The universal approximation property uniformly with respect to weakly compact families of measures is established for several classes of neural networks. To that end, we prove that these neural networks are dense in Orlicz spaces, thereby extending classical universal approximation theorems even beyond the traditional $L^p$-setting. The covered classes of neural networks include widely used architectures like feedforward neural networks with non-polynomial activation functions, deep narrow networks with ReLU activation functions and functional input neural networks.
Most modern optimizers form their momentum as an exponential moving average (EMA) of past gradients, forgetting every direction at one fixed rate. However, the inputs a deep network sees during training can be highly anisotropic, with a few directions queried frequently while most are seen rarely. Preconditioning methods address this anisotropy by wrapping extra processing around this buffer and leave the momentum update itself unchanged. We propose Activation-Keyed Momentum (AK-Momentum), which builds direction-awareness into the momentum update rule. The gradient of a linear layer splits into an input activation that acts as a $\textit{key}$ and an output-side error that acts as a $\textit{value}$. Keying on that activation, AK-Momentum updates the momentum buffer by the canonical delta rule, so each direction is forgotten at a rate set by how often it appears. We prove that it is a valid momentum, that it applies the input-side curvature correction without matrix inversion, and that it clears stale directions faster than EMA under both a fixed and a drifting optimum. It is a drop-in replacement for the momentum buffer of any optimizer, its coefficient transfers across widths under $μ$P, and its extra compute stays between $22.2\%$ and $25.0\%$ of a gated-MLP block's linear cost with no persistent memory. In FineWeb-Edu pretraining, AdamW with AK-Momentum (AK-AdamW) reaches AdamW's validation loss in up to $46.64 \pm 4.52\%$ fewer steps at 67M and $22.14 \pm 0.81\%$ at 370M over three seeds, and the gain persists at 1B on a Chinchilla-optimal budget. A Muon baseline tuned under the same protocol sits above AK-AdamW at both language-model scales, and the gain holds for SGD, ResNet-18, and ViT-Tiny on CIFAR-10. Training-time diagnostics confirm the predicted mechanism, better gradient tracking and healthier input directions. Code is available at https://github.com/euijinh/AK-Momentum.
Roman Rausch, David Jansen, Sukhbinder Singh, Román Orús
Low-rank factorization of weight matrices is a promising approach to compress large language models (LLMs), but its practical effectiveness depends on allocating ranks across layers and efficiently representing the resulting factors. We introduce three complementary methods for improving SVD-based LLM compression. First, FermiGrad formulates global layer-wise rank allocation as a differentiable optimization problem by replacing discrete singular-value truncation with continuous Fermi gates. Because FermiGrad optimizes only rank parameters over precomputed low-rank factors, it avoids repeated activation decompositions and backpropagation through SVD operations. Second, the Targeted Compressor extends FermiGrad with a task-specific objective that directs limited model capacity toward a selected domain. At 60% parameter density for Llama-3.1-8B-Instruct, SciQ-targeted compression achieves 60% MMLU accuracy, exceeding the reported quantized Dobi-SVD result by 10 percentage points, while deliberately trading performance on unrelated tasks for domain specialization. Third, PivGa provides a lossless secondary compression of low-rank factors by exploiting the gauge freedom in their parametrization. The proposed framework is modular, computationally efficient, and applicable at scale, as demonstrated by compressing Llama-3.3-70B-Instruct to 50% parameter density.
We show that the transformer architecture emerges as the score of last-token laws in a natural statistical model of language, where sentences are empirical measures of tokens, a language is a law on $P(S^d)$, and our prior belief is that sentences carry no structure at all. By the maximum-entropy principle, we select the law closest to this prior in relative entropy among those reproducing an observed sentence-level statistic. We discuss the relevance of the statistic in the posterior interaction between tokens and we build a statistic, able to capture that interaction, with Luce's choice axiom truncated at the lowest relational order. This procedure gives the softmax choice rule, with the attention cross-moment as its mean parameter. The query--key matrix is then the Lagrange multiplier of a maximum-entropy problem inside each sentence, and the value matrix the multiplier of the constraint over the language, so that the two are untied. Under the optimal entropic law we then show that a transformer head appears in the score of the last-token law. For ordered sentences, causal masking follows from the masked reference choice rule, and with a next-token cross-moment the score equals one causally masked attention head. We also discuss multi-head attention, weight-tied unembedding, and a heuristic reading of residual blocks as projected Langevin steps.
