Papers

Large language models (LLMs) hallucinate numerical values when generating high-stakes meteorological text, posing risks for weather communication. We present AFDBench, an AI meteorologist that generates professional Area Forecast Discussions (AFDs) by reasoning through structured AI weather forecast data from Google's WeatherNext 2. We introduce AFDBench, the first benchmark for evaluating generative meteorological reasoning, comprising 7,732 expert written discussions from 13 National Weather Service (NWS) offices paired with real AI weather forecast inputs, and three complementary metrics: Met-Align (numerical accuracy), Style-Align (professional dialect adherence), and Input-Grounding (fidelity to source weather data). Zero-shot evaluations reveal that open-source LLMs achieve low Style-Align (~0.33) and moderate Input-Grounding (~0.88), failing to write in the professional NWS register or faithfully use their input data. We apply Group Relative Policy Optimization (GRPO) with domain-specific rewards targeting temperature accuracy, synoptic correctness, and format compliance. On 1,033 held-out samples from two unseen NWS offices, GRPO nearly doubles Style-Align from 0.318 to 0.619 and improves Input-Grounding from 0.881 to 0.940, demonstrating that reinforcement learning teaches a 7B-parameter model to write like a professional meteorologist and faithfully interpret AI weather data.

Generative models are commonly ranked by Fréchet Inception Distance (FID) and Kernel Inception Distance (KID), yet FID's first-two-moment summary can miss distributional differences, and a reported scalar gap alone is not a calibrated test against sampling variation. FID's moment restriction has concrete consequences: on ImageNet, visually unrecognizable images optimized only to match the reference Inception mean and covariance obtain FID $24.7$ versus $58.6$ for held-out real images (lower is better). Moreover, FID and KID are scalar discrepancies that are unchanged when the two samples are exchanged and therefore do not encode the direction of a dispersion change: under-dispersion, as can occur in mode collapse, versus over-dispersion. We introduce ZID (Z-resolved Integrated Diagnostic), which combines six standardized location- and dispersion-sensitive arms from a rank graph (RISE) and Gaussian kernels (GPK at two bandwidths). Rather than asking one scalar to serve incompatible roles, ZID reports three linked outputs: an index for ranking departure magnitude, a permutation $p$-value for testing distributional equality, and a signed dispersion readout for diagnosis. In controlled experiments, ZID detects a broad range of departures, and its score tracks increasing severity along the corresponding sweeps, including cases in which FID is flat or reversed. On DiT-XL/2 and SiT-XL/2 guidance sweeps, ZID detects departure from real data, and its signed readout labels the high-guidance diversity collapse as under-dispersion.

Lipschitz constants are a standard way to quantify the sensitivity of neural networks to small input perturbations, but computing them is difficult even for shallow ReLU networks. We study this problem for two-layer input-convex neural networks (ICNNs), a restricted architecture where nonnegative output weights enforce convexity. Computing the $L_p$-Lipschitz constant for these networks is equivalent to maximizing the dual norm over a zonotope. While $L_1$- and $L_\infty$-norm maximization on zonotopes admit fixed-parameter and polynomial-time algorithms, respectively, the parameterized complexity of the remaining $L_p$-norms was open. We prove that, for every fixed $p\in (1,\infty)\cap \mathbb{Q}$, maximizing the $L_p$-norm over a zonotope in $\mathbb{R}^d$ is W[1]-hard with respect to the dimension $d$. Moreover, our hardness results imply that brute-force enumeration algorithms are essentially optimal for this problem under the Exponential Time Hypothesis. By duality, the same hardness results hold for computing the $L_p$-Lipschitz constant of two-layer ReLU ICNNs. Our proof first establishes the result for the $L_2$-norm and then transfers the construction to arbitrary fixed $p\in (1,\infty)\cap\mathbb{Q}$ using a suitable Taylor approximation. These results resolve the corresponding questions regarding the parameterized complexity status for zonotope norm maximization and two-layer ICNN Lipschitz constants. Our paper resolves an open problem posted at COLT'25. There are several independent concurrent papers resolving the same problem. Our paper prioritizes a clear exposition of the underlying mathematics and conceptual intuitions behind the proof. Additionally, we explicitly describe our research process including the use of LLMs.

