Papers

Conformal prediction (CP) has attracted broad attention as a simple and flexible framework for uncertainty quantification through prediction sets. In this work, we study how to deploy CP under differential privacy (DP) in a statistically efficient manner. We first introduce differential CP, a non-splitting conformal procedure that avoids the efficiency loss caused by data splitting and serves as a bridge between oracle CP and private conformal inference. By exploiting the stability properties of DP mechanisms, differential CP establishes a direct connection to oracle CP and inherits corresponding validity behavior. Building on this idea, we develop Differentially Private Conformal Prediction (DPCP), a fully private procedure that combines DP model training with a private quantile mechanism for calibration. We establish the end-to-end privacy guarantee of DPCP and investigate its coverage properties under additional regularity conditions. We further study the efficiency of both differential CP and DPCP under empirical risk minimization and general regression models, showing that DPCP can produce tighter prediction sets than existing private split conformal approaches under the same privacy budget. Numerical experiments on synthetic and real datasets demonstrate the practical effectiveness of the proposed methods.

In computational paralinguistics, detecting cognitive load and deception from speech signals is a heavily researched domain. Recent efforts have attempted to apply these acoustic frameworks to corporate earnings calls to predict catastrophic stock market volatility. In this study, we empirically investigate the limits of acoustic feature extraction (pitch, jitter, and hesitation) when applied to highly trained speakers in in-the-wild teleconference environments. Utilizing a two-stream late-fusion architecture, we contrast an acoustic-based stream with a baseline Natural Language Processing (NLP) stream. The isolated NLP model achieved a recall of 66.25% for tail-risk downside events. Surprisingly, integrating acoustic features via late fusion significantly degraded performance, reducing recall to 47.08%. We identify this degradation as Acoustic Camouflage, where media-trained vocal regulation introduces contradictory noise that disrupts multimodal meta-learners. We present these findings as a boundary condition for speech processing applications in high-stakes financial forecasting.

Clinical value set authoring – the task of identifying all codes in a standardized vocabulary that define a clinical concept – is a recurring bottleneck in clinical quality measurement and phenotyping. A natural approach is to prompt a large language model (LLM) to generate the required codes directly, but structured clinical vocabularies are large, version-controlled, and not reliably memorized during pretraining. We propose Retrieval-Augmented Set Completion (RASC): retrieve the $K$ most similar existing value sets from a curated corpus to form a candidate pool, then apply a classifier to each candidate code. Theoretically, retrieve-and-select can reduce statistical complexity by shrinking the effective output space from the full vocabulary to a much smaller retrieved candidate pool. We demonstrate the utility of RASC on 11,803 publicly available VSAC value sets, constructing the first large-scale benchmark for this task. A cross-encoder fine-tuned on SAPBert achieves AUROC~0.852 and value-set-level F1~0.298, outperforming a simpler three-layer Multilayer Perceptron (AUROC~0.799, F1~0.250) and both reduce the number of irrelevant candidates per true positive from 12.3 (retrieval-only) to approximately 3.2 and 4.4 respectively. Zero-shot GPT-4o achieves value-set-level F1~0.105, with 48.6% of returned codes absent from VSAC entirely. This performance gap widens with increasing value set size, consistent with RASC's theoretical advantage. We observe similar performance gains across two other classifier model types, namely a cross-encoder initialized from pre-trained SAPBert and a LightGBM model, demonstrating that RASC's benefits extend beyond a single model class. The code to download and create the benchmark dataset, as well as the model training code is available at: \href{https://github.com/mukhes3/RASC}{https://github.com/mukhes3/RASC}.

We give an algorithm for PAC learning intersections of $k$ halfspaces with a $\rho$ margin to within error $\varepsilon$ that runs in time $\textsf{poly}(k, \varepsilon^{-1}, \rho^{-1}) \cdot \exp \left(O(\sqrt{n \log(1/\rho) \log k})\right)$. Notably, this improves on prior work which had an exponential dependence on either $k$ or $\rho^{-1}$ and matches known cryptographic and Statistical Query lower bounds up to the logarithmic factors in $k$ and $\rho$ in the exponent. Our learning algorithm extends to the more general setting when we are only promised that most points have distance at least $\rho$ from the boundary of the polyhedron, making it applicable to continuous distributions as well.