Tabular foundation models use in-context learning to estimate conditional expected outcomes and average treatment effects for specific groups of individuals. Using these estimates to guide scientific decisions requires reliable uncertainty quantification. The challenge is to account simultaneously for two sources of uncertainty: limited-context bias and context variation arising from the small labeled sample supplied at inference. Predictive spread does not directly quantify both sources of uncertainty, and we observe that the ordinary bootstrap may substantially overestimate context variation, whereas subsampling may underestimate it. We introduce TabCUQ, a perturbation-based framework that injects noise into observed outcomes and applies the foundation model to the perturbed contexts. TabCUQ estimates context variation using wild-bootstrap dispersion and accounts for limited-context bias using the maximum deviation between perturbed and original estimates. For a kernel-regression analogue whose estimation error admits an exact decomposition into these two components, we show that our framework provides asymptotically valid pointwise coverage. In simulations, TabCUQ substantially improves coverage, with average interval widths typically about twice those of an oracle bias-aware benchmark. Across seven real-world tasks, it also improves coverage of reference functions learned from much larger samples.
This study proposes a statistically grounded framework for real-time win probability evaluation and player assessment in score-based team sports, based on minute-by-minute cumulative box-score data. We introduce a continuous dominance indicator (T-score) that maps final scores to real values consistent with win/lose outcomes, and formulate it as a time-evolving stochastic representation (T-process) driven by standardized cumulative statistics. This structure captures temporal game dynamics and enables sequential, analytically tractable updates of in-game win probability. Through this stochastic formulation, competitive advantage is decomposed into interpretable statistical components. Furthermore, we define a latent contribution index, STATS X, which quantifies a player's involvement in favorable dominance intervals identified by the T-process. This allows us to separate a team's baseline strength from game-specific performance fluctuations and provides a coherent, structural evaluation framework for both teams and players. While we do not implement AI methods in this paper, our framework is positioned as a foundational step toward hybrid integration with AI. By providing a structured time-series representation of dominance with an explicit probabilistic interpretation, the framework enables flexible learning mechanisms and incorporation of high-dimensional data, while preserving statistical coherence and interpretability. This work provides a basis for advancing AI-driven sports analytics.
Pekka Malo, Lauri Viitasaari, Antti Suominen, Eeva Vilkkumaa, Olli Tahvonen
In many control problems, actions consume a resource budget, such as a wear allowance, and the probability of exhausting it within a given horizon must stay below a tolerance. A simple, widely used approach is to augment the state with the remaining budget, subtract an interior barrier from the utility and penalize exhaustion, so that policy-gradient methods can be applied unchanged. We clip the combined one-step reward from below at a level m to keep the returns of exploratory policies finite. The clip level is a price of failure risk, since the clipped value decomposes exactly into the utility before failure, the cost of the interior margin and m times a geometric failure risk. Consequently, an approximately optimal policy maximizes utility net of margin cost among policies with no greater risk. Its risk exceeds that of any comparison policy by at most order 1/m and can be certified from its value or from validation data. Our main result proves exponential convergence of the population policy-gradient flow for every parameter regularizer that is convex along generalized geodesics and sufficiently strong relative to the price. This extends the mean-field analysis from relative entropy to weight decay and to finite networks. Relative entropy adds parameter noise and a pull toward a prior, whereas weight decay provides only the pull. Action regularization bounds changes in failure risk by the relative entropy to a reference policy. The bounds are evaluated on an exact maintenance benchmark, a fuel-tank controller, a solvable flow and a benchmark-suite task.
Szabolcs Szentpéteri, Péter Kovács, Balázs Csanád Csáji
This paper studies robust uncertainty quantification for separable nonlinear least squares (SNLLS) problems. We propose two identification methods, (i) Residual-Perturbed Gradient (RPG) and (ii) Global-Fit-Gap (GFG), to construct joint distribution-free confidence regions for both the linear and nonlinear parameters. By exploiting assumed group invariance of the noise (such as symmetry or exchangeability), we apply random perturbations (such as sign-changes or permutations) to build the regions based on resampling and ranking. We prove that both methods guarantee non-asymptotically exact coverage (type I error). Finally, we illustrate the constructions through two numerical experiments: (i) identifying generalized finite impulse response (GFIR) systems using Laguerre polynomials with an adjustable decay rate and (ii) fitting Gaussian radial basis function (RBF) models with variable knots and widths.