Multi-task vehicle routing problem (VRP) solvers seek to handle multiple VRP variants within a single unified model, avoiding the need to train a separate model for every variant. In spite of recent progress, current approaches remain limited on two fronts. On the training side, reinforcement learning suffers from reward-scale disparities and shrinking advantage signals as policies improve, whereas preference optimization stagnates once sampled tours become near-identical and thus fundamentally limited by the quality of the policy's own generated solutions, leaving both paradigms with weak supervision as training progresses. On the architecture side, existing fully shared encoders entangle constraint-dependent representations across heterogeneous variants, which limits generalization. We address these gaps with two model-agnostic contributions. First, we propose Preference Optimization with Locally Augmented Refinement (POLAR), a novel training algorithm that applies a local search refinement pass to the best decoded tour before forming preference pairs, yielding much more informative pairwise margins. Second, a Progressive Layered Extraction (PLE) encoder routes each encoder layer through one shared expert and a set of task-specific experts via a gating mechanism, progressively separating common routing structure from constraint-specific encodings. Through extensive experiments on various VRP variants, we show that POLAR and PLE together elevate the current state-of-the-art among neural multi-task solvers. We reduce the average gap to reference solutions by 21.3% relative to the strongest published baseline on 16 in-distribution variants, and outperform prior neural methods on 27 out of 32 unseen variants. Ablation studies confirm the efficacy of each contribution, showing that both improve cross-problem generalization across multiple backbone model architectures.

Marginalized importance weighting evaluates a target policy by reweighting offline state-action samples with its discounted occupancy ratio, characterized by an adjoint Bellman equation. Existing minimax, primal-dual, and fitted fixed-point estimators can leave residual occupancy-balance violations because of function-class approximation, regularization, or incomplete optimization. These violations are difficult to diagnose and reduce because the objectives generally lack a direct supervised validation loss for hyperparameter tuning, model selection, and early stopping. We introduce isotonic Bellman calibration, a one-dimensional, model-agnostic post-processing method that reduces these violations while preserving the ranking information in any initial occupancy-ratio estimate. The method corrects the estimate's scale and shape by applying fitted occupancy-ratio evaluation (FORE) over a one-dimensional class of nondecreasing transformations. We characterize Bellman calibration as a conditional fixed-point property equivalent to occupancy-balance against every test function of the calibrated ratio. More generally, we derive a calibration-refinement bound showing that any fitted ratio with small calibration error performs nearly as well as the best post-processing based on its fitted values. For isotonic Bellman calibration, we establish finite-sample calibration guarantees and a KL oracle inequality relative to the best monotone transformation of the initial estimate. Consequently, isotonic Bellman calibration achieves small calibration error and KL risk within statistical error of the best monotone correction, with guarantees for downstream target-occupancy functionals, including policy-value estimation.

We present LAION-BVD, a large-scale open video dataset for multimodal learning, which contains 1.3B platform-specific video URLs collected from CommonCrawl. From these, we download 80M videos with a total duration of 10 million hours. The dataset is designed for multimodal pre-training across the video, audio, and image modalities. Using content-aware scene detection, we extract clips for which we synthetically generate video and audio captions. Models trained on these data achieve competitive performance on standard video-text and audio-text benchmarks, with consistent improvements as training or model scale increases. Additionally, we explore video frames as an alternative source of image-text data by extracting scene-changing frames. These frames exhibit a visual distribution distinct from standard web image corpora, and models trained on this dataset achieve strong image-text retrieval performance. We release LAION-BVD to the research community. It significantly expands open access to multimodal videos at an unprecedented scale.

Spatially resolved biology requires representations that preserve biological neighborhood structure rather than only exact cross-modal correspondences. Existing histology–transcriptomics objectives can emphasize instance-level matching even when non-paired spots share molecular or spatial context. We introduce BioKERN, a multimodal spatial representation-learning framework that incorporates biological structure as an explicit, learnable inductive bias. BioKERN constructs a training-time biological kernel by combining transcriptomic similarity and spatial proximity, then uses it to provide graded neighborhood supervision and regularize embedding geometry. Evaluation uses a fixed, model-independent biological neighborhood definition shared by all methods. Across Mouse Brain Visium and Human Liver GSE240429, BioKERN consistently improves biological-neighborhood retrieval over BLEEP in both single- and multi-scale settings. Controlled shared-architecture experiments show that most of the improvement arises from biological-kernel regularization rather than increased model capacity. These results support explicit biological geometry as an interpretable inductive bias for multimodal learning in spatial biology.