Self-speculative decoding is an inference technique for large language models designed to speed up generation without sacrificing output quality. It combines fast, approximate decoding using a compact version of the model as a draft model with selective re-evaluation by the full target model. Some existing methods form the draft model by dynamically learning which layers to skip during inference, effectively creating a smaller subnetwork to speed up computation. However, using heuristic-based approaches to select layers to skip can often be simpler and more effective. In this paper, we propose ConfLayers, a dynamic plug-and-play approach to forming the draft model in self-speculative decoding via confidence-based intermediate layer skipping. The process iteratively computes confidence scores for all layers, selects layers to skip based on an adaptive threshold, evaluates the performance of the resulting set, and updates the best selection until no further improvement is achieved or a maximum number of iterations is reached. This framework avoids the overhead and complexity of training a layer skipping policy and can provide more consistent speed-quality trade-offs while preserving the adaptivity of the draft model to diverse tasks and datasets. The performance evaluation of ConfLayers across different models and datasets shows that our novel approach offers up to 1.4x speedup over vanilla LLM generation.

The fundamental limit of natural signal compression has traditionally been characterized by classical rate-distortion (RD) theory through the tradeoff between coding rate and reconstruction distortion, while the rate-distortion-perception (RDP) framework introduces a divergence-based measure of perceptual quality as a modeling principle, leaving its theoretical origin unclear. In this paper, motivated by a synonymity-based semantic information perspective, we reformulate perceptual reconstruction as recovering any admissible sample within an ideal synonymous set (synset) associated with the source, rather than the source sample itself, and establish a synonymous source coding architecture. On this basis, we develop a synonymous variational inference (SVI) analysis framework with a synonymous variational lower bound (SVLBO) for tractable analysis of synset-oriented compression. Within this framework, we establish a synonymity-perception consistency principle, showing that optimal identification of semantic information is theoretically consistent with perceptual optimization. Based on this result, we further derive a tight-bound synonymous source coding rate characterization and show that its Jensen-limit relaxation leads to a synonymous rate-distortion-perception form for practical optimization. These analytical results show that the distributional divergence term arises naturally from the synset-based reconstruction objective, clarify its compatibility with existing RDP formulations and classical RD theory, and suggest the potential advantages of synonymous source coding.

Lion optimizer is a popular learning-based optimization algorithm in machine learning, which shows impressive performance in training many deep learning models. Although convergence property of the Lion optimizer has been studied, its generalization analysis is still missing. To fill this gap, we study generalization property of the Lion via algorithmic stability based on the mathematical induction. Specifically, we prove that the Lion has a generalization error of $O(\frac{1}{N\tau^T})$, where $N$ is training sample size, and $\tau>0$ denotes the smallest absolute value of non-zero element in gradient estimator, and $T$ is the total iteration number. In addition, we obtain an interesting byproduct that the SignSGD algorithm has the same generalization error as the Lion. To enhance generalization of the Lion, we design a novel efficient Cautious Lion (i.e., CLion) optimizer by cautiously using sign function. Moreover, we prove that our CLion has a lower generalization error of $O(\frac{1}{N})$ than $O(\frac{1}{N\tau^T})$ of the Lion, since the parameter $\tau$ generally is very small. Meanwhile, we study convergence property of our CLion optimizer, and prove that our CLion has a fast convergence rate of $O(\frac{\sqrt{d}}{T^{1/4}})$ under $\ell_1$-norm of gradient for nonconvex stochastic optimization, where $d$ denotes the model dimension. Extensive numerical experiments demonstrate effectiveness of our CLion optimizer.

Decentralized Finance (DeFi) lending protocols like Aave v3 rely on over-collateralization to secure loans, yet users frequently face liquidation due to volatile market conditions. Existing risk management tools utilize static health-factor thresholds, which are reactive and fail to distinguish between administrative "dust" cleanup and genuine insolvency. In this work, we propose an autonomous agent that leverages time-to-event (survival) analysis and moves beyond prediction to execution. Unlike passive risk signals, this agent perceives risk, simulates counterfactual futures, and executes protocol-faithful interventions to proactively prevent liquidations. We introduce a return period metric derived from a numerically stable XGBoost Cox proportional hazards model to normalize risk across transaction types, coupled with a volatility-adjusted trend score to filter transient market noise. To select optimal interventions, we implement a counterfactual optimization loop that simulates potential user actions to find the minimum capital required to mitigate risk. We validate our approach using a high-fidelity, protocol-faithful Aave v3 simulator on a cohort of 4,882 high-risk user profiles. The results demonstrate the agent's ability to prevent liquidations in imminent-risk scenarios where static rules fail, effectively "saving the unsavable" while maintaining a zero worsening rate, providing a critical safety guarantee often missing in autonomous financial agents. Furthermore, the system successfully differentiates between actionable financial risks and negligible dust events, optimizing capital efficiency where static rules fail.