A streaming estimator must compute from the evidence that fits in memory. We determine when storing an accurate mean is easier than computing it. For Gaussian mean estimation over the Euclidean unit ball in $\mathbb{R}^d$, we establish the minimax rate under squared error with $K$ persistent states, a known horizon, and a public time index: $\min\{1, K^{-2/d}+\frac{d}{n} (1+\frac{\log(ed)}{1+(\log K)/d})\}$. Unrestricted computation followed by $K$-level encoding has rate $\min\{1,d/n+K^{-2/d}\}$. At a fixed sufficiently small error, statistically optimal sample size therefore requires order $d\log d$ working bits, while order $d$ bits suffice to store the answer. The lower bound tracks repeated state compression through an entropy potential obtained by Gaussian smoothing. A matching construction combines a multiscale modular counter with workspace reused across coordinate groups. In one dimension, it attains the terminal rate $n^{-1}+K^{-2}$. Any fixed confidence requirement that the realized loss not exceed the zero baseline preserves these rates. Increasing confidence creates a further memory requirement: at fixed nontrivial risk in every fixed dimension, failure probability $e^{-γn}$ requires and permits order $\log n$ working states for sufficiently small fixed $γ>0$. In the scalar problem, a fixed terminal alphabet suffices under the same protection. These bounds separate the memory needed to represent an answer from the memory needed to accumulate evidence and validate a decision.
Julien Grand-Clément, Joon Kwon, Francesco Orabona
A central topic in online and stochastic optimization is to design algorithms that do not require tuning of hyperparameters. Successful strategies in online learning include AdaGrad-style learning rates, Conic Blackwell algorithms, and parameter-free algorithms. However, these classes require different, often difficult, analyses. We develop a unifying family of Online Convex Optimization (OCO) algorithms that require no user-tuned learning rate on convex domains that may be unbounded. The construction lifts the domain to a cone by adding one scalar coordinate, an idea going back to Blackwell (1954). The lifted subgradients are orthogonal to the corresponding lifted decisions, leaving the comparator scale free to optimize in the regret analysis. Applying Follow-the-Regularized-Leader (FTRL) and Online Mirror Descent (OMD) on the cone recovers Conic Blackwell Algorithm (CBA) and Conic Blackwell Algorithm Plus (CBA$^+$) on bounded domains in Hilbert spaces and extends them to finite-dimensional normed spaces with a strongly convex squared norm. For arbitrary nonempty closed convex domains in these spaces, a known gradient bound yields parameter-free regret guarantees. Surprisingly, on an unconstrained Hilbert space, the construction recovers the Krichevsky--Trofimov (KT) coin-betting algorithm exactly. We also give a parameter-free algorithm for unconstrained domains that requires no gradient bound.
Exhaustively evaluating many large language models (LLMs) on a large suite of benchmarks is expensive. We cast benchmarking as finite-population inference and, under a fixed query budget, seek tight confidence intervals (CIs) for model accuracy with asymptotic frequentist coverage. We propose Factorized Active Querying (FAQ), which (a) leverages historical information through a Bayesian factor model; (b) adaptively selects questions using a hybrid variance-reduction/active-learning sampling policy; and (c) maintains validity through Pro-Active Inference---a finite-population extension of active inference (Zrnic & Candès, 2024) that enables direct question selection while preserving asymptotic frequentist coverage. In our experiments, FAQ delivers up to $5\times$ effective sample size gains over strong baselines on two benchmark suites, across varying historical-data missingness levels, with negligible computational overhead: this means that it matches the CI width of uniform sampling while using up to $5\times$ fewer queries. We release our source code and our curated datasets to support reproducibility and future research.
We show that established performance metrics in binary classification, such as Matthews' correlation coefficient (MCC), Cohen's $κ$, the F-score or the Jaccard similarity coefficient are not robust to class imbalance in the sense that if the proportion of the minority class tends to $0$, the true positive rate (TPR) of the Bayes classifier under these metrics {becomes very small or even tends to $0$}. Thus, in imbalanced classification problems, these metrics favour classifiers which ignore the minority class. To alleviate this issue we introduce robustified modifications of the MCC, of Cohen's $κ$ and of the F-score with an additional tuning parameter which allows to adapt the amount of robustness against class imbalance. As theoretical guarantee we show that the Bayes-optimal classifier for these robustified performance metrics, when expressed in terms of the density ratio $f_1/f_0$ of the class-conditional densities $f_i$, has a threshold parameter which is upper-bounded in terms of the tuning parameters. Therefore, even in strongly imbalanced settings, the TPR associated to this classifier will be bounded away from $0${, provided that $f_1/f_0$ exceeds this bound with positive probability under the positive class}. We numerically illustrate the behaviour of the various performance metrics and the effect of the tuning parameters in simulations as well as on a credit default data set. We also discuss connections to the receiver operating characteristic and precision-recall curves, which provide an alternative perspective on the proposed notion of robustness, and give recommendations on how to combine their usage with performance metrics.