Neighborhood-based fairness audits evaluate individual fairness by comparing predictions among similar individuals in feature space. Despite their widespread use, little is known about the robustness of the auditing procedure itself. Because these audits rely on nearest neighbor relationships, small perturbations in feature space can alter local neighborhoods and produce different fairness assessments even when model predictions remain unchanged. We develop a geometric framework for analyzing the robustness of neighborhood-based fairness audits under bounded perturbations. Our analysis establishes sufficient conditions for neighborhood invariance, quantifies how neighborhood replacement propagates to audit instability, and introduces audit volatility, a measure of the expected sensitivity of fairness audits under repeated perturbations. Experiments on benchmark datasets support the theoretical analysis and show that the proposed framework explains the observed stability of neighborhood-based fairness audits.

We uncover ELR collapse in language model pretraining: learning rate (LR) and parameter norm govern loss dynamics primarily through their ratio, the effective learning rate (ELR). When ELR is matched across runs, their loss trajectories collapse throughout training despite substantially different LRs and parameter norms. Across optimizers, architectures, datasets, and model scales, mean collapse errors are typically a few x 10^-3, below the seed-to-seed variation measured in a representative configuration. Systematic ablations identify normalization design and the timescale of LR-norm variation as key determinants of collapse precision. Controlled interventions further show that weight decay and Hyperball shape loss dynamics primarily through the ELR schedules they induce. Replacing LR with ELR enables a fitted functional scaling law (FSL) to transfer across norm-control methods. The resulting ELR-based FSL also explains delayed acceleration, a recurring effect of norm control. Together, these results establish ELR as a common coordinate linking LR scheduling, norm control, and loss dynamics.

Class-imbalanced node classification on temporal graphs is challenging because majority-dominated temporal propagation progressively assimilates minority representations, while conventional node and neighborhood information provides insufficient discriminative evidence for minority classes. To address these issues, we propose MDTE, a minority-aware diffusion framework that reconstructs stable and discriminative temporal edge-event representations through conditional diffusion denoising. Specifically, MDTE introduces Distribution-Aware Selective Propagation, which combines Local Outlier Factor (LOF)-based propagation filtering with cluster-aware low-frequency propagation. The module preserves informative neighborhood dependencies while mitigating harmful propagation and majority-class information assimilation. It further develops Multi-View Discriminative Fusion, which exploits feature reconstruction and topology prediction to characterize class-wise differences in distribution learning and extracts complementary discriminability signals to guide denoising. Experiments on five real-world datasets demonstrate that MDTE consistently achieves the best performance on minority-class-oriented metrics, improving minority-class recall by up to 23.53 percentage points, minority-class F1 by 8.68 percentage points, and AUPRC by 2.67 percentage points over the strongest baselines.

Recently, the rapid advancement of multimodal domains has driven a data-centric paradigm shift in graph ML, transitioning from text-attributed to multimodal-attributed graphs. This advancement significantly enhances data representation and expands the scope of graph downstream tasks, such as modality-oriented tasks, thereby improving the practical utility of graph ML. Despite its promise, limitations exist in the current neural paradigms:(1) Neglect Context in Modality Alignment: Most existing methods adopt topology-constrained or modality-specific operators as tokenizers.These aligners inevitably neglect graph context and inhibit modality interaction, resulting in suboptimal alignment.(2) Lack of Adaptation in Modality Fusion: Most existing methods are simple adaptations for 2-modality graphs and fail to adequately exploit aligned tokens equipped with topology priors during fusion, leading to poor generalizability and performance degradation.To address the above issues, we propose LION (c\underline{LI}ff\underline{O}rd \underline{N}eural paradigm) based on the Clifford algebra and decoupled graph neural paradigm (i.e., propagation-then-aggregation) to implement alignment-then-fusion in multimodal-attributed graphs. Specifically, we first construct a modality-aware geometric manifold grounded in Clifford algebra.This geometric-induced high-order graph propagation efficiently achieves modality interaction, facilitating modality alignment.Then, based on the topology-aware Clifford components of aligned tokens, we propose adaptive holographic aggregation. This module integrates component-wise energy and propagation-scale information with learnable parameters to improve modality fusion. Extensive experiments on 9 text-image MAG datasets demonstrate that LION significantly outperforms SOTA baselines across 3 graph and 3 modality downstream tasks.

