Papers

E-commerce search ranking must balance multiple objectives–relevance, user engagement, and platform revenue–when allocating impression slots to competing listings. Estimating the expected revenue component is well understood for fixed-price items, but becomes challenging when marketplace inventory includes mixed listing formats such as pure auctions and hybrid "Auction with Buy It Now" (ABIN) items, where prices evolve dynamically and the final transaction value is unknown at ranking time. Yet auction and ABIN listings account for a meaningful share of inventory and transaction volume on platforms such as eBay, and are a popular format for individual sellers and for unique items with unclear value. We extend the standard Expected Cost-per-Mille (eCPM) framework to auction and ABIN listings by deriving a marginal eCPM (meCPM) that captures the incremental value of showing one more impression of an item whose price is still evolving. The resulting formulation extends the familiar fixed-price eCPM–which is already inherently marginal–to auction dynamics, allowing unified ranking of fixed-price, auction, and ABIN listings under a single objective. We then describe a practical production implementation that approximates this objective, addressing cold-start challenges by bootstrapping from existing engagement models. Online A/B tests at a large e-commerce platform showed positive revenue gains and statistically significant improvements to user metrics, and the system was deployed to production.

Heterogeneous federated systems require agents to learn and exchange informative representations despite differences in data distributions, sensing modalities, model architectures, latent dimensionalities, and local learning objectives. To address this challenge, we propose Sheaf-based Federated Representation Learning (SFRL), a general framework that jointly optimizes local objectives with a manifold-constrained geometric alignment regularizer based on learnable sheaf restriction maps. Unlike most existing approaches, SFRL does not assume a shared global latent space. Instead, global consistency emerges from the alignment of neighboring latent representations through orthogonal transformations and isometric embeddings. This alignment is enforced by a quadratic gluing regularizer induced by the sheaf Laplacian, whose learnable restriction maps adapt the geometry to the observed data. The penalty is evaluated on a small set of shared pilot samples, ensuring scalability and communication efficiency. We develop a decentralized algorithm for solving SFRL, termed Sheaf-FRL, which alternates between gradient updates of the local models and closed-form Procrustes updates of the edge-wise restriction maps. We further establish convergence of Sheaf-FRL to first-order stationary points in both deterministic and stochastic settings. As an application, we consider a cooperative classification task in the context of semantic communication, under model and data heterogeneity. Our results show that Sheaf-FRL outperforms baseline approaches in terms of local and post-communication classification accuracy across different levels of local distribution shift and exhibits greater robustness to latent-space dimensionality compression.

Continuous hemodynamic monitoring guides treatment decisions in surgery and intensive care. However, gold-standard signals are only measured in severe cases due to risks associated with invasive measurement. In this work, we introduce HIPNO (Hemodynamic Inference via Physics-informed Neural Operators) to recover hemodynamic state from ubiquitous, non-invasive signals and expand access to advanced monitoring. HIPNO addresses a problem of scale symmetry in physics-informed hemodynamic inference, where different combinations of flow, resistance, and compliance can generate the same observed pressure. We identify the symmetry group of the observation model and parameterize the network in its quotient space. For the 3-element Windkessel model, the quotient coordinates are the compliance-normalized flow $U=Q/C$, the decay time constant $\tau_{WK}=R_2 C$, and the characteristic-impedance coordinate $\kappa=R_1 C$. Across 945499 intraoperative windows from 2562 patients, HIPNO predicts $\tau_{wave}$, a proxy for vascular decay derived from pressure, with 32% lower error on the log scale than a population baseline while preserving mean arterial pressure accuracy. Because vascular decay and flow drive occupy separate coordinates, counterfactual perturbations produce the expected directional responses in at least 90% of windows in almost all prespecified scenarios, a separation unavailable to pressure-only baselines. The coordinates are also used as inputs to a calibration model for monitored cardiac output. Finally, the formulation identifies the external compliance or flow reference required to recover absolute physical scale.