Chenxi Li, Yao Shu, Siyang Gao, Shuang Qiu, Guangxin Jiang
Many applications of multi-objective reinforcement learning (MORL) require policies that balance conflicting objectives while satisfying safety constraints. We study safe MORL with zero constraint violation throughout learning, covering both cumulative and stepwise constraints. A key challenge is to explore sufficiently under unknown dynamics and constraint costs without violating the constraints. For a given preference over objectives, we develop an algorithm to learn a safe Pareto optimal policy by combining optimistic reward estimates with pessimistic cost bounds. With high probability, every policy executed over $T$ episodes incurs zero constraint violation. The algorithm returns a safe Markov policy induced by the averaged occupancy measure and achieves an $\widetilde{O}(1/\sqrt{T})$ suboptimality gap under smooth Tchebycheff scalarization. We further develop a preference-free algorithm that first explores safely without a prespecified preference and then computes a safe policy for any given preference without further interaction. Over $T$ exploration episodes and $K$ planning iterations, it retains the same safety guarantee and achieves an $\widetilde{O}(1/\sqrt{T}+1/\sqrt{K})$ suboptimality gap. Both algorithms require knowledge of a policy that strictly satisfies the constraints and its costs. We then relax these requirements, incurring at most constant cumulative violation. To our knowledge, we provide the first sample-efficient MORL algorithms that guarantee safety for every executed policy.
LLM agentic workflows increasingly employ different models for specialized roles. Users may prefer different trade-offs between task quality and deployment cost, sometimes also considering latency. Finding good assignments incurs a separate search cost from evaluating multiple candidates. We introduce Cost-Coupled Gittins (CC-Gittins), a multi-objective extension of single-objective cost-aware Gittins search for Pareto recommendation. It exploits the coupling between evaluation and deployment costs: assignments that cost more to deploy also cost more to evaluate. The search alternates between seeking higher quality and lower deployment cost, reusing each evaluation to inform both searches. After each batch, CC-Gittins recommends assignments that are nondominated under uncertainty-adjusted estimates of quality and deployment cost. We evaluate CC-Gittins on eight quality--cost matrices from workflows with two to five roles and up to 1,681 candidate assignments. We also evaluate the method with latency as a third objective. Across the eight benchmarks, CC-Gittins recovers a substantial fraction of the true Pareto frontier at roughly 10% of the exhaustive-evaluation cost, and recovers most of the frontier with few false positives by 30%.
Nan Lu, Ethan Lee, Ruicheng Ao, James M. Robins, David Simchi-Levi, Junwei Lu
We study offline inference for the optimal value in reinforcement learning under finite state and action spaces. Two new nuisances are derived as fixed points of a self-induced Bellman equation, in which we approximate the maximum Bellman operator by its softmax correspondence. We propose a debiased estimator through the Neyman orthogonality and establish its asymptotic normality under diverging horizons even when the behavior policy changes with time, as long as the nuisances have the statistical rates that can be achieved by many machine learning methods. We provide a concrete estimating procedure for these nuisances and show they can lead to valid inference. Synthetic experiments validate the numerical performance of our inference method, and we implement it in real-life decision-making problems, including bike repositioning and AI agentic tool use.
Toni J. B. Liu, Jiajun Bao, Gurbir Arora, Mitchell Ostrow, Yizhou Liu, Raphaël Sarfati, Nicolas Boullé, Christopher J. Earls
What does a language model predict when it has few clues?The answer lies in its unembedding geometry: a single direction of the unembedding matrix encodes the unigram distribution of the training corpus, which serves as the Bayesian prior the model falls back on when uncertain. This structure, termed the direction of ignorance, appears across all model families examined (Llama, Qwen, Gemma, and Pythia), ranging from 0.4B to 405B parameters. Projecting the final prediction state onto this direction yields a per-token prior loading factor $λ$, which declines steadily as the context becomes more informative. Formally, the same projection decomposes the prediction state into two orthogonal vectors interpretable as the two components of a tempered Bayesian update: a unigram prior raised to the exponent $λ$ and a context-driven likelihood. This geometric probabilistic connection makes $λ$ meaningfully comparable across model sizes and families. Larger models generally exhibit lower prior reliance, and reasoning models exhibit lower reliance than their base counterparts. Finally, we show that $λ$ serves as a causal, inference-time control: reducing it can promote context-driven continuations and improve Llama-3.2-3B-Instruct on GSM8K, whereas amplifying it or applying matched perturbations along random directions degrades performance.
