-aware evaluation and adaptation strategies.

AdamW is a default optimizer for modern deep learning, but its first and second moment states add roughly two parameter-sized buffers to training memory. We propose Gefen, a memory-efficient optimizer that automatically shares second-moment estimates across parameter blocks and quantizes the first moment using a learned codebook, thereby reducing AdamW's memory footprint by ~8x while maintaining the same performance, corresponding to a reduction of 6.5 GiB per billion parameters. The method is motivated by a theoretical result showing that large mixed Hessian entries constrain the ratio of squared gradients toward one, suggesting that Hessian-aligned parameters are natural candidates for sharing second-moment statistics. Since computing Hessians is impractical at scale, Gefen infers block structure from the initial squared gradients, requiring no architecture-specific metadata or hyperparameters beyond AdamW defaults. Gefen learns an exact histogram-based dynamic-programming quantization codebook and reuses the same blocks for first-moment scaling. Across diverse experiments, Gefen achieves the lowest peak optimizer memory among the compared AdamW-like methods while maintaining AdamW-level performance. In FSDP and DDP training, the reduced memory footprint enables larger microbatches and improves throughput significantly over AdamW, providing a practical drop-in replacement with lower memory usage that can increase throughput and enable training larger models or using larger batch sizes. We provide the complete Python implementation, including fused CUDA kernels at https://github.com/ndvbd/Gefen

Vision-Language-Action (VLA) models excel at mapping visual inputs and natural language instructions directly to robotic control policies. However, because they are trained primarily to fit behavioural demonstration data, they do not explicitly enforce fundamental physical principles such as rigid-body dynamics or contact constraints. This exposes a critical physics gap: standard temporal smoothing applied on top of single-step or chunked VLAs trades trajectory quality for added failures that short-term memory cannot resolve. To bridge this gap, we introduce PhysVLA (Physics-VLA), a plug-and-play, inference-time framework designed to wrap any frozen VLA backbone without retraining, fine-tuning, or weight access, with less than 1 ms of overhead per control step. PhysVLA intercepts the predicted control action, captures only the simulator or system state, and applies a dual-layered correction: (i) a phase-aware finite-state machine that structures discrete task segments (approach, grasp, transport, and place), and (ii) a selective Euler-Lagrange gate that activates only when a dynamics oracle detects kinodynamic inconsistency. Evaluated across OpenVLA, OpenVLA-OFT, Force-VLA, and Generalist-VLA on LIBERO-Spatial with a 7-DoF Franka Panda, the framework delivers absolute success rate increases of up to 17% and stability increases of up to 19% with no per-task regressions, improves trajectory efficiency by up to 15% across all four backbones, and shows up to a 10x improvement in trajectory jerk robustness on a Robosuite Lift cross-simulator sweep. We further validate the framework on a real Agilex Piper arm with a pick-and-place task, confirming that PhysVLA transfers to physical hardware without retraining, with success-rate improvements of up to 50%, establishing physical awareness as a composable, backbone-agnostic runtime module.

Autonomous UAV navigation is conventionally solved by pipelines that separate perception, mapping, and planning into distinct stages, which propagates errors, accumulates latency, and requires environment-specific retuning. End-to-end generative models remove these interfaces by mapping raw observations directly to trajectories, but inherit a subtle failure mode: trained on clean data, they cannot recognise when an observation is unreliable, and treat degraded regions such as glass, mirrors, and overexposed surfaces as valid evidence for planning. We present a reliability-aware diffusion planner for 3D UAV navigation. It conditions trajectory generation on the observation together with a scene-level reliability heatmap that marks where perception cannot be trusted, produced by a lightweight network that distils the open-vocabulary reasoning of a vision-language model within the real-time planning budget. To generalise to unseen environments without retraining, we steer the denoising process with a differentiable two-stage ESDF cost that treats physical obstacles from depth and virtual obstacles from highly unreliable regions on equal footing. In simulation and on a real quadrotor, our planner produces markedly safer trajectories than a state-of-the-art diffusion baseline, reducing the obstacle-violation rate from 40.3% to 9.6% and raising the mean reliability of traversed regions from 0.588 to 0.925. Ablating the reliability term alone drops mean reliability from 0.898 to 0.783, confirming it as the decisive component, while distillation runs the framework up to 2 times faster than the full vision-language model.