Low-precision formats usually optimize scalar fidelity while inheriting conventional product arithmetic. We introduce CurveFP, a block-scaled family that distributes magnitudes across interleaved logarithmic curves. Uniform curve indices make every nonzero product an exact sign and integer-index update, while a rational radix exposes the finite phase schedule required for accumulation. We instantiate the algebra as CurveFP8 E4C3/E5C2 for training and CurveFP7 E3C3 for compact inference. On four 7B-9B models, CurveFP7 beats tensorwise FP8 perplexity with one fewer element bit and stays within 1.32% of native quality. CurveFP8 lowers error in all 36 paired training-GEMM comparisons. Across three matched 3B-token pretraining triplets, it reaches mean BF16-inference perplexity 22.5366 versus 22.5407 for FP8 and has a lower format penalty in every seed. Downstream evaluation shows transfer parity and a consistent WikiText-103 gain. In a preliminary 4x4 Nangate45 spatial accelerator tile, CurveFP8 uses one fewer product register and 4.6% less area than timing-closing FP8 at 500 MHz. These results support CurveFP as a numerical and arithmetic co-design, while leaving system-level efficiency to future study.

Data-driven health-state estimators for SiC (Silica-Carbide) power modules typically report their performance on a single accelerated-aging campaign, and how that performance transfers to a different failure mechanism is rarely tested. We benchmark five reference methods from the prognostics and condition-monitoring literature against a physics-informed NODE (Neural Ordinary Differential Equation) on two SiC power-cycling campaigns driven by structurally different failure mechanisms, solder-layer fatigue and wire-bond lift-off, under a per-module $k$-fold protocol. The NODE is evaluated under two input regimes that share the rest of the pipeline: the baseline electrical precursors and a set of cumulative thermoelectric features. Every reference method degrades on the wire-bond campaign, with average errors growing and precision decreasing with respect to their performance on the soldered campaign. The NODE fed with the cumulative features keeps its soldered-campaign metrics on both mechanisms, with differences inside the fold-to-fold variance, while the same architecture fed with the baseline precursors falls back to the reference-method cluster. The input representation contributes at least as much as the architecture to failure-mechanism transferability of a health-state estimator.

Silicon carbide (SiC) power modules are increasingly deployed in automotive traction inverters, where condition monitoring is essential to prevent in-service failures. Despite extensive qualification under AQG 324, no consolidated approach exists for in-field health state estimation: physics-of-failure lifetime models lack real-time applicability, purely data-driven architectures require large labeled datasets and generalize poorly, and physics-informed frameworks remain too demanding for embedded deployment. We address SiC MOSFET modules assembled with sintered packaging, which suppresses solder degradation and produces aging behavior distinct from previously studied devices. Instead of the smooth quasi-exponential drift of solder-based modules, the forward voltage drop $V_{DS}$ exhibits multi-regime profiles, with wirebond liftoff events introducing abrupt, non-monotonic perturbations. We propose a condition monitoring framework combining three elements. First, physics-informed features replace raw sensor signals with cumulative damage indicators derived from junction temperature swing, mean junction temperature and a Miner rule accumulator, encoding degradation history in an interpretable form. Second, a monotonicity constraint enforced by gradient penalty regularization embeds the expected degradation direction as a physics-guided prior. Third, a heavy-tailed output distribution replaces the point estimate, giving calibrated uncertainty robust to the out-of-distribution variance introduced by liftoff. On an industrial power cycling dataset from Infineon Technologies, several neural architectures are compared under a strict cross-validation protocol. The full configuration reduces mean absolute error by approximately 70% over purely data-driven baselines and stays stable across all folds, while remaining lightweight enough for embedded deployment.

Ordinal regression, also called ordinal classification, is classification of ordinal data, in which the underlying target variable is categorical and considered to have a natural ordinal relation. Previous works have indicated that, in many real-world ordinal data, the conditional probability distribution (CPD) of the target variable given a value of the explanatory variable would be unimodal in a large domain of the explanatory variable and close to be unimodal even in a remaining domain. Therefore, unimodality-promoting regularized learning (UPRL), which promotes a predicted CPD closer to be unimodal with the aim of decreasing a prediction variance without inducing much bias for ordinal data of the unimodality, is promising to improve the prediction performance especially with small-size training data. In this study, we show that previous UPRL methods promote a predicted CPD to not only become closer to be unimodal but also have a larger scale (in other words, be smoother or less-confident). Therefore, we develop a novel method that more strictly reflects the idea of UPRL and evades a scale-related bias, and verify through experimental comparison that the unimodality-promotion indeed contributes to improve the prediction performance. Additionally, while our proposed UPRL method could perform better for smaller-scale data or with larger-size training data compared to a previous UPRL method, our analysis explains this experimental observation in terms of the presence or absence of an unexpected scale-related bias.