Marketing research often relies on parameters estimated from costly human-generated data, such as conjoint survey responses, purchase decisions, and field experiment outcomes. Recent advances in large language models (LLMs) and other AI systems offer inexpensive auxiliary data, but introduce a new challenge: AI outputs are not direct observations of the target outcomes, but could involve high-dimensional representations with complex and unknown relationships to human labels. Conventional methods leverage AI predictions as direct proxies for true labels, which can be inefficient or unreliable when this relationship is weak or misspecified. We propose Generative Augmented Inference (GAI), a general framework that incorporates AI-generated outputs as informative features for estimating models of human-labeled outcomes. GAI uses an orthogonal moment construction that enables consistent estimation and valid inference with a flexible, nonparametric relationship between LLM-generated outputs and human labels. We establish asymptotic normality and a key dominance result: under random labeling, GAI is optimal within a unified class of debiased estimators-including human-data-only estimators and state-of-the-art debiasing methods-and delivers strict improvements under a mild informativeness condition. Even when the labeled sample is not representative of the target population, an extended variant of GAI still dominates the weighted human-data-only estimator. Empirically, GAI outperforms benchmarks across diverse marketing research settings. In a conjoint analysis, it halves estimation error and reduces human labeling requirements by over 75%. In a pricing study, it consistently outperforms alternative estimators when all methods receive identical auxiliary inputs. In a health insurance study, it saves over 90% of labels while preserving decision accuracy.

Accurate temperature estimation of pouch cells with indirect liquid cooling is essential for optimizing battery thermal management systems for transportation electrification. However, it is challenging due to the computational expense of finite element simulations and the limitations of data-driven models. This paper presents a physics-informed machine learning (PIML) framework for the efficient and reliable estimation of steady-state temperature profiles. The PIML approach integrates the governing heat transfer equations directly into the neural network's loss function, enabling high-fidelity predictions with significantly faster convergence than purely data-driven methods. The framework is evaluated on a dataset of varying cooling channel geometries. Results demonstrate that the PIML model converges more rapidly and achieves markedly higher accuracy, with a 49.1% reduction in mean squared error over the data-driven model. Validation against independent test cases further confirms its superior performance, particularly in regions away from the cooling channels. These findings underscore the potential of PIML for surrogate modeling and design optimization in battery systems.

Accurate temperature field prediction in metal additive manufacturing (AM) is essential for understanding the process-structure-performance relationship. While prior studies have explored generalization to unseen process conditions, they often require extensive datasets, costly retraining, or pre-training. Generalization across different materials also remains relatively unexplored due to the challenges posed by distinct material-dependent thermal behaviors. This paper introduces a parametric physics-informed neural network (PINN) framework for generalization across unseen materials without labeled data, retraining, or pre-training. The framework adopts a decoupled parametric PINN architecture that separately encodes material properties and spatiotemporal coordinates, fusing them through conditional modulation to better align with the multiplicative role of material parameters in the governing equation and boundary conditions. Physics-guided output scaling derived from Rosenthal's analytical solution and a hybrid optimization strategy are further incorporated to enhance physical consistency, training stability, and convergence. Experiments with numerical simulations across diverse metal alloys, including both in-distribution and out-of-distribution cases, demonstrate effective generalizability along with superior training efficiency. Specifically, the proposed framework achieved up to a 64.2% reduction in relative L2 error compared to the non-parametric baseline while surpassing its performance within only 4.4% of the baseline training epochs. Ablation studies clarify each component's contribution and scrutinize the severe training instability prevalent in conventional parametric PINNs. Overall, the proposed framework provides an efficient and scalable material-agnostic solution for temperature field modeling, contributing to more flexible and practical deployment in metal AM.

Modern datacenters increasingly rely on low-power, single-slot inference accelerators to balance performance, energy efficiency, and rack density constraints. The NVIDIA T4 GPU has become widely deployed due to strong performance per watt and mature software support. Its successor, the NVIDIA L4 GPU, introduces improvements in Tensor Core throughput, cache capacity, memory bandwidth, and parallel execution capability. However, limited empirical evidence quantifies the practical inference performance gap between these two generations under controlled and reproducible conditions. This work introduces DEEP-GAP, a systematic evaluation extending the GDEV-AI methodology to GPU inference. Using identical configurations and workloads, we evaluate ResNet18, ResNet50, and ResNet101 across FP32, FP16, and INT8 precision modes using PyTorch and TensorRT. Results show that reduced precision significantly improves performance, with INT8 achieving up to 58x throughput improvement over CPU baselines. L4 achieves up to 4.4x higher throughput than T4 while reaching peak efficiency at smaller batch sizes between 16 and 32, improving latency-throughput tradeoffs for latency-sensitive workloads. T4 remains competitive for large batch workloads where cost or power efficiency is important. DEEP-GAP provides practical guidance for selecting precision modes, batch sizes, and GPU architectures for modern inference deployments.