The Bayesian Ideal Observer (IO) establishes the theoretical upper bound on task performance for binary detection tasks. However, analytical computation of the IO test statistic is generally intractable. Numerical approaches based on Markov-chain Monte Carlo (MCMC) methods, including their recent deep generative model-based extensions, typically require extensive posterior sampling for each test image. Supervised learning has also been investigated to approximate the IO performance. However, such methods are typically trained for a specific detection task and signal and may require retraining when the task or signal changes. The score function, defined as the gradient of the log probability density, encodes the local geometry of the data distribution and is a fundamental quantity in modern score-based generative modeling. This work reformulates the IO test statistic in terms of the score function and introduces a score-based ideal observer (SIO). The proposed SIO uses a denoising convolutional neural network trained exclusively on signal-absent images to estimate the signal-absent score function. Once trained, the resulting score model can be used to approximate the IO test statistic for detection tasks involving arbitrary additive signals, without per-image posterior sampling or signal-specific retraining. Numerical studies consider a signal-known-exactly (SKE) detection task with a stochastic lumpy-background model. The results demonstrate that the proposed SIO can closely approximate the IO performance.

Procedural video-language models must solve heterogeneous tasks from the same visual evidence, including action recognition, forecasting, and procedure prediction. Dense transformer decoders share the same feed-forward networks across tasks, which can entangle task behavior and make controlled capability expansion difficult. Sparse Mixture-of-Experts (MoE) decoders provide conditional computation, but token-level learned routing is not naturally aligned with task-level procedural objectives. We propose MoTE (Mixture of Task Experts), a decoder architecture that converts large language model feed-forward networks into task-specific experts while keeping the multimodal backbone shared. Each example follows one sample-level task route, so active task-expert computation remains independent of the number of stored task experts. We instantiate this design as VideoLLM-MoTE and evaluate it on five COIN benchmarks using explicit task routes. The five-expert model activates ~2B LLM parameters per sample and achieves higher average top-1 accuracy than recent VideoLLM baselines. Under the same expert topology, it improves over dense all-expert activation and learned sparse-routing controls. These results show that task-structured routing provides an interpretable and compute-efficient decoder alternative for multi-task video-language learning.

Many disentanglement methods represent generative factors using Euclidean product coordinates, although the underlying factor spaces may wrap, collapse, or have position-dependent geometry. We introduce factor-space structure, which combines factor domains, generator-induced identifications, and position-dependent scales to distinguish topologically equivalent spaces with different factor geometries. We show that statistical independence does not imply uniform product geometry: independent hue and scale can affect observations at different, scale-dependent rates, producing anisotropy that no fixed rescaling can remove. We propose the Factor-Space Topographic Map (FactoMap), which learns interpretable prototypes indexed by a structured factor-space lattice. Topographic learning transfers the lattice's periodicity, collapses boundaries, and non-uniform extent to the representation. Experiments show that matching this structure better preserves factor continuity and improves disentanglement of the underlying factors.

Seagrass meadows are crucial blue-carbon habitats, and mapping their extent is a prerequisite for coastal management and carbon inventory. Optical satellite sensors cover large areas but cannot reach deep or turbid water, whereas side-scan sonar (SSS) images the seabed at high resolution and at any depth. Interpreting SSS, however, still relies on dense manual annotation, which is slow and costly. We address this by adapting a weakly supervised semantic segmentation framework to SSS benthic habitat mapping, so that pixel-level maps are learned from image-level labels alone. The framework couples a ViT-based encoder-decoder with a classification branch, extracts class activation maps, and refines them into pseudo-labels with a dense conditional random field that we tune for the noise and weak boundaries of acoustic imagery. It follows an iterative self-training scheme, together with a sampling strategy to cope with the strong class imbalance of the data. We also study the effect of different loss functions on segmentation quality, finding Lovász-Softmax loss the most effective. On a held-out transect, the refined pseudo-labels reached an mIoU of 89.3% against the ground truth, and the segmentation branch, trained without any pixel-level labels, reached 87.6%. Self-supervised pretraining on unlabelled SSS added a further 3% in mean intersection-over-union. Field trials further demonstrate the generalizability of the trained model. These results show that accurate and label-efficient benthic habitat mapping from side-scan sonar is feasible at the scale needed for coast-wide seagrass monitoring.