The initialization of deep neural networks determines whether information and gradients can propagate across depth, yet a unified theory connecting these properties to learning dynamics remains elusive. Combining mean-field theory and random matrix theory, we establish a direct link between correlation propagation and the Neural Tangent Kernel (NTK) that governs learning in the sequential limit of infinitely wide, infinitely deep networks. Correlation propagation to infinite depth is possible only at a single, critical point in the weight-bias variance plane. At this point, we leverage the algebraic decay of the end-to-end Jacobian with depth to prove that the NTK becomes exactly proportional to the output correlation at infinite depth, tying together information propagation and learning dynamics. We further show that orthogonal initialization suppresses the leading finite-size corrections present under Gaussian initialization, clarifying the respective roles of the two initialization ensembles in this limit. These theoretical predictions are validated quantitatively on finite-width, finite-depth networks. Together, these results demonstrate that orthogonal initialization and criticality are required to control the asymptotic dynamics of deep learning.

Permutation optimization arises whenever the components of a system are fixed but their ordering affects performance. We introduce PRISM, a predictive protocol for permutation optimization that measures a fitness landscape before selecting a search strategy. PRISM uses inexpensive landscape diagnostics, including one-step move autocorrelation and fitness-distance correlation, to predict useful mutation operators, identify when structured search is likely to outperform random sampling, and detect regimes in which search provides little advantage. Across synthetic permutation landscapes, neural architecture benchmarks, scientific machine learning pipelines, and large-language-model instruction ordering, the protocol makes testable predictions about search behavior before optimization begins. Exhaustive instruction-ordering experiments reveal substantial performance variation induced solely by permutation, while cross-model experiments show that useful ordering structure can transfer across model families and task difficulty. Additional experiments demonstrate that instruction ordering remains consequential after prompt wording is optimized, indicating that content optimization and ordering optimization are complementary. The results position PRISM not as a universally superior optimizer, but as a framework for determining when permutation search is useful, which representation and operator should be used, and when simpler alternatives are preferable.

Adaptive loss-balancing schemes for physics-informed neural networks rest on a premise that every residual should be driven to zero. For level-set advection with an eikonal regulariser that premise fails: the eikonal term penalises deviation of $\lVert\nabla\phi\rVert$ from unity, a property transport preserves only under rigid motion; where the exact solution departs from a signed-distance function the eikonal residual of the correct answer is nonzero, and driving it to zero moves the network away from that answer. We show that standard gradient-norm balancing fails in exactly this way, its weight remaining near its initial value throughout training on benchmarks where the property is violated, and we introduce SDF-Aware Weighting (SAW), which combines a residual-quantile gate with a gradient-norm ratio so that points exhibiting legitimate departure are excluded before the surviving term is scaled. Across four three-dimensional benchmarks SAW selects an eikonal weight within an order of magnitude of the value located by an eighteen-run manual sweep, spanning four decades from $10^{-1}$ to $10^{-5}$ with a single fixed configuration. On the slotted sphere, where the initial field is non-differentiable at reentrant edges, SAW attains a lower error than any weight in that sweep. Two smooth rigid benchmarks serve as controls: SAW is worse there, as expected when its premise does not hold. An ablation with the gate disabled shows the slot is nearly entirely filled while the relative $L_2$ error reads $1.06\%$, indistinguishable from a field that never represented the slot. We give a feature-restricted measure that separates the two cases.

Floods and landslides often co-occur, but their relationships with environmental controls vary spatially. This study develops a spatial heterogeneity-aware framework for flood-landslide susceptibility and relative-risk mapping in Kerala, India, and Nepal. It combines 15 km x 15 km grid cells with region-specific contextual zones and compares proximity-gated cross-zone training (S1) and ecology-gated zone-constrained training (S2). S1 permits geographically nearby models to be assigned across contextual boundaries, whereas S2 restricts model development and assignment to the same zone. Random Forest models for each hazard use strategy-specific predictor sets and are evaluated on spatially held-out test samples. Susceptibility surfaces are integrated with CRITIC-weighted exposure and vulnerability indices to produce hazard-specific and nine-class bivariate relative-risk maps. S1 achieved higher mean accuracy, precision, recall, F1-score, AUC-ROC, and PR-AUC for both hazards and regions. The largest difference occurred for Nepal flood susceptibility, where AUC-ROC increased from 0.728 under S2 to 0.886 under S1 and PR-AUC from 0.512 to 0.823. S2 produced lower Brier scores for both Nepal hazards and retained zone-specific differences in predictor selection, SHAP rankings, and response patterns, particularly in Kerala. Both strategies reproduced flood-prone lowland and landslide-prone upland patterns but differed in susceptibility and risk classes. Bivariate risk-map agreement was 0.521 in Kerala and 0.711 in Nepal, with allocation disagreement exceeding quantity disagreement in all S1-S2 comparisons. Susceptibility-to-risk correspondence remained below 0.350, showing that exposure and vulnerability changed priority locations. Overall, cross-zone learning strengthens regional discrimination, while zone-constrained learning preserves environmental differences, supporting their integration.