Generating synthesizable Verilog for large, hierarchical hardware designs remains a significant challenge for large language models (LLMs), which struggle to replicate the structured reasoning that human experts employ when translating complex specifications into RTL. When tasked with producing hierarchical Verilog, LLMs frequently lose context across modules, hallucinate interfaces, fabricate inter-module wiring, and fail to maintain structural coherence - failures that intensify as design complexity grows and specifications involve informal prose, figures, and tables that resist direct operationalization. To address these challenges, we present VeriGraphi, a framework that introduces a spec-anchored Knowledge Graph as the architectural substrate driving the RTL generation pipeline. VeriGraphi constructs a HDA, a structured knowledge graph that explicitly encodes module hierarchy, port-level interfaces, wiring semantics, and inter-module dependencies as first-class graph entities and relations. Built through iterative multi-agent analysis of the specification, this Knowledge Graph provides a deterministic, machine-checkable structural scaffold before code generation. Guided by the KG, a progressive coding module incrementally generates pseudo-code and synthesizable RTL while enforcing interface consistency and dependency correctness at each submodule stage. We evaluate VeriGraphi on a benchmark of three representative specification documents from the National Institute of Standards and Technology and their corresponding implementations, and we present a RV32I processor as a detailed case study to illustrate the full pipeline. The results demonstrate that VeriGraphi enables reliable hierarchical RTL generation with minimal human intervention for RISC-V, marking a significant milestone for LLM-generated hardware design while maintaining strong functional correctness.

As speech language models (SLMs) transition from personal devices into shared, multi-user environments, their responses must account for far more than the words alone. Who is speaking, how they sound, and where the conversation takes place can each turn an otherwise benign request into one that is unsafe, unfair, or privacy-violating. Existing benchmarks, however, largely focus on basic audio comprehension, study individual risks in isolation, or conflate content that is inherently harmful with content that only becomes problematic due to its acoustic context. We introduce VoxSafeBench, among the first benchmarks to jointly evaluate social alignment in SLMs across three dimensions: safety, fairness, and privacy. VoxSafeBench adopts a Two-Tier design: Tier1 evaluates content-centric risks using matched text and audio inputs, while Tier2 targets audio-conditioned risks in which the transcript is benign but the appropriate response hinges on the speaker, paralinguistic cues, or the surrounding environment. To validate Tier2, we include intermediate perception probes and confirm that frontier SLMs can successfully detect these acoustic cues yet still fail to act on them appropriately. Across 22 tasks with bilingual coverage, we find that safeguards appearing robust on text often degrade in speech: safety awareness drops for speaker- and scene-conditioned risks, fairness erodes when demographic differences are conveyed vocally, and privacy protections falter when contextual cues arrive acoustically. Together, these results expose a pervasive speech grounding gap: current SLMs frequently recognize the relevant social norm in text but fail to apply it when the decisive cue must be grounded in speech. Code and data are publicly available at: https://amphionteam.github.io/VoxSafeBench_demopage/

Objective: Post-traumatic epilepsy (PTE) is a debilitating neurological disorder that develops after traumatic brain injury (TBI). Early prediction of PTE remains challenging due to heterogeneous clinical data, limited positive cases, and reliance on resource-intensive neuroimaging data. We investigate whether routinely collected acute clinical records alone can support early PTE prediction using language model-based approaches. Methods: Using a curated subset of the TRACK-TBI cohort, we developed an automated PTE prediction framework that implements pretrained large language models (LLMs) as fixed feature extractors to encode clinical records. Tabular features, LLM-generated embeddings, and hybrid feature representations were evaluated using gradient-boosted tree classifiers under stratified cross-validation. Results: LLM embeddings achieved performance improvements by capturing contextual clinical information compared to using tabular features alone. The best performance was achieved by a modality-aware feature fusion strategy combining tabular features and LLM embeddings, achieving an AUC-ROC of 0.892 and AUPRC of 0.798. Acute post-traumatic seizures, injury severity, neurosurgical intervention, and ICU stay are key contributors to the predictive performance. Significance: These findings demonstrate that routine acute clinical records contain information suitable for early PTE risk prediction using LLM embeddings in conjunction with gradient-boosted tree classifiers. This approach represents a promising complement to imaging-based prediction.