A long-lived assistant that brings up memories without being asked must decide at every turn whether anything in its store is worth an interruption. We show that this decision is a multiple-testing problem whose difficulty grows with the store size N. On streaming replays of LongMemEval, a similarity threshold tuned to 5% false alarms at N=8 reaches 84-100% at N=2048 across embedding models and cue types, and LoCoMo and LLM judges show it too. We prove that a gate which thresholds the best of N independent irrelevant scores at false-alarm rate alpha operates at a per-item false-positive rate of about -ln(1-alpha)/N, which pins its hit rate to the far left of the single-comparison ROC curve and, under regular variation, makes it regularly varying in N. A finite-pool form of this bound predicts the hit rate of a label-tuned threshold on held-out questions to within 6.1 points. At matched false alarms the deployed score recalls 14-26 points more than raw cosine. A cue-swap conformal gate, which ranks the current cue among other cues on the same store, has a finite-sample false-alarm guarantee under exchangeability. With a background drawn from the user's unfiltered history it keeps false alarms at 3-6% on LongMemEval at the hit rate of a label-tuned threshold; on LoCoMo it overshoots or turns conservative depending on where its reference cues come from. In a chronological replay with the user's own earlier messages as references, it speaks on 6-8% of session-opening turns, against 57% for the deployed gate at N=2048, at about 50 ms per decision on one CPU thread. A cross-encoder reranker pays the same price of scale, and on a prospective-memory benchmark, similarity gates recall at most 9% of constraints at controlled false alarms. One human and two LLM annotators judge most false alarms unhelpful for a response. Code and data: https://github.com/RiverChenStillCompiling/price-of-scale
We study satisficing regret minimization in bandits with knapsacks (BwK), where a learner sequentially selects randomized policies, i.e., distributions over arms, to obtain mean rewards above a satisficing threshold as frequently as possible while respecting finite resource budgets. The performance is measured by satisficing regret, the cumulative deficits of the selected policies' mean rewards below this threshold. Resource constraints introduce two difficulties: an algorithm can incur linear satisficing regret even with small standard regret as deficits to the threshold cannot be compensated, and random consumption can cause early resource depletion even when policies are feasible in expectation. To address these challenges, we propose SELECTwK, a general algorithmic template that uses a BwK learning oracle algorithm with sub-linear expected (standard) regret upper bound to search for and test candidate satisficing policies (i.e., policies whose mean rewards exceed satisficing threshold and whose expected resource consumption is feasible over the time horizon). Our regret analysis, supported by an LP-duality argument, establishes that SELECTwK attains constant expected satisficing regret when a satisficing policy exists. We further introduce a stronger regret condition that bounds the expected sum of positive reward deficits to the optimal one at each time step, under which SELECTwK achieves a sharper constant bound together with a sub-linear bound in the time horizon. As a complement, SELECTwK still preserves the oracle's sub-linear expected standard regret bound if the satisficing threshold is unattainable. Finally, we establish the oracle conditions and corresponding satisficing regret bound for existing finite-armed BwK and dynamic pricing algorithms, and conduct numerical experiments on dynamic pricing with limited inventory to demonstrate the empirical performance of SELECTwK.
Nonlinear structural system identification remains challenging when governing physics is only partially known and measurements are sparse. Deep Bayesian filtering addresses these challenges by combining learnable dynamical representations with probabilistic state-space models, physical inductive biases, and uncertainty-aware inference. This paper presents, for the first time, a unified tutorial treatment of deep Bayesian filtering for nonlinear structural system identification, connecting model formulation, learning, posterior inference, and evaluation. Amortized variational inference and differentiable Kalman filtering are introduced as complementary strategies within this broader framework. The Deep Markov Model (DMM) provides a data-driven reference, while the Physics-guided Deep Markov Model (PgDMM) and Neural Extended Kalman Filter (Neural EKF) serve as illustrative instances through which methodological choices and their implications are examined. Dynamic state estimation and joint state and parameter estimation are demonstrated on a three-dimensional shear-frame benchmark with hysteretic Bouc-Wen connections and sparse acceleration measurements. The comparative analysis yields new observations concerning the relationships among reconstruction accuracy, predictive fidelity, physical interpretability, and estimation robustness. Physical structure improves latent-state interpretability without necessarily ensuring faithful predictive dynamics, while comparable typical reconstruction accuracy can conceal substantial differences in estimation robustness. These findings motivate practical guidelines for model construction, training, and evaluation that distinguish reconstruction, forecasting, state recovery, and parameter identification. The tutorial concludes by outlining challenges in uncertainty calibration, identifiability, and validation on real structures.