Brain magnetic resonance imaging (MRI) is central to neuroscience and clinical assessment, but models are commonly developed for individual diseases, populations or imaging protocols. Foundation models promise more general representations, yet they are usually pretrained once and can lose earlier capabilities when updated with new data. Here we show that Alcmaeon, a three-dimensional brain MRI foundation model pretrained without manual labels on more than 425,000 volumes and derived imaging maps, can be expanded sequentially across clinical domains. Alcmaeon combines volumetric encoding and latent diffusion generation with Graph-Blueprint Pruning (GBP), which protects network modules important to earlier domains while leaving the remaining capacity trainable. Across expansion from healthy ageing and neurodegeneration to developmental, psychiatric and tumour imaging, GBP showed less forgetting than sequential adaptation and elastic weight consolidation across voxel-level reconstruction measures, with its largest advantage after adaptation to tumour imaging. The blueprints provided an inspectable record of how model capacity was protected and reused. Representations from different model levels supported image synthesis, disease classification, survival modelling and postoperative prediction, although no single representation was optimal for every task. These findings provide a route towards brain MRI foundation models that can grow with emerging data while retaining earlier capabilities.

Biological neural systems achieve high efficiency and robustness through compartmentalized architectures. In contrast, modern artificial neural networks rely on globally entangled structures, which obscure decision logic and suffer from catastrophic forgetting. Here, we report a Decomposable Spiking Neural Network (D-SNN) that eliminates global synaptic entanglement by structurally isolating classification pathways into independent experts. Optimized via a bio-inspired push-pull loss function, the D-SNN achieves competitive accuracies on MNIST, Fashion-MNIST, and CIFAR-10/100 benchmarks. This modular approach matches the performance of fully dense networks while utilizing an order of magnitude fewer parameters. In addition, our networks operate with up to several orders of magnitude lower firing rates and fewer synaptic operations. Furthermore, physically severing connections between experts provides inherent protection against catastrophic forgetting during sequential learning. Crucially, these isolated pathways generate auditable neural signals, increasing decision transparency. This biomimetic, verifiable architecture establishes an efficient foundation for deploying deterministic neuromorphic intelligence in resource-constrained edge environments.

We argue that learning visual representations without labels requires a training signal jointly complete across three non-overlapping objectives: semantic invariance across augmented views, patch-level spatial prediction, and representational non-degeneracy. We formalize these as the observation, prediction, and regularization principles and prove (i) that combining observation and prediction without regularization admits the constant encoder as a global minimizer under negative-free alignment; (ii) that the two objectives are gradient-complementary and structurally non-conflicting at the encoder output; and (iii) that the momentum encoder converges to the same fixed point as the online encoder and provides no collapse guarantee at convergence. Contrastive alignment provides only self-limiting collapse resistance, formalized via an explicit gradient-decay argument. Dropping prediction withholds the spatial training signal by construction; dropping observation forfeits cross-view semantic invariance by construction; at the scale we study, no pair substitutes for the third. Every major self-supervised method is a special case of a single unified energy decomposition. We pair every theoretical claim with a controlled experiment, including a patch-retrieval evaluation for the spatial consequence of prediction.

Reliable identification of railway wheel defects is important for safety and maintenance. This study develops a machine-learning-based diagnostic framework for multi-class defect identification using passive air-coupled ultrasonic acoustic emission signals. Data were collected from eleven full-scale railway wheelsets representing nine health states. Time- and frequency-domain features were evaluated using Kruskal-Wallis statistical testing and mutual-information analysis to identify the most discriminative indicators. A Random Forest classifier was then trained using the selected features with stratified 5-fold cross-validation. The model achieved a balanced accuracy of approximately 0.66 and a Macro-F1 score of 0.65 across the nine classes. Decay rate, kurtosis, skewness, and envelope low-frequency power emerged as the most influential features, while a compact subset of features retained most of the classification performance. The results demonstrate the feasibility of combining passive ultrasonic sensing, statistical feature selection, and supervised machine learning for non-contact railway wheel defect classification and provide a foundation for future field-deployable inspection systems.