Purpose. Athlete monitoring is constrained by small cohorts, heterogeneous biomarker scales, limited feasibility of repeated sampling, and the lack of reliable injury ground truth. These limitations reduce the interpretability and utility of traditional univariate and binary risk models. This study addresses these challenges by proposing an unsupervised multivariate framework to identify latent physiological states in athletes using real data. Methods. We propose a modular computational framework that operates in the joint biomarker space, integrating preprocessing, clinical safety screening, unsupervised clustering, and centroid-based physiological interpretation. Profiles are learned exclusively from amateur soccer players during a competitive microcycle. Synthetic data augmentation evaluates robustness and scalability. Ward hierarchical clustering supports monitoring and etiological differentiation, while Gaussian Mixture Models (GMM) enable structural stability analysis in high-dimensional settings. Results. The framework identifies coherent profiles that distinguish mechanical damage from metabolic stress while preserving homeostatic states. Synthetic data augmentation demonstrates feasibility and detection of latent silent risk phenotypes typically missed by univariate monitoring. Structural analyses indicate robustness under augmentation and higher-dimensional settings. Conclusion. The framework enables interpretable identification of latent physiological states from multivariate biomarker data without injury labels. By distinguishing mechanisms and revealing silent risk patterns not captured by conventional monitoring, it provides actionable insights for individualized athlete monitoring and decision making.

Accurate prediction of future risk and disease progression in sepsis is clinically important for early warning and timely intervention in intensive care. However, short-window sepsis prediction remains challenging, because shorter observation windows provide limited historical evidence, whereas longer prediction horizons reduce the number of patient trajectories with valid future supervision. To address this problem, we propose CSRA, a Controlled Spectral Residual Augmentation framework for short-window multi-system ICU time series. CSRA first groups variables by clinical systems and extracts system-level and global representations. It then performs input-adaptive residual perturbation in the spectral domain to generate structured and clinically plausible trajectory variations. To improve augmentation stability and controllability, CSRA is trained end-to-end with the downstream predictor under a unified objective, together with anchor consistency loss and controller regularization. Experiments on a MIMIC-IV sepsis cohort across multiple downstream models show that CSRA is consistently competitive and often superior, reducing regression error by 10.2% in MSE and 3.7% in MAE over the non-augmentation baseline, while also yielding consistent gains on classification. CSRA further maintains more favorable performance under shorter observation windows, longer prediction horizons, and smaller training data scales, while also remaining effective on an external clinical dataset~(ZiGongICUinfection), indicating stronger robustness and generalizability in clinically constrained settings.

Catastrophic forgetting remains a fundamental challenge in continual learning, in which models often forget previous knowledge when fine-tuned on a new task. This issue is especially pronounced in class incremental learning (CIL), which is the most challenging setting in continual learning. Existing methods to address catastrophic forgetting often sacrifice either model interpretability or accuracy. To address this challenge, we introduce ClassIncremental Concept Bottleneck Model (CI-CBM), which leverage effective techniques, including concept regularization and pseudo-concept generation to maintain interpretable decision processes throughout incremental learning phases. Through extensive evaluation on seven datasets, CI-CBM achieves comparable performance to black-box models and outperforms previous interpretable approaches in CIL, with an average 36% accuracy gain. CICBM provides interpretable decisions on individual inputs and understandable global decision rules, as shown in our experiments, thereby demonstrating that human understandable concepts can be maintained during incremental learning without compromising model performance. Our approach is effective in both pretrained and non-pretrained scenarios; in the latter, the backbone is trained from scratch during the first learning phase. Code is publicly available at github.com/importAmir/CI-CBM.

As a classic vision task, anomaly detection has been widely applied in industrial inspection and medical imaging. In this task, data scarcity is often a frequently-faced issue. To solve it, the few-shot anomaly detection (FSAD) scheme is attracting increasing attention. In recent years, beyond traditional visual paradigm, Vision-Language Model (VLM) has been extensively explored to boost this field. However, in currently-existing VLM-based FSAD schemes, almost all perform anomaly inference only by pairwise feature matching, ignoring structural dependencies and global consistency. To further redound to FSAD via VLM, we propose a Heterogeneous Hypergraph Vision-Language Reasoning (H2VLR) framework. It reformulates the FSAD as a high-order inference problem of visual-semantic relations, by jointly modeling visual regions and semantic concepts in a unified hypergraph. Experimental comparisons verify the effectiveness and advantages of H2VLR. It could often achieve state-of-the-art (SOTA) performance on representative industrial and medical benchmarks. Our code will be released upon acceptance.