Whether models trained on different modalities converge to the same representation of the world is central to AI. The Platonic Representation Hypothesis (PRH) conjectures that they do as they scale, but without a proof it offers neither a guarantee nor a recipe. We turn the hypothesis into a theorem. First, we prove a \emph{Platonic Representation Theorem}: Joint-Embedding Predictive Architectures (JEPAs) trained separately on any number of modalities learn the same representation up to one rotation whenever their positive pairs are two observations of one scene--\emph{good views}--and their misalignment is bounded otherwise. Second, we show that next-token prediction is a JEPA with \emph{bad views}, parts of a single document, which explains why language models align with vision only through scale. Third, we validate the theory from 145k captions to the texts of 940M web images, and uncover a scaling law of alignment that predicts unseen runs within seed noise. Good views need 42 times fewer training samples than next-token prediction on the same corpus to reach the same alignment, and a 109M-parameter text JEPA aligns with DINOv2 and CLIP's image encoders better than every language model we evaluated, up to 72B parameters, and than CLIP's text encoder: next-token prediction is a remarkably inefficient route to platonic representations. As a showcase, one rotation turns separately trained encoders into an image captioner (CIDEr .838 on COCO). Even that rotation needs no pair: on the latest pair-free benchmark, our text JEPA beats every zero-shot accuracy reported with language models up to 8B parameters, e.g., 64\% against at most 56\% top-5 on ImageNet-100. Positive views, and not scale alone, are thus the lever to build platonic representations.
Reinforcement learning with verifiable rewards (RLVR) makes a language model solve more problems on the first try and fewer when given many tries: it raises pass@1 while the base model overtakes it at large sampling budget k. Is this a failure of RL to add reasoning, or a transient that more training removes? We settle the question with a theorem. The reversal is a price, and we compute its exact rate. Modeling RLVR as an exponential tilt of a base policy over strategies shared across prompts, we prove that under exclusive specialization every update that raises pass@1 strictly lowers the success of some prompt; a strictly positive price on prompts (Stiemke duality) is the exact condition, and being merely the best strategy on a prompt is provably not enough. When the loss reaches the hardest prompts, it forces a pass@k crossover, and the number of crossings is at most the sign variation of the shift in per-prompt success, a totally positive bound that is tight and that recovers the known single-crossover result as its order-preserving case. The crossover is not a bug to patch: a constrained update traces the exact pass@1 versus pass@k frontier, and removing the crossover entirely costs the entire pass@1 gain. The same tilt identity yields entropy collapse and the entropy-performance law with an identifiable exponent, unifying two literatures under one mechanism. Reinforcement cannot widen the reasoning frontier; it can only move along it, at a price we make measurable by a diagnostic that reads the bound, the order-preserving test, and the price from any evaluation log. There is no free lunch, only a price.
Large language model (LLM) distillation can propagate teacher bias under covariate shift when target reward feedback is unavailable. We propose Iterative Calibration and Learning (IterCaL), which alternates source-reward teacher calibration through token-level branching with student learning on target prompts. We prove polynomial convergence in expected average KL to the oracle student, the reference-regularized target-reward maximizer within a potentially misspecified student class; a quantified source-to-target calibration discrepancy gives an explicit additive true-return term. Our separation shows that regularized direct teacher matching retains positive oracle-KL error at its population optimum for every shared reference penalty, even when the teacher is stronger than every student. On the same instance, IterCaL converges to the oracle using source rewards alone.
Rotary Position Embedding (RoPE) typically uses uniform coverage across Transformer layers despite heterogeneous positional requirements. Phase-off interventions reveal non-uniform, model-specific rotary-phase dependence, while coverage interventions relate it to coverage-upgrade benefits. We propose profiling-guided layer-wise RoPE allocation, assigning high/low coverage under a fixed budget using early coverage-benefit profiles. The framework supports fixed-width query/key representations and physical rotary-dimension allocation in Multi-head Latent Attention (MLA). Normalized-depth mapping transfers profiles within model families, enabling small-model measurements to guide larger-model allocations before pretraining. A brief analysis relates selection and transfer to score calibration and inter-layer interactions. From-scratch experiments demonstrate validation-loss improvements over available budget-matched baselines, while analytical resource accounting identifies favorable quality--storage--compute trade-offs for physical MLA. These results connect positional demand to transferable, efficient architecture design.