How can humans make sense of the rapid takeoff of artificial intelligence (AI)? We studied the sensemaking dynamics of AI through an open-ended, mixed-methods study with computational text analysis of millions of AI-related newspaper articles and social media posts grounded in 57 semi-structured interviews with AI professionals in 2021 and 2023–before and after the recent surge of public interest. We identify a range of sociological frames (interpretive schemas that structure collective cognition) and show how AI professionals use frames to address significant cognitive challenges, such as assigning responsibility for societal impacts. We develop a framework of three primary debates across which frames are adopted and contested: (i) the $\textit{method}$ of AI development, between frames of top-down expert systems and bottom-up emergent capabilities, (ii) the $\textit{mind}$ of an AI system, ranging from a passive tool to a humanlike "digital mind," and (iii) the $\textit{morality}$ of how AI is used, particularly the decision of whether to slow down or speed up AI development. As humanity enters the era of transformative AI, technologists and policymakers must account for the framing dynamics that will circumscribe our beliefs, values, and actions.

Existing linear program (LP) and semidefinite program (SDP) relaxations for rectified linear unit (ReLU) neural network (NN) verification yield overly-conservative safety guarantees due to significant relaxation gaps. While the completely positive program (CPP) formulation closes this gap, it is NP-hard to solve. Its cheapest tractable relaxation, the doubly non-negative program (DNN), retains critical constraints as an SDP, but one whose size exceeds the reach of interior-point methods at practical scale. While Burer-Monteiro (BM) factorization has been applied to make SDP-based verification scalable, no such result exists for the strictly tighter DNN formulation. A key obstacle is that additional non-negativity constraints in the DNN cause dual multipliers for optimality certification to be non-unique, making standard certification methods inapplicable. We propose a novel eigenvalue maximization procedure that searches the non-unique multiplier space for a valid certificate, i.e. a global optimality guarantee. Experiments demonstrate that our approach $(\text{DNN})^2$ produces bounds consistently tighter than the standard SDP method, often matching the exact solution, and that our certification procedure confirms global optimality when a valid certificate exists. These results are a key step toward providing tight, certifiable, and computationally scalable verification guarantees needed to deploy neural network controllers and perception modules in safety-critical autonomous systems.

We settle the minimax-optimal alternating regret, a regret notion motivated by alternating learning dynamics in games, for both online linear optimization (OLO) and online convex optimization (OCO). For OLO over the probability simplex $\Delta_d$, we give an algorithm with $O(\log d)$ alternating regret that remains a constant for any time horizon $T$, and a matching lower bound. Our constant regret bound significantly improves previous results with $O(\log ^{2/3}d \cdot T^{1/3})$ regret [Cevher, Cutkosky, Kavis, Piliouras, Skoulakis, Viano, NeurIPS 2023, Hait, Li, Luo, Zhang, COLT 2025]. As a result, we obtain alternating learning dynamics with $O(\log d /T)$ convergence to Nash equilibria in two-player zero-sum games and $O(\log d /T)$ convergence to coarse correlated equilibria in two-player general-sum games. This is the first uncoupled learning dynamics with $O(1/T)$ convergence to CCE in two-player general-sum games, while all prior works suffer additional $\log T$ factors. For general OCO over a $d$-dimensional compact convex set, we give an algorithm with $O(d\log (1+T/d))$ alternating regret, improving the previous best of $\widetilde{O}(d^{2/3}T^{1/3})$. We also prove a matching lower bound of $\Omega(d\log (1+T/d))$, showing that the $\Omega(\log T)$ factor is unavoidable.