Accurate estimation of long-term care transition probabilities is central to disability insurance pricing, reserving, and solvency assessment. Classical actuarial multi-state models commonly rely on Markov, semi-Markov, or proportional-hazard specifications, which provide a direct connection to cohort projection but may be restrictive for irregular longitudinal health data with nonlinear aging patterns and heterogeneous covariate histories. This paper develops a well-calibrated estimator of multi-state transition probabilities for irregular longitudinal health data. The model learns from individual health history, incorporates the time elapsed between observations, and conditions transition probabilities on demographic and socioeconomic attributes. It produces a valid probability distribution over the next observed health state, with four possible states: healthy, mild disability, severe disability, and death. Individual probabilities are aggregated by age group and origin state to form transition matrices compatible with actuarial cohort projection. Using longitudinal data from the Health and Retirement Study, we compare the proposed estimator with logistic regression, gradient-boosted trees, a recurrent neural network, and a last-state persistence benchmark. The evaluation considers probabilistic accuracy, endpoint discrimination and calibration for severe disability and death, risk concentration, and transition matrix error after aggregation. The proposed estimator improves severe disability discrimination relative to logistic regression and gradient-boosted tree benchmarks, maintains strong calibration, and yields the lowest transition matrix error among the evaluated models in the held-out test analysis. Results show that a structured machine learning estimator can support long-term care transition modeling when judged by calibration and projection fidelity, beyond discrimination.

End-to-end navigation policies trained on large simulation corpora degrade sharply when transferred to out-of-distribution scenes, categories, or goal modalities. Modular pipelines such as Modular GOAT are bottlenecked by closed-set object detection recall, while 3D snapshot-memory systems (e.g. 3D-Mem) accumulate dense, view-dependent representations that are heavy to maintain. We present AnyGoal, a training-free multi-robot architecture that places a Vision-Language Model (VLM) at the core of frontier-based exploration and coordinates agents through a shared 2D Gaussian Bayesian Value Map (BVM). The BVM maintains a per-pixel (mu, sigma^2) posterior over goal relevance, updated via precision-weighted fusion of VLM scores through a depth-cone mask, and is never reset between subtasks, yielding lifelong evidence accumulation. Frontiers are ranked by a convex blend of a VLM-as-judge softmax and a Bayesian UCB term on the BVM. A greedy allocator with spatial-separation penalty and commitment hysteresis distributes frontiers across agents without a centralized controller. On the full GOAT-Bench val unseen split (360 episodes, 2,669 subtasks), our dual-agent system achieves 52.4% Subtask SR at 12.7% SPL--state of the art under the strict physical regime (discrete 0.25 m steps, no teleportation, 42 deg HFOV) and a +27.5 pp improvement over Modular GOAT (24.9%). Single-agent AnyGoal achieves 41.9% Subtask SR, showing gains arise from the decision architecture. A four-way perception ablation shows that open-vocabulary detectors shift the dominant failure mode from exploration to goal verification.

Contact-rich manipulation requires world models to reason over complex contact dynamics from multimodal sensory observations. However, it remains unclear which representation properties fundamentally support stable long-horizon planning in contact-rich settings. In this paper, we present ContactWorld, a benchmark and systematic empirical study of vision-tactile world models spanning 12 contact-rich manipulation tasks, including insertion, disassembly, screwing, and exploratory interaction. Across extensive experiments, we find that representations that are both spatially structured and temporally continuous consistently achieve the strongest planning performance. In particular, point-cloud observations improve average planning success rates from 20.7% with wrist-view observations and 22.0% with front-view observations to 32.1%. We further find that the effectiveness of tactile sensing depends critically on cross-modal representation compatibility rather than modality scaling alone. Combining point-cloud observations with tactile force-field representations, which preserve richer spatial structure and interaction dynamics, further improves performance to 36.1%, yielding the strongest overall planning performance across all evaluated tasks. Moreover, tactile sensing becomes increasingly important under long-horizon planning objectives, where compounding prediction errors and contact uncertainty accumulate over time. Together, these findings highlight the importance of representation structure, multimodal compatibility, and long-horizon robustness in vision-tactile world models for contact-rich robotic manipulation.

Unlearning aims to remove the influence of specific training data sources, but this has proved challenging because the contributions of different sources are entangled within the model. Isolating source contributions to disjoint parameters makes removal easier, though it obstructs joint learning across sources. We propose NULLs (Natively Unlearnable LLMs), a model class that satisfies the two opposing goals of isolating source-specific contributions and learning jointly across sources, by training a set of shared backbone neurons alongside a pool of sparsely activated sinks. During training, information specific to a source naturally concentrates in its sinks while information shared across sources accumulates in the backbone. A source is then unlearned at deployment by disabling its corresponding sinks, with no gradient updates and no access to the retained data. We show that NULLs scales to Wikipedia's ~6M articles, isolating each as an independent source. Unlearning a single article removes knowledge specific to it while preserving facts shared with semantically related articles, closely matching retraining from scratch. We note that unlearning with NULLs is also robust: in a case study of unlearning the Harry Potter books, NULLs resists both adversarial extraction and relearning that reverses post-hoc unlearning. Finally, NULLs preserves general language capabilities, matching a standard transformer on downstream benchmarks. Together, these results suggest that source-level unlearning need not be an afterthought. It can be built natively into LLM training while retaining the benefits of shared representation learning.