Modeling asynchronous event sequences is crucial in numerous real-world applications such as healthcare monitoring, financial transaction analysis, and so on. Traditional temporal point processes, including Hawkes Processes, often fail to capture complex dependencies due to their parametric limitations. While neural approaches like RNNs and Transformers have improved flexibility, they struggle with computational inefficiency, and attention saturation. In this paper, we introduce the Mamba Hawkes Process (MHP), the first framework to integrate selective state space model (Mamba) with temporal point processes. MHP leverages time-varying state transitions and input-dependent gating to efficiently encode event history and capture long-term dependencies with linear complexity. Importantly, we provide theoretical guarantees showing that MHP generalizes both classical multi-exponential Hawkes processes and exponential-decay gated RNNs, underscoring its expressive power and theoretical soundness. To address the inherent constraints of pure state space models in handling heterogeneous event interactions, we further develop Adaptive Mamba Hawkes Process (A-MHP) that incorporates two novel mechanisms: a Time-Scaling Mechanism that adaptively weights time intervals based on event type and history, and a Dual-Channel State Transition that adaptively processes event content and temporal dynamics for more refined state updates. Extensive experiments on synthetic and realworld datasets demonstrate that MHP and A-MHP consistently outperform state-of-the-art baselines in event prediction tasks, particularly in long-sequence scenarios. Our work establishes a scalable and theoretically grounded paradigm for event sequence modeling, with practical implications for predictive maintenance, anomaly detection, and dynamic system analysis.
Parameter estimation and classification in binary logistic regression deteriorate severely when the number of minority class observations ($n_1$) is smaller than the feature dimension ($p$). Under such rank-deficient rarity ($n_1 < p$), maximum likelihood estimators suffer from quasi-complete separation, penalized likelihood methods exhibit severe intercept bias, and convex minority oversampling techniques such as SMOTE collapse into an $(n_1 - 1)$-dimensional affine subspace. In this paper, we propose a novel \textit{Self-Calibrated Zero-Complement} estimator that shifts the paradigm from minority interpolation to majority fringe extrapolation. Grounded in Stein's Lemma, we prove that the conditional expectation shift of the abundant majority class ($Y=0$) induced by logistic truncation uniquely identifies the true hyperplane normal $\boldsymbolβ$ after Mahalanobis whitening. We further establish that extrapolating synthetic minority instances outward from the high-quantile density fringe of the majority distribution along the conjugate direction preserves full-rank covariance regularity ($\mathrm{rank} = p$). To accommodate both semi-supervised environments (with an unlabeled background mean) and strictly supervised subsamples, we introduce a hybrid estimator incorporating the intrinsic third-moment skewness of the majority class alongside a location-shift-invariant zero-boundary intercept calibration. Extensive Monte Carlo simulations and real-world benchmark applications (Wisconsin Diagnostic Breast Cancer and Banknote Authentication) under extreme rarity ($n_1 \in \{3, 5\}$) demonstrate that our approach prevents parameter divergence, recovers the oracle hyperplane with high fidelity, and substantially outperforms Firth's regression and SMOTE in F1-score and estimation stability.
Understanding the loss-landscape geometry near a minimum is key to explaining the implicit bias of gradient-based methods in non-convex optimization problems such as deep neural network training and deep matrix factorization. A central quantity for characterizing this geometry is the maximum eigenvalue of the Hessian of the loss. Currently, its precise role has been obfuscated because no exact expressions for this sharpness measure were known in general settings. In this paper, we present the first exact expression for the maximum eigenvalue of the Hessian of the squared-error loss at any minimizer in deep matrix factorization problems, resolving an open question posed by Mulayoff and Michaeli [2020]. This expression reveals a fundamental property of the loss landscape in deep matrix factorization: It is sufficient, but not necessary, for flatness that, across all layers, the product of the spectral norms of the left and right intermediate factors equals a constant determined by the depth and the spectral norm of the target matrix. Most notably, in both depth-2 matrix factorization and deep overparameterized scalar factorization, we show that this condition is both necessary and sufficient for flatness, which implies that flat minima are spectral-norm balanced and not necessarily balanced in depth-2 matrix factorization. To complement our theory, we provide empirical verification, relying on our exact expression, of an escape phenomenon during gradient-based training near a minimizer of a deep matrix factorization problem.