EEG foundation models pretrained via self-supervised learning promise transferable representations, but their generalization remains limited, especially across diverse clinical datasets. Full fine-tuning is impractical for resource-constrained clinical settings due to high computational requirements. In this work, we investigate whether parameter-efficient self-supervised adaptation, updating only 9% of parameters suffices to align representations to target tasks. We evaluate our method on two state-of-the-art models with different pretraining objectives: BIOT (contrastive) and CBraMod (masked reconstruction), and evaluate on three clinical EEG datasets for abnormality detection (TUAB), event classification (TUEV), and seizure detection (CHB-MIT) under both in-distribution and out-of-distribution conditions. SSL adaptation yields consistent gains over linear probing, up to 20x AUCPR. Under a fixed compute budget, peak performance requires only 20–50% of available unlabeled data. Critically, when total window count is fixed, performance remains invariant to patient count, suggesting that performance is dependent on overall temporal window diversity only. Our findings demonstrate that parameter-efficient adaptation enables effective deployment of EEG Foundation models (EEG-FM) with minimal computational overhead and data collection burden. Code available at: https://github.com/c3n-group/efficient-eeg-adapt

Hyperparameter selection remains a key challenge in Bayesian optimization (BO) and Bayesian active learning (AL), as model misspecification can lead to suboptimal performance, while more accurate fully Bayesian treatments typically rely on computationally expensive MCMC sampling. This paper proposes a unified framework, KENDO (Kernel ENsemble Disagreement-aware Operator), that integrates Ensemble Gaussian Processes (EGP) with disagreement-aware acquisition strategies. The central idea is to replace hyperparameter sampling with a kernel ensemble and adaptive Bayesian weighting, combined with disagreement-aware acquisition strategies. Within this unified framework, we instantiate KENDO-BO for BO and KENDO-AL for Bayesian AL, demonstrating that both arise from a common self-correcting mechanism with task-specific acquisition objectives. We further extend the approach to multi-objective optimization via random scalarization that preserves the single-optimizer conditioning structure. Thorough numerical tests on synthetic and real-world benchmarks across single-objective optimization, multi-objective optimization, and active learning demonstrate that (i) KENDO-BO achieves competitive or superior optimization performance compared to state-of-the-art methods while reducing computational overhead by up to $5\times$ and (ii) KENDO-AL achieves superior predictive calibration over MCMC-based active learning baselines with up to $27\times$ speedup.

Lifted inference algorithms enable scalable probabilistic inference even for large object domains by leveraging the indistinguishability of objects in a probability distribution. An essential prerequisite for constructing a lifted representation is to identify commutative factors, i.e., functions whose output values are invariant under permutations of a subset of their input values, in a potential-based factorisation. In practice, however, parameters learned from data inevitably deviate even if associated objects are indistinguishable, causing their corresponding factors to be only approximately commutative instead of being exactly commutative. We address this problem by introducing the concept of {\epsilon}-commutativity, a relaxation of commutativity where output values are only approximately invariant under permutations of input values. Specifically, we show how {\epsilon}-commutativity can be exploited for lifted model construction, downstream probabilistic inference, and prove strict bounds on the induced approximation error, thereby ensuring the practical applicability of lifted model construction while maintaining highly accurate query results. These theoretical guarantees are confirmed empirically, demonstrating comparable query accuracy at lower runtime.

Optimization of hyperparameters is a critical factor to obtain optimal model performance. While existing research has predominantly concentrated on batch-learning scenarios, addressing the complexities inherent in data streams presents a challenge. The deployment of sophisticated methodologies to manage data streams becomes highly important. Consequently, the capacity for self-adjusting hyperparameters during on-line learning phases emerges as a goal. Many hyperparameters exhibit constraints and are confined within bounded search spaces, rendering specific solutions unacceptable upon applying optimization operators. To solve this issue, employing boundary constraint- handling techniques becomes imperative to rectify invalid solutions. This paper presents strategies for effectively managing boundary constraints within constrained numerical optimization problems. Recent methodologies, including heuristic and evolutionary-based optimization, employ a "boundary" strategy, wherein values that surpass boundary thresholds for a given hyperparameter are realigned to the respective limits. Our study introduces four strategies to navigate boundary constraints in online optimization algorithms. Through empirical investigations conducted on established datasets, we demonstrate that adopting boundary strategies outperforms the "boundary" strategy.