Knowledge distillation trains a smaller student to match the outputs of a larger teacher. Feature-based methods also align intermediate representations, but this extra constraint may affect students differently. We study this question on CIFAR-100 using a ResNet-50 teacher, a width-controlled CustomResNet family and MobileNetV2 as a cross-design comparison. For each student, we evaluate each feature method against a matched logit-KD run using the same teacher, optimizer settings, training schedule and seed. We repeat the main comparisons across multiple seeds. Logit KD improved every tested student over its scratch baseline. Attention Transfer showed no clear relationship with size inside the CustomResNet family, but its average effect was negative for that family and positive for MobileNetV2. FitNets was below logit KD in all 15 paired runs. Within the constant-depth width sweep, its gap increased for wider students, although the different-depth w=48 student did not follow this trend. Finally, the same auxiliary coefficient produced different gradient scales across students, showing that a fixed coefficient does not create a uniform training condition.

Estimating counterfactual outcomes over time from longitudinal observational data is central to clinical decision support. Existing methods rely on domain confusion – adversarial training that renders representations invariant to treatment assignment – yet this invariance creates a mutual information conflict: it suppresses treatment-correlated covariate signals necessary for accurate outcome prediction. We formalise this tension via a Jensen-Shannon divergence bound on counterfactual prediction error and develop two complementary models. CSSD (Causal State-Space model with Direct decoder) adapts selective State Space Models with a parallel multi-step decoder that eliminates accumulated rollout error by producing all prediction horizons simultaneously in a single forward pass. CSSPD (Causal State-Space model with Predictive regularisation and Direct decoder) augments CSSD with Contrastive Predictive Coding and Local Information Maximisation to reinforce temporal predictability in the balancing representation and recover local covariate information destroyed by domain confusion. On MIMIC-III, CSSPD achieves lower counterfactual RMSE than the Causal Transformer at every horizon tau >= 2 at O(T) encoder cost, with gains from 0.02 (2-step) to 0.07 (6-step). On Cancer Simulation across confounding strengths gamma in {0,1,2,3,4}, CSSPD outperforms CT at gamma <= 3 (margins 25.9%–37.0%), and CSSD achieves the lowest overall average RMSE (12.7% reduction over CT), confirming the MI conflict analysis. To our knowledge, this is the first work to formalise the balancing-prediction MI conflict and propose a structured resolution through complementary predictive and information-theoretic training objectives.

Tool-using language-model agents face constraints whose meaning changes with observations and prior actions. We study exact sampling from the model distribution conditioned on a hard stateful validator while reusing invalidity certificates across histories. Stateful CARS freezes a bank of sound state–continuation schemas within each attempt and removes every trajectory containing a certified continuation at a matching abstract state. An exact residual Doob transform samples from the resulting proposal. We give a checkable future-validity bisimulation condition, prove schema soundness, adaptive exactness, i.i.d.\ outputs, almost-sure termination, monotone acceptance, and compression invariance, and characterize computation by the number of reachable full-history product states. This number can be exponential for a history-dependent language model; the evaluated method therefore makes no generic finite-trie scalability claim. On enumerable workflows, its analytic law matches the valid conditional to $10^{-16}$ at validity probability $6\times10^{-8}$, whereas state-aware local decoding can be $0.97$ away. A matched comparison is negative: observation-keyed official CARS is cheaper in sampler steps (root/Stateful ratio $0.942$ $[0.934,0.951]$), and the Qwen comparison is null ($0.99$ $[0.90,1.08]$). Cross-history transfer helps only in an internal matched-key ablation ($1.27\times$). Thus the evidence supports exact schema-induced conditioning, not a systems advantage over CARS.

As firms increasingly deploy machine learning for strategic decision-making, understanding algorithmic interactions has become central to operations research and economics. This paper studies learning in infinite-horizon, nonzero-sum linear-quadratic stochastic games under a radically uncoupled information structure, where players are either unaware of opponents or strategically oblivious, observing only a common state and their own action history. Under this minimal information, we analyze an asynchronous decentralized learning process in which each player independently runs a single-agent $\epsilon$-greedy iterated least-squares algorithm. We prove that, despite being unable to identify the system parameters, players' learning dynamics converge almost surely to the complete-information Nash equilibrium and characterize the convergence rate. We then apply the framework to a dynamic Cournot competition with sticky prices. Numerical experiments validate the theoretical results and show that learning under limited information reduces firm profits under both low and high price stickiness, while total surplus declines and market concentration increases when price stickiness is high. Publicly revealing aggregate market output substantially accelerates convergence and mitigates these welfare losses.