We study how depth, finite precision, state dimension, and chain-of-thought (CoT) affect the expressive power of multi-layer state-space models (SSMs). For the explicit-table $K$-function-composition problem, a canonical benchmark for sequential information propagation, we prove that any $L$-layer SSM solving $(L+3)$-function composition must satisfy $d^2p=\Omega(N/L^3)$, where $d$ is the state dimension and $p$ is the per-scalar precision. Conversely, $K$-function composition is solved exactly by a $(K+1)$-layer generalized SSM with $d=1$ and $p=\Theta(\log N)$. This gives a worst-case depth hierarchy for this formal problem family. We then distinguish post-input reasoning, in which all thought tokens are generated after the input, from input-interleaved reasoning, in which thought tokens may be inserted while the input stream is being read. Post-input reasoning does not circumvent our communication-based lower-bound pipeline, whereas input-interleaved reasoning admits bidirectional simulations with general deterministic one-pass streaming algorithms at the granularity of persistent memory. Finally, width and precision are not interchangeable under exact step-preserving simulation in the base affine-state model, but become interchangeable through the streaming-memory characterization once input-interleaved reasoning is allowed.

Synthetic augmentation is increasingly used to mitigate data scarcity in financial machine learning, yet its statistical role remains poorly understood. We formalize synthetic augmentation as a modification of the effective training distribution and show that it induces a structural bias–variance trade-off: while additional samples may reduce estimation error, they may also shift the population objective whenever the synthetic distribution deviates from regions relevant under evaluation. To isolate informational gains from mechanical sample-size effects, we introduce a size-matched null augmentation and a finite-sample, non-parametric block permutation test that remains valid under weak temporal dependence. We evaluate this framework in both controlled Markov-switching environments and real financial datasets, including high-frequency option trade data and a daily equity panel. Across generators spanning bootstrap, copula-based models, variational autoencoders, diffusion models, and TimeGAN, we vary augmentation ratio, model capacity, task type, regime rarity, and signal-to-noise. We show that synthetic augmentation is beneficial only in variance-dominant regimes, such as persistent volatility forecasting-while it deteriorates performance in bias-dominant settings, including near-efficient directional prediction. Rare-regime targeting can improve domain-specific metrics but may conflict with unconditional permutation inference. Our results provide a structural perspective on when synthetic data improves financial learning performance and when it induces persistent distributional distortion.

At matched accuracy, open-weight LLMs differ substantially in the shape of their error severity distribution – a difference invisible to the scalar error rate. Hallucination benchmarks report a single error count and treat all errors as equivalent, yet a wrong date and a fabricated court ruling differ by orders of magnitude. We introduce Errorquake-10k, a 10,000-query benchmark scoring each response on a continuous 0-4 severity scale across 8 domains and 5 difficulty tiers, and we fit per-model severity distributions for 21 open-weight models. For each model we estimate a severity distribution index (b, the Gutenberg-Richter upper-tail slope) with 95% bootstrap confidence intervals. Headline: across the 210 model pairs, 85 have disjoint 95% b confidence intervals at matched accuracy (|Delta epsilon| < 0.05) on human-consensus scoring, e.g. deepseek-v3.2 vs. ministral-14b at epsilon = 0.586 and Delta b = 0.47. A 519-item three-rater human validation study confirms measurement reliability (ICC(2,k=3) = 0.85), validates the LLM-judge ranking (rho = 0.89), and confirms the dense-model scaling correlation on human data (rho_s = -0.86). We prove a Non-Reducibility Theorem showing that severity profile and error rate are informationally non-redundant (I(b; model | epsilon) = 1.56 bits; 64.5% of cross-model b variance is unexplained by epsilon). A severity mechanism taxonomy (kappa = 0.83) reveals that error type shifts categorically with severity: low-severity errors are retrievals (71%); high-severity errors are fabrications (39%) – and this composition differs by model size (p < 0.0001). Severity distribution should be reported alongside accuracy; it carries discriminative information that the error rate cannot.