When a network has learned a function with a known symmetry, can that symmetry be moved through the parametrisation—is there a motion in parameter space realising the group action in function space? We formulate this as a lifting problem for the realisation map $\Phi:\theta\mapsto f_\theta$, and show that a smooth parameter-space action exists only if the tangent space to the function's symmetry orbit lies within the image of $\mathrm d\Phi_\theta$, whose columns are the functional sensitivities of individual parameters. This condition is also sufficient for pointwise first-order lifting. Relaxing it in least squares yields two local parameter directions: one following the symmetry orbit, one descending towards the equivariant subspace, with residuals measuring what the parametrisation cannot reach. On a rotationally invariant classifier we find these directions induce their predicted function-space motion, but only locally: recomputed directions track the orbit and reduce the equivariance defect, while directions held fixed depart from both after training. The same holds for Hamiltonian neural networks trained on a rotationally symmetric potential, even though the architecture does not explicitly enforce the symmetry.

Predictive Coding (PC) is a neural learning paradigm that enables parallelizable neural network layer updates. However, the main bottleneck of PC Networks (PCN) is the sequential backwards error propagation. To tackle this, we introduce a training technique that pairs a Generative PCN with a support Encoding PCN. The two PCNs are trained in parallel to match their neural activations, without sequential propagation. We apply this to time series anomaly detection and show that our approach results in more stable, continuous, online learning.

Reinforcement Learning with Verifiable Rewards (RLVR) and on-policy distillation (OPD) have become two widely adopted paradigms for post-training large language models. However, RLVR suffers from sparse task-level feedback, while OPD provides dense token-level guidance but ignores trajectory correctness, limiting its performance to that of the teacher. Combining them is a promising direction: OPD supplies dense supervisory signals, while RLVR provides task-level correctness. Nevertheless, existing integrations often rely on weighted combination or heuristic switching, introducing extra hyperparameters and trade-offs. We propose On-policy Distillation with Verifiable Reward (OPDVR), a simple yet effective method that seamlessly combines OPD and RLVR without adding any hyperparameters. We first reformulate the implicit reward of sampled-token OPD based on trajectory correctness, then apply a ReLU gating mechanism to ensure that correct trajectories receive non-negative rewards and incorrect ones receive non-positive rewards—thereby aligning the distillation signal with task success while preserving the teacher's distributional guidance. Furthermore, our modification transforms sampled-token OPD into a proper RLVR method, making it readily combinable with any policy gradient algorithm, such as GRPO. Experiments on six reasoning benchmarks show that OPDVR consistently outperforms standard OPD. Our code is available at https://github.com/LeapLabTHU/OPDVR.

Agentic systems increasingly gate actions on a model's own stated confidence, which assumes confidence tracks correctness at the moment of acting. We test this in a hidden-information chess variant where royal status can be secretly, repeatedly relocated between pieces, and where an agent's stated probability distribution over the opponent's hidden royal piece – elicited every turn, separately from the move it chooses – is scored against ground truth recoverable after the game. Across two independent batches, captures made at high stated confidence ($\geq 0.5$) about the hidden piece's location were correct in 1 of 62 cases. The calibration deficit is concentrated almost entirely in these events: 99.3% of it in the original batch, 98.7% in the replication. The same pattern, in weaker form, orders consistently (point estimates only; most pairwise gaps are not statistically distinguishable at this sample size) across four further model configurations spanning a second provider – reported as scope for the finding, not as evidence that capability predicts calibration: a same-model comparison at a fixed external leaderboard score shows a deliberation-budget change alone moves the metric by nearly as much as a large cross-model gap. In a separate seat, conventional evaluation axes – legality, cost, latency, completion rate – can dissociate entirely from belief quality, with the configuration winning on every conventional axis producing the worst belief quality tested. A model exhibiting this pattern can still win the game its belief was about, which is why outcome-only evaluation would not detect it.