Code generated by modern language models often reads naturally. Yet, it also often fails to implement what was asked. This should be no surprise, as research shows the models' own confidence signals are poorly calibrated with actual correctness. A promising way to assess correctness looks inside the model: by contrasting the hidden states of correct and incorrect programs, recent work captured an internal signal of code correctness that is able to judge candidate solutions better than the model's token-level or stated confidence, with no test execution. However, this signal was captured under one particular way, leaving open an important question: whether it reflects a robust property of the model or an artifact of that choice. We study this question systematically, varying how the signal is extracted from the model internals. Besides this, we also ask if the signal's quality is limited by the data used to extract it, by constructing program pairs that differ only in the fault that makes them incorrect. Our results show that no single configuration is best, and that isolating the fault does not help.

Oracle routing measures how much a pool of language models could gain from per-query selection, but the diagnostic has two flaws: testing against a best fixed model selected on the same examples invalidates paired inference, and a full-information oracle sees outcomes no deployable router observes. We separate three estimands (outcome-oracle opportunity, the Bayes-optimal gain from a declared pre-answer signal, and the held-out gain of a learned router) and prove selection-valid confidence intervals that survive choosing the best fixed model or the best member of a router family, a signal-information sandwich, and a $(1-1/e)$ greedy guarantee for building compact pools from submodular complementary coverage. On eight checkpoints from six families over four benchmarks, selection-valid intervals certify a population oracle gap of $9.7$–$30.7$ points on every task, yet the strongest deployable prompt router recovers only $7.5$–$14.4\%$ of it, and the simultaneous interval for the best of eleven tested policies has lower limit zero throughout. The realizable share of oracle opportunity is small and certifiable: strong routers beat the best fixed model, and most of the gap remains.

Agentic reinforcement learning (RL) often suffers from delayed and sparse rewards in real-world environments. A promising solution to this challenge is credit assignment, which aims to decompose trajectory-level rewards and provide more fine-grained supervision for intermediate decisions. However, existing credit assignment approaches ignore the rich process information naturally generated during environment interaction, e.g., interaction history. We argue that such information provides valuable supervision for identifying the contribution of individual actions. To this end, we propose Environmental Feedback-based Credit Assignment (EFCA), a multi-timescale credit assignment approach for long-horizon agentic RL. EFCA complements the long-term outcome signal with two environment-grounded process signals: a short-term feedback signal that captures the immediate effect of the current action and a medium-term state-history signal that identifies ineffective patterns from recent interactions. Both signals are directly extracted from environment feedback and integrated through a return reweighting mechanism. Experiments on ALFWorld and WebShop demonstrate that EFCA consistently improves both task success and task quality over strong baselines, highlighting the effectiveness of environment-grounded multi-timescale credit assignment for long-horizon agentic RL.

LLM routers promise efficiency by matching each request to the cheapest adequate model, and are increasingly applied per step inside multi-step agents. Yet agentic routers are evaluated like single-turn routers: by replaying logged trajectories and substituting another model's recorded outputs, assuming the rest of the trajectory is unaffected. We test this assumption with branching rollouts: we fork live SWE-bench agent trajectories at controlled points, rebuild the environment, continue each fork with a different model, and compare against same-model control forks that isolate sampling and replay noise. Across six paired runs (~900 rollouts), swaps exceed their matched control floors by +0.25 to +0.66 normalized edit distance (multiplicity-corrected CIs exclude zero), rewriting 61-94% of post-fork actions; 74-77% of early swaps diverge at the first post-fork action, versus 6-35% of controls, leaving only 3% of replayed states valid. Divergence decreases with fork depth in both directions. All five outcome flips we observe occur in swap arms, upgrades rescuing unsolved instances and a downgrade losing the sole solve, and zero occur across 359 control forks. Scoring these same swaps with a log-stitching replay evaluator, replay mispredicts every success-relevant outcome call and predicts patches with 0.00-0.11 similarity to reality. Auditing the noise floor, temperature-0 "determinism" is configuration-dependent: FP8-served controls diverge on over 90% of forks while AWQ-served ones remain near-identical; and under tight budgets the stronger model more often exhausts its steps without submitting. Replay-based benchmarks score the wrong world for agentic routing; we release our harness and all trajectories.