We give a stereological theory of LLM benchmark coverage. For any suite with effective dimensionality d_eff, the visible Hausdorff distance between two convex capability profiles consistent with the same scores is bounded by epsilon + C R m^(-1/(d_eff-1)), with matching Lipschitz lower bound. Empirically, three independent leaderboards (Open LLM v2, an extended 12-benchmark suite, LiveBench) all have d_eff in [2.86, 4.80] on their competitive frontier; the structural blind spot exceeds the observed runner-up score gap by two orders of magnitude and dominates statistical noise by 52-127x. Under a chi-squared projection model, the isotropic prior is the optimistic case; across six hidden-capability priors and four ambient dimensions the simulated half-split swap rate of the top two models stays in [0.38, 0.49], and a 500-trial random visible/held-out split shows that 92% of trials swap the top-1 ranking with on average 2.83 of 5 top-5 models changing. A submodular greedy algorithm with the Nemhauser (1 - 1/e) guarantee finds a stable core of 4 benchmarks; 7 of 12 suffice for 90% coverage, and the trained subset transfers across temporal quarters with 93-97% retention. A counterfactual validation across 12 internal benchmarks and 27 Chatbot Arena categories confirms that the eigenstructure predicts which evaluations are irreplaceable (rho = -0.69, p = 0.013 for removal disruption) and which external evaluations bring new information (rho = +0.38). As a second, independent theoretical contribution, we resolve Gardner's Problem 1.5 (1995) for C^2 support functions, establishing the minimax rate Theta(R/(kappa m^(2/(D-1)))) in general dimension via optimal recovery theory on S^(D-1).

In this paper, we study the efficient numerical solution of Darcy equations in strongly heterogeneous media with high-contrast permeability and propose a hybrid framework that combines learning with multiscale numerical methods. The learning component is used for the prediction of multiscale basis functions in the mixed generalized multiscale finite element method (mixed GMsFEM), with the goal of reducing the repeated local computations required in the offline stage. Once these basis functions are predicted, the global system is assembled and the pressure field is computed by a two-grid preconditioned solver. The resulting method accelerates the costly local basis-construction stage while retaining the multiscale discretization and preconditioned iterative structure of the underlying solver. Numerical experiments on two-dimensional heterogeneous Darcy problems show that the proposed framework yields more accurate final pressure reconstruction than several representative learning-based methods and remains stable under strong heterogeneity and high-contrast coefficients. In comparison with the traditional mixed GMsFEM, its main advantage lies in the efficiency of the basis-generation stage, while the quality of the global solve is still ensured by the two-grid preconditioner. These results indicate that accelerating multiscale basis construction through learning, while preserving a mature numerical solver for the global problem, provides a viable approach for high-resolution Darcy-type simulations.

The rapid development of Language Diffusion Models (LDMs) challenges the dominant position of auto-regressive competitors in language processing. However, their flexible, any-order decoding strategies not only enable fast decoding speed but also potentially bring new trustworthiness challenges. To better understand the risks behind their pipelines, we introduce a comprehensive trustworthiness benchmark tailored to LDMs (TrustLDM), evaluating safety, privacy, and fairness across different LDM architectures with multiple categories of static post contexts. Our empirical results show that although LDMs generally exhibit strong trustworthiness with only the user prompts, their alignment behavior degrades noticeably when the malicious post contexts are attached to the masked responses. We further observe that longer contexts do not necessarily induce stronger effects, and both decoding order and generation length affect the evaluation outcomes. Finally, we propose TrustLDM-Auto, an automatic evaluation framework that leverages LDM decoding flexibility to systematically identify vulnerable configurations, revealing substantial trustworthiness weaknesses across all evaluated models and dimensions. Our work may potentially help the community build more trustworthy LDMs. Our code is available at https://github.com/PKU-ML/TrustLDM.

The AI inference industry keeps models loaded in GPU memory around the clock to avoid cold-start latency, implicitly treating idle power as a fixed cost of readiness. Yet the structure of this cost has never been empirically decomposed - and never across GPU architectures. We present the first cross-architecture measurement of idle GPU power as a function of VRAM allocation, combining 18 days of production telemetry (335,267 samples, 14 H100 GPUs) with controlled dose-response experiments on three GPU architectures spanning three memory technologies: NVIDIA H100 (HBM3, 80 GB), A100 (HBM2e, 80 GB), and L40S (GDDR6, 48 GB). We observe that idle power is piecewise constant on all three architectures: the CUDA context forces a discrete DVFS transition consuming +26-66 W over bare idle (26-50 W on HBM architectures, 66 W on GDDR6), while the marginal VRAM effect is bounded below measurement relevance ($|\beta| < 0.02$ W/GB) on every device tested. The CUDA context accounts for >98% of the parking tax regardless of memory technology. We validate this finding with a real HuggingFace model (Qwen2.5-7B) on all three architectures, confirming <0.5 W difference from empty tensors on every device, and capture cold-start power profiles during model loading. We derive a cold-start breakeven model showing energy-optimal behavior depends on request arrival rate and loading latency - not model size - with breakeven intervals of 1-5 minutes. Our results identify a constraint consistent across all tested architectures: idle-with-context power is determined by DVFS state, not memory occupancy.