Generating avatar videos that are not merely visually similar to a target individual but behaviorally recognizable, faithfully reproducing their talking rhythm, gestural tendencies, and expression dynamics, remains an open challenge. Existing methods predominantly condition on single static images, which provide insufficient identity information and cannot capture dynamic motion traits, while standard pixel-level objectives underserve the perceptually critical facial regions that determine avatar fidelity. We present Avatar V, a production-scale framework that addresses these limitations through video-reference-conditioned identity modeling. Rather than compressing identity into fixed-size embeddings, the model conditions directly on the full token sequence of a reference video, learning to reproduce both static identity attributes (facial geometry, skin texture) and dynamic behavioral patterns (talking rhythm, micro-expressions) through attention over the reference context. We introduce Sparse Reference Attention, an asymmetric mechanism achieving linear-complexity conditioning on arbitrarily long references; a motion representation stream enabling closed-loop talking style transfer; and an identity-aware super-resolution refiner inheriting the full reference conditioning. These are supported by a data engine curating 100M+ training clips from 50M raw videos, and a five-stage training pipeline with flow matching pre-training, personality fine-tuning, two-phase distillation (>10x acceleration), and RLHF alignment, deployed across thousands of GPUs. Avatar V generates 1080p videos of unlimited duration, achieving state-of-the-art identity preservation, lip synchronization, and generation quality on our cross-scene benchmark, consistently outperforming leading systems including Seedance 2.0, Kling O3 Pro, Veo 3.1, and OmniHuman 1.5 in both automated metrics and human evaluation.

Tabular data embeddings have become a cornerstone of data profiling and data integration pipelines, enabling tasks such as entity annotation and resolution; schema matching; column type detection; and table search, among others. Existing approaches embed rows, columns, or entire tables into a vector space and rely on nearest-neighbor search to retrieve candidate matches. A fundamental limitation of current embedding methods is the lack of interpretable similarity scores: the concrete similarity value between a query and its nearest neighbour carries no intrinsic meaning, making it impossible to determine whether that neighbour is a true match or simply the least-dissimilar item in a corpus that contains no valid answer. This inability to set principled thresholds for retrieval undermines practical deployment, particularly for zero-match detection. We investigate the use of HyperDimensional Computing (HDC), specifically the Holographic Reduced Representations (HRR) model, as a framework for tabular row embeddings when the retrieval task corresponds to answering structured select-project queries in vector space. Exploiting the algebraic properties of HDC operations, we derive closed-form expected similarity values for both equality and non-equality retrieval predicates, which converge to interpretable values as dimensionality increases, and use these to identify suitable retrieval thresholds. We evaluate HDC against EmbDI, a graph-based baseline, on two real-world datasets across varying table sizes and predicate lengths. Our results show that HDC matches or outperforms EmbDI for row retrieval across all configurations, handles non-equality predicates more robustly, and achieves perfect attribute projection accuracy at sufficient dimensionality -- while uniquely enabling reliable identification of zero-match predicates through its principled thresholds.

Vision-language models (VLMs) can answer image-based questions confidently, and often correctly, even when no image is provided. This mirage behavior inflates benchmark scores without reflecting visual grounding. Prior work treats this as a single failure mode. We argue it is two. Using Mirage Probes, a contrastive probing framework that pairs paraphrased question variants with matched mirage and non-mirage labels on the same image, we show that mirage behavior is linearly decodable from internal activations across residual stream, MLP, post-attention, and attention-head sites in two open-source VLMs. We demonstrate that a Naive Bayes text baseline cannot recover this signal, ruling out surface lexical confounds. Cross-benchmark separability patterns, together with a novel Prior Harnessing Index (PHI) measuring how much a model can answer from text alone, expose two distinct regimes: textual biases, where the model answers from language priors without engaging visual representations, and spurious images, where it constructs false visual content in latent space and answers as if grounded. The distinction has direct mitigation consequences: text-distribution cleaning can address the first regime but cannot reach the second, since spurious-image mirages live in the model's visual representations rather than its text. Faithful visual grounding will require interventions at the representational level.