Retrieval-Augmented Generation (RAG) systems in production operate under strict service level objectives (SLOs) on tail latency and infrastructure cost. However, standard retrieval pipelines rely on fixed retrieval budgets that ignore query difficulty, over-retrieving for easy queries and under-serving hard ones, forcing operators to trade answer quality against SLO compliance. This paper proposes SAGE, a learned SLO-aware adaptive retrieval policy that dynamically selects the number of passages k per query. SAGE uses lightweight features derived from initial retrieval (e.g., score distributions, rank gaps, lexical signals) and is trained offline via imitation learning from an oracle that approximates optimal latency-quality trade-offs. At inference, it adds no LLM calls and minimal overhead. On Natural Questions, under a 5s P95 latency SLO, SAGE achieves 95% SLO compliance versus 30% for the best static baseline (k=20), reduces P95 latency by 36% and retrieval cost by 51% with only 2 percentage points Exact Match (EM) loss. A single policy trained on Natural Questions generalizes across HotpotQA, UnSeenTimeQA, and four LLM families (Llama, Qwen, Mistral, Gemma), consistently yielding +45-52 point SLO improvements without quality degradation.

Reinforcement learning post-training unlocks complex reasoning in LLMs. Yet benchmark scores reveal only whether a model improved, not what changed inside it, nor how it splits finite capability across tasks. A representative interpretability line attributes the success of RL fine-tuning to stronger and more diverse circuit activation. We challenge this activation-centered account by separating activation from control: an activated circuit need not control the post-training reward gain. Adapting Metabolic Control Analysis, we define the Post-training Control Coefficient to measure component control over reward gain and arrange these coefficients by task family into a control matrix, paired with an activation-magnitude matrix. We call cross-task control concentration the Shared Control Bottleneck and the difference between activation and control concentration the Activation-Control Gap. This reveals that highly shared activations can coexist with task-specific control, while a small gap indicates that control has collapsed onto a shared direction and lost task specificity. To reduce this collapse, we regularize the post-training loss with the Shared Control Bottleneck and propose Control-Diverse Reinforcement Fine-Tuning (CD-RFT). The exact regularizer gradient requires second-order automatic differentiation incompatible with flash attention, so we derive a first-order proxy with worst-case overhead below eight percent. On Qwen2.5-7B, CD-RFT achieves the largest control decoupling and improves multi-task capability over matched GRPO across mathematics, code, and logic. The no-KL variant leads on pass@1, and the KL-penalized variant leads on large-k pass@k coverage that KL otherwise degrades. Together, these results show that the Shared Control Bottleneck is both a mechanistic diagnostic and a training regularizer, and that control decoupling and capability gains transfer to Llama-3.2-3B.

Multivariate Time Series Classification (MTSC) demands models that can effectively capture complex temporal patterns across multiple scales while remaining computationally efficient. However, existing approaches generally struggle to reconcile fine-grained representation learning, especially under class imbalance and real-world constraints. In this paper, we present FreSH, a Frequency-Segmented Hierarchical Multi-Expert Framework designed to address these challenges. FreSH introduces a new perspective for MTSC by enabling adaptive, multi-scale analysis of temporal signals, allowing different aspects of the data to be modeled in a complementary and coordinated manner. By combining localized specialization with holistic context modeling, FreSH achieves strong representational capacity without incurring excessive computational overhead. An adaptive fusion strategy further enhances flexibility, enabling the model to dynamically emphasize the most informative components of the input. In addition, we incorporate a more robust optimization objective that improves learning stability across varying sample difficulties and class distributions. Extensive evaluations on 30 UEA benchmark datasets and real-world vibration data demonstrate that FreSH consistently outperforms state-of-the-art methods in classification accuracy, while substantially reducing model size and efficiency.