Overparametrized models can exhibit an excellent generalization performance, although they should be prone to overfitting according to classical statistical theory. The discovery of the "double descent", indicating that the generalization error decreases after a certain model complexity has been reached, opened a new line of research. Robust statistics considers statistical estimation on contaminated data, which, due to assumptions that do not hold on real data, let data points appear as outliers w.r.t. the assumed "ideal" distribution, potentially severely distorting any classical estimator. We address the question whether a double descent phenomenon can be observed in a linear regression setting with contaminated training data. We compare the performance of the highly non-robust least-squares interpolation estimator with several robust alternatives. It turns out that large overparametrization indeed allows for a double descent phenomenon, resulting in a very good generalization performance of the least-squares interpolator, surpassing that of the robust alternatives.

Achieving precise, versatile whole-body character control in physics-based animation remains challenging. Recent diffusion-based policies generate rich and expressive motions but typically rely on gradient-based test-time guidance to satisfy task objectives, which is slow and can reduce robustness. We introduce NaP-Control (Navigating Diffusion Prior for Versatile and Fast Character Control), abbreviated as NaP. Our method uses reinforcement learning to manipulate the latent noise of a task-agnostic diffusion policy prior, steering it toward task-specific behaviors for fast, robust control with high motion fidelity. In contrast to methods that rely solely on offline training, NaP interacts with the environment during training to correct motions and optimize task rewards, improving success rates and enabling adaptation to challenging scenarios. By directly predicting task-optimized diffusion noise, NaP eliminates iterative guidance during denoising and enables efficient inference. Experiments show that NaP attains higher success rates and faster inference while preserving natural motion across diverse tasks.

Much of the focus in RL today is on evaluation design: building meaningful evals that serve simultaneously as benchmarks and as well-defined reward signals for post-training. Yet, many real-world tasks are governed by subjective, procedural, and domain-specific requirements that are difficult to encode as exact-match targets or open-ended preference judgments frequently used in RL pipelines today. In this work, we present AsymmetryZero, a framework for operationalizing human expert preferences as semantic evals. AsymmetryZero represents each task as a stable evaluation contract that makes grading criteria explicit: what is being graded, how each criterion is judged, and how criterion-level decisions are aggregated into a task outcome. The same contract can be executed using Inspect for model-only evaluations, as well as the Harbor Framework for agentic evaluations, enabling comparable scores and shared audit artifacts across both settings. We argue that the central challenge in post-training today is the faithful encoding of expert requirements into the evaluation itself. To that end, we present a study using Harbor that holds task contracts fixed and compares a five-model frontier jury against a five-model compact jury across four frontier-class solvers (Claude Opus 4.6, GPT-5.4, Grok-4.20, Gemini-3.1-Pro). We find that criterion-level frontier-vs-compact agreement ranges from $75.9\%$ to $89.6\%$ (strict common-subset agreement: $77.8\%$ to $92.1\%$), while compact juries exhibit substantially higher internal dissent (3–2 split rate $28.7\%$–$32.4\%$) than frontier juries ($6.1\%$–$11.5\%$). Verifier traces further show that compact juries reduce per-criterion judging cost to roughly $4.2\%$–$5.6\%$ of frontier and latency to roughly $21.7\%$–$27.1\%$, even as aggregated task-level outcomes often remain comparatively stable.

Model-based solutions for nitrous oxide (N2O) emissions from wastewater treatment plants (WWTP) are informed by operational datasets designed to control nutrient levels in liquid waste, coupled with dedicated campaigns for N2O measurements. We analysed how machine learning (ML) models predict disturbances to WWT operation and spatially variable N2O emissions. A real dataset was investigated to validate the modelling framework from N2O emissions predicted by four ML models (R2 = 0.79 - 0.89). Monitoring campaigns for N2O were simulated with a plant-wide mechanistic model to include additional sensors, site-level N2O datasets, and wastewater disturbances (n = 16). ML models were highly accurate (0.97 +- 0.02, n = 80), but the feature importance depended on the model, the scenario and the N2O measurement scale (reactor vs. WWTP). We argue that N2O soft sensor model predictions are limited to the measuring location and the methodological uncertainty of the dataset, which affect the interpretability of the model. Lastly, the analysis of the mechanistic model structure exposed interactions between autotrophic and heterotrophic pathways over nitric oxide which can overestimate aerobic nitrite production and bias the N2O pathway contributions.