Muon is an increasingly widely used optimizer that replaces a gradient $G=USV^\top$ with its polar factor $UV^\top$, thereby flattening the singular spectrum. However, full flattening discards singular-value information that may matter for adaptation. We introduce Muon$^p$, a Muon-style optimizer that instead uses fractional spectral-power updates $US^pV^\top$ for rational $p\in(0,1)$, interpolating between Muon and gradient descent. To make it practical, we prove that fractional spectral powers cannot be computed by any fixed univariate polynomial iteration, and furthermore derive low-degree odd bivariate recurrences that approximate $US^pV^\top$ using only matrix multiplications, preserving Muon's matrix-multiplication-only structure and compute complexity. We show that Muon$^p$ maximizes the linear improvement in loss under the Schatten $q$-norm for $q=1+\frac{1}{p}$. Empirically, Muon$^p$ is especially effective for finetuning: on billion-scale models, Muon$^p$ improves validation perplexity and downstream task performance. We further analyze when Muon$^p$ is less suitable, through the lens of spectral geometry. Our results reveal important insights on when preserving the singular spectrum can bring significant gains, and introduce a principled way to achieve them.

Long Chain-of-Thought (CoT) reasoning improves LLM problem-solving but is computationally expensive due to sequential token generation. While recent works explore reasoning in continuous latent spaces to bypass discrete token generation, they often struggle with training stability and fail to scale to complex, long-horizon tasks due to lack of supervision signal. We propose SuperThoughts, which compresses pairs of consecutive CoT tokens into single latent representations and decodes two tokens per step via a lightweight Multi-Token Prediction (MTP) module. This preserves discrete token supervision at training time while doubling throughput at inference time. We finetune Qwen2.5-Math-1.5B-Instruct, Qwen2.5-Math-7B-Instruct, Qwen2.5-Math-14B-Instruct, and evaluate on MATH500, AMC, OlympiadBench, and GPQA-Diamond. With a confidence-based adaptive mechanism that falls back to standard decoding when uncertain, SuperThoughts achieves $\sim$20--30\% CoT length reduction while maintaining accuracy with minimal degradation (1-2 points accuracy drop on most tasks).

While test-time scaling has revolutionized reasoning in large language models, generative video reasoning remains bottlenecked by a single-shot paradigm. We demonstrate that searching over denoising steps cannot rescue logically flawed rollouts because spatial trajectories commit early in the diffusion process. Root-level Best-of-N (BoN) sampling is similarly inefficient: reasoning errors cluster early in the temporal axis, and resampling blindly discards verified upstream progress. To unlock effective test-time scaling for video models, we introduce Temporal Backtracking Search (TBS), which shifts the search space to the temporal axis. TBS transforms video generation into an iterative generate-verify-restart loop via three core mechanisms: (1) variable-K conditioning to resume generation from arbitrary clean prefixes; (2) temporal process verification to localize failures and extract valid restart anchors; and (3) prefix-based search to reallocate compute toward extending correct trajectories rather than root resampling. Across algorithmic, navigation, and robotics domains, TBS Pareto-dominates matched-budget BoN. In a strict out-of-distribution setting where one-shot generation collapses (0.7% for BoN), TBS achieves 22.7%, with every solved episode stemming from a restarted branch. Ultimately, TBS reveals that the local reasoning competence of video models far exceeds what single-shot rollouts indicate, providing a scalable test-time framework to unlock it.

Fine-tuning a vision-language-action model (VLA-JEPA) on a single GPU should be simple: load a pretrained checkpoint, run training, deploy. There is a hidden danger. Run the same fine-tuning code thirteen times -- same data, same architecture, different random seed -- and twelve runs produce a robot succeeding 91--94% of the time, while one run silently degrades to 65.2%: a 29 pp gap with no error message, no warning, and no way to predict which seed will fail. We call this the seed lottery. We trace the cause to output collapse: the action predictor quietly learns to produce nearly identical outputs regardless of what the robot sees. Existing weight-level methods (L2, EWC) are structurally blind to this collapse -- they penalize weight changes, but collapse occurs in directions weights can move freely without affecting outputs, a gap we formalize via the Jacobian null-space. Across 7 methods x up to 13 seeds x 3 LIBERO benchmarks, three output-level regularizers -- VICReg (n=12 seeds), Dropout (n=4), and a halved learning rate (n=5) -- each eliminate every catastrophic seed (0/21 combined collapses vs. 1/13 Baseline; F(12,11)=28.7, p<0.001), while weight-level methods (L2, EWC) preserve the lottery. The simplest fix is changing one number in your optimizer config.