This work proposes deep nonparametric Instrumental variable quantile regression (IVQR), a two-stage estimator that combines conditional diffusion modeling with a kernel-smoothed conditional moment formulation. In the first stage, we estimate the joint conditional distribution of the outcome and endogenous covariates given the instrument using a variance-preserving conditional diffusion model. In the second stage, we approximate the conditional moment operator through Monte Carlo sampling and a kernel-smoothed surrogate for the indicator function, and then estimate the structural quantile function by empirical risk minimization over deep neural networks. We establish an excess-risk bound for the proposed estimator and derive end-to-end total variation guarantees for the conditional diffusion model under unbounded support, explicitly accounting for score estimation, early stopping, and discretization errors. Our theory is developed under a polynomial-tail envelope on the data distribution and degenerates continuously to the exponential setting: as the tail index grows, the obtained excess-risk rate converges to the minimax-optimal rate of nonparametric regression, thus our heavy-tailed theory covers the classical light-tailed nonparametric guarantees as a limiting case. Simulation studies and a real-data application demonstrate that the proposed method outperforms existing nonparametric IVQR approaches, with gains that become increasingly pronounced as the dimensionality of the covariates and instruments increases.

Data-centric curation pipelines frequently rely on model confidence scores to flag and filter noisy or mislabeled training instances. Evaluating this filtering convention on a large-scale consumer lending sample (LendingClub, N = 1,344,936) uncovers an underlying demographic asymmetry: high-income defaulters are disproportionately classified as label noise relative to low-income defaulters (Cramer's V approximately 0.03-0.07). Re-examining this behavior through the lens of equal opportunity [Hardt et al., 2016] reveals a far more severe discrepancy: a 16.86 percentage point gap in true positive rate (recall) between high- and low-income borrowers who ultimately defaulted. Implementing a sequential feature-blinding methodology allows us to isolate the drivers of this disparity across three distinct mechanisms: (1) direct reliance on self-reported applicant income; (2) algorithmic absorption of upstream institutional bias encoded within origination interest rates; and (3) a residual disparity (3.55 percentage points in cross-validation; 2.56 percentage points on a held-out test partition, Z = -4.04, p < 0.0001) that remains even after purging both income and interest rates from the model. Out-of-sample signed SHAP valuations demonstrate that this residual gap is maintained by structural proxies, most notably loan amount and home ownership status. These empirical findings show that simply blinding an algorithm to sensitive attributes fails to ensure fairness when institutional pricing decisions and behavioral proxy variables collectively reconstruct the omitted signals. We outline the practical implications of these findings for auditing data-centric AI workflows within regulated financial institutions.

Post-training quantization reduces the deployment cost of large language models, yet how severely a quantized model degrades is not determined by bit-width alone. We systematically study weight-only post-training quantization across bit-widths, quantization methods, model scales and downstream tasks on multiple model families. We observe that such degradation varies substantially across these factors: 4-bit quantization usually preserves performance, 2-bit often causes broad degradation, and at 3-bit, degradation becomes apparent but varies markedly with task type, quantization method and model scale. To explain this variability, we use the signal-to-noise ratio (SNR) to measure how strongly quantization perturbs full-precision representations. We trace degradation back to two linked processes: how quantization errors arise within individual modules, and how they accumulate across layers. First, a source SNR decomposition shows that newly introduced errors depend on three factors: the magnitude of the weight error, the strength of the task-specific signal, and how strongly the quantization error aligns with task-specific activations. Different factors affect these components in distinct ways. Second, a cross-layer propagation analysis shows that these errors can be attenuated, preserved, or amplified as they pass across layers, and that larger models benefit from weaker error amplification. Together, these results establish that quantization degradation is governed by how errors are introduced at the source and how they accumulate across the network.

Machine learning models for MALDI-TOF mass spectrometry have shown considerable promise for clinical microbiology tasks such as microbial identification and antimicrobial resistance prediction. However, their deployment across institutions remains limited by domain shift, as acquisition-specific variability often leads models to capture technical artifacts rather than transferable biological information. Existing representation learning approaches primarily address this problem through statistical domain alignment while largely overlooking the biological supervision naturally available in microbiology datasets. We introduce DALMA, a probabilistic representation learning framework that jointly models acquisition-specific variability and biological supervision to learn biologically structured latent representations. By combining domain-specific reconstruction with biologically guided representation learning, DALMA learns transferable representations that generalize across heterogeneous clinical centers without requiring institution-specific components at inference, enabling zero-shot deployment on previously unseen sites. We evaluate DALMA on a multi-center benchmark comprising seven datasets from three countries. DALMA consistently achieves state-of-the-art zero-shot microbial identification across two held-out clinical centers, while the learned representations also transfer effectively to antimicrobial resistance prediction. Furthermore, latent-space novelty estimation enables reliable selective prediction under previously unseen domain shifts. These results demonstrate that biologically informed representation learning provides an effective strategy for robust and transferable ML in clinical microbiology.