Precision oncology necessitates a longitudinal model of patient state that captures cancer evolution and treatment over time, integrating multimodal observations. We introduce the oFM, a foundation model developed on a real-world oncology cohort of 1.67 million cancer patients that integrates clinical trajectories with DNA, RNA, and H&E pathology. Patient-level partitions were reserved for training, validation, and testing, with over one million patients used for training. The oFM encodes daily clinical and molecular episodes and, along with pathology images, integrates them over time to produce a patient state embedding. We evaluate frozen oFM embeddings against expert-curated clinical and molecular baseline features. In prognostic benchmarks, the oFM improved AUC for treatment response, progression-free survival, and overall survival (0.774 vs. 0.563 for overall survival). Across 11 comparative-treatment cohorts, the oFM embeddings achieved a three-fold higher pooled and scale-normalized treatment-benefit AUTOC than baseline features with improved benefit ranking in 9 of 11 cohorts, and provided stronger prognostic discrimination within both treatment arms. We also evaluated a mechanism discovery framework that interprets downstream models built on oFM embeddings by linking their predicted outcomes to clinically and biologically grounded mechanisms through an evidence-grounded temporal graph, enabling evaluation in clinical and drug-development applications.

Power outage prediction models are increasingly used in assessments of climate-driven infrastructure risk, yet current evaluation practices obscure whether these models generalize to the novel conditions such applications require. We identify three common methodological choices in power outage prediction models that influence their ability to generalize across spatial, temporal, and event-based settings. We compare the predictive performance impacts of different methodological decisions using publicly available data for the U.S. East Coast from 2018 to 2023 and feature sets derived from weather reanalysis and land-cover data, and embeddings from a GeoAI foundation model (Prithvi WxC). Specifically, we assess model performance under multiple test selection strategies, including unfiltered random splits, leave-one-state-out, and leave-one-event-out designs, which increasingly approximate real-world deployment conditions. While random train-test splits yield strong performance, we show that these results are inflated by spatial and temporal autocorrelation. Under spatial and temporal holdout experiments, predictive accuracy degrades substantially, with models often failing to outperform a simple null baseline. Incorporating GeoAI foundation model embeddings yields limited and inconsistent improvements, primarily for spatial generalization, and does not resolve poor event-level transferability. These findings suggest that, given current data availability and evaluation practices, publicly trained outage prediction models offer limited and uncertain operational value. Progress will likely require improved data coverage, more realistic evaluation protocols, and a shift in focus from marginal modeling advances toward addressing structural data constraints.

We introduce Maia 200, an advanced AI accelerator delivering high performance-10 145 Tflop/s FP4 and 5072 Tflop/s FP8 within a 750W TDP and 7 TB/s HBM bandwidth. Maia exemplifies a new class of Software Defined Locally Accessed Dataflow Architectures (SDLA), which explicitly program dataflow engines to orchestrate highly specialized memories and data movement engines. This approach shifts the focus from today's thread-centric to data-movement-centric architecture, improving efficiency and scalability. Our taxonomy of data management, inspired by Flynn's classification, highlights how SDLA addresses challenges in modern AI computing. Maia 200 achieves significant cost and energy savings while supporting massive parallelism for AI inference workloads, making it a compelling solution for next-generation high-performance computing systems.

Similarity in many decision systems is governed not by distance alone but by interactions among variables. In fraud and anomaly detection, small local perturbations can cross interaction-sensitive decision boundaries while leaving ambient distance almost unchanged. Motivated by this setting, we introduce a thin-slab interaction model and an interaction-driven quantum kernel constructed from entangled Pauli-string feature maps. The feature map explicitly encodes sparse high-order block interactions. We show that the resulting fidelity kernel is positive semidefinite, admits an exact block-factorized formulation, and induces a geometry sensitive to changes in interaction regime. Across balanced and imbalanced synthetic experiments spanning third-, fourth-, sixth-, and eighth-order interactions, the proposed kernel consistently outperforms linear, radial basis function, Laplacian, and polynomial kernels, as well as an engineered-interaction linear baseline supplied with the planted block products. On real fraud-detection benchmarks, it achieves the highest mean accuracy and F1 on Credit Card Fraud Detection and ranks second on IEEE-CIS Fraud Detection. Executed on a 156-qubit IBM Quantum processor in a fourth-order setting, the hardware-estimated kernel matches the noise-free simulator within seed-to-seed variability and retains its advantage over the baselines. These findings show that quantum-kernel performance depends on alignment between feature-map geometry and the underlying predictive structure, rather than on Hilbert-space dimension alone. Because the prescribed block-factorized kernel can also be evaluated exactly on a classical computer, the results establish predictive and representational value rather than computational quantum speedup.