Mame Diarra Toure, Quang Hung Pham, Mohamed Ryad Zemouri, David Stephens, Luc Vouligny, Martin Gagnon
Industrial systems continuously generate time series that vary in sampling frequency, scale, dynamics, and operating regime. Consequently a single time-series foundation model may not be able to represent all regimes effectively and retraining a large model for every new dataset or regime shift is computationally expensive. Modular foundation models address this issue by reusing specialised encoders trained on diverse datasets and by allowing the expansion of capabilities through the addition of new experts. This approach requires compact experts to minimize storage cost and a proper selection mechanism of experts for unseen data. We propose a modular foundation model framework that combines end-to-end Bayesian encoder training with uncertainty based routing. Using Singular Bayesian Neural Networks (SBNNs), we build a collection of compact Bayesian encoders whose epistemic uncertainty guides label-free selection. At the selected ranks, our Bayesian collection uses 63.8\% less storage than its deterministic counterpart. We evaluate this framework on time-series data through distribution shift, encoder selection, and probabilistic forecasting. Across seven incoming datasets, uncertainty-based routing is competitive with distance-based routing in terms of forecast MSE, while reducing average ECE from 0.086 to 0.072 and average NLL from 0.907 to 0.836. It also increases PICP90 from 82.28\% to 89.03\%. In a separate model-family comparison, we evaluate the SBNN pipeline against full-rank Bayes by Backprop, deep ensembles, and MC Dropout. The SBNN pipeline achieves an average MSE of 0.390, comparable with 0.396 for deep ensembles, 0.389 for MC dropout, and 0.443 for the full-rank Bayesian model. It uses 63.9\% fewer parameters than MC dropout, 81.9\% fewer than the full-rank model, and 92.8\% fewer than the deep ensemble.
Survival analysis provides statistical methods to model the time until an event occurs. Reporting delays arise when event times are not observed at their occurrence but are only revealed upon reporting. This issue is particularly critical for timely risk evaluation when the observation window is short due to administrative censoring. In this study, we incorporate right-censored reporting delays by jointly modeling parametric hazards for the event and reporting processes. We then construct a consistent estimator for the model parameters and develop a Monte Carlo expectation-maximization algorithm to compute it. To address the challenges posed by administrative censoring, we leverage these findings and propose a transfer-learning procedure. Experimental results demonstrate that our method improves the accuracy of timely risk evaluation under administrative censoring.
Learning independence is fundamental to statistical inference and machine learning, yet learning independence between entire time series by a test remains challenging when only a single observation record is available. Existing guarantees often rely on structural or mixing assumptions, while stationarity and ergodicity alone do not permit consistent testing. In this work, we identify a class-level boundary governed by shared nontrivial Koopman eigenvalues, which represent common persistent oscillatory frequencies of the two processes. When no such eigenvalues are shared, we extend shift-based rank calibration with finite-window dependence witnesses and a data-adaptive threshold to construct a deterministic, weakly consistent test with finite-sample false-positive control, without quantitative mixing assumptions. Conversely, we show that no possibly randomized test can achieve vanishing false-positive probabilities under every independent jointly totally ergodic law while consistently detecting every dependent jointly totally ergodic law with a shared nontrivial eigenvalue. This impossibility already holds for binary observations and persists within any prescribed nonempty open eigenvalue window, identifying a spectral obstruction to universal learning of full-process independence.
Estimating conditional parameters, such as heterogeneous treatment effects, often requires rich covariate adjustment, even when the parameters are reported along only a few dimensions. Fitting nuisance functions on the full sample makes all observations available for adjustment, but creates dependence between fitted scores and the observations at which they are evaluated, which is typically avoided by cross-fitting. We develop pointwise inference with full-sample nuisance predictions for exactly identified conditional moment equations, recovering conditional average treatment effects as a special case. We localize the estimating equations around the reporting point using distributional nearest neighbors (DNN), a U-statistic-based averaging procedure. Under sufficient conditions on the local design, orthogonality, prediction accuracy and quantitative deletion stability, the estimator is asymptotically normal, and a frozen-score jackknife consistently estimates its limiting covariance with nuisance predictions held fixed. In simulations, full-sample and cross-fitted inference perform similarly, with coverage shortfalls from nuisance approximation error, spatial bias and heavy studentized tails. An application to 401(k) eligibility using Survey of Income and Program Participation data finds similar estimates across nuisance prediction methods, with wide normal-reference intervals.