Neural topic models enable scalable semantic discovery, but their integration with quantum hardware remains largely unexplored. We present a proof-of-concept hybrid classical-quantum variational autoencoder (VAE) for topic modeling, embedding parameterized quantum circuits within the VAE inference network while retaining a classical topic-word decoder. To address the resource constraints of quantum hardware, we propose a modified Gaussian Softmax posterior that decouples latent space dimensionality from the number of topics to be extracted, enabling the model to operate with a low-resource 10-qubit quantum device. On the AgNews dataset, the hybrid VAE outperforms state-of-the-art neural topic models (NTMs), reaching a $C_v$ coherence score of 0.71 and an NPMI score of 0.20 while preserving high topic diversity. For comparison, we also construct a fully classical variant, which also outperforms state-of-the-art models on AgNews and exhibits clear class separation in the latent space. These results demonstrate that hybrid VAEs are computationally viable even on NISQ-era devices and represent a promising direction for quantum-enhanced topic modeling.

We present an algorithm for efficient domain-adaptive policy learning via kernel representations. Learning domain-adaptive policies is challenging since it requires an environment representation that is both sufficiently expressive to model complex sim-to-real gaps during offline training, and computationally efficient enough to support rapid online adaptation during deployment. For instance, a quadrotor may encounter time-varying, non-stationary disturbances, such as sudden gusts of wind, payload shifts, or transitions between distinct flight regimes with and without ground effects. To address these challenges, we model unknown disturbances using a differentiable kernel approximation based on random Fourier features. During the offline training phase, we randomly sample kernel coefficients and bandwidth parameters to generate a rich diversity of disturbance profiles. We then optimize the control policy via differentiable simulation with analytical gradients, a process that takes only 50 seconds of training time on an RTX 4090 GPU. During hardware deployment, the policy adapts to non-stationary environments in real time by updating both the kernel coefficients and bandwidth through online least-squares estimation. We evaluate our method on quadrotor trajectory tracking tasks across high-fidelity numerical simulations and hardware experiments using Crazyflie, subjected to various disturbances, including complex aerodynamic effects, wind, ground effects, and payload fluctuations.

Autonomous driving is shifting from isolated vehicle intelligence toward multi-agent embodied systems that share perception, infer intent, and coordinate action under uncertainty. This survey examines this transition through the lens of Shared World Models (SWMs): predictive cross-agent representations maintained across vehicles, infrastructure, and other traffic participants. We review more than 380 publications spanning vehicle-to-everything (V2X) communication, collaborative perception, inter-agent cognition, cooperative planning, end-to-end cooperative driving, and simulation and data engines for closed-loop validation. The organizing question is how exchanged observations become aligned state, intent-aware interaction, and coordinated downstream action. Across the surveyed literature, evaluation remains concentrated in simulation, curated benchmarks, and offline protocols. Foundation-model-based coordination also lacks verified real-time safety guarantees in open traffic. These gaps motivate key research priorities for multi-agent embodied autonomous driving (MAEAD): verifiable shared-state maintenance, robust intent and plan alignment, and safe coordinated action under communication, latency, and deployment constraints.

Remote photoplethysmography (rPPG) transformers achieve low heart-rate error on benchmarks, yet their decisions remain opaque--a growing concern as rPPG moves toward clinical heart rate estimation. Existing rPPG XAI is dominated by qualitative heatmap inspection without quantitative faithfulness metrics or physiology-grounded validation, leaving a gap between visual plausibility and auditable evidence. We address this gap. First, we adapt four attribution methods (raw attention, rollout, flow, Beyond Intuition) to RhythmFormer's bi-level routing attention with top-$k$ selection. Second, we introduce a skin coverage metric quantifying how much attribution mass falls on skin regions. Third, we adapt the SaCo faithfulness coefficient from its original classification setting to rPPG regression by using the MAE between original and perturbed predicted rPPG waveforms as the perturbation impact. Applying these tools, we quantify a multi-hop leakage effect under sparse top-$k$ routing: attention rollout and flow almost completely restores the connections that individual refined-attention layers explicitly set to zero. Beyond Intuition mitigates this via its value-projection-weighted rollout and gradient-supported mask, attaining the highest median refined skin coverage ($0.83$ vs. $0.57$ for vanilla rollout) and faithfulness ($F=0.92$) among the evaluated methods on UBFC-rPPG. Validation across diverse datasets and model variants is needed. A case study on a low-SaCo outlier further shows all four methods recovering consistently once an artefactual region is replaced, suggesting consistent SaCo behavior across attribution families in this illustrative case. Together, these metrics move XAI for rPPG toward auditable numerical evidence about spatial alignment and perturbation faithfulness, i.e. trustworthy rPPG XAI.