Papers

We study a hybrid proposal–correction architecture for reduced-order PDE forecasting in which a pure-state quantum reservoir computer (QRC) predicts latent coefficient dynamics and a PINN-based physics-informed corrector (PIC) refines local rollout windows. The method is evaluated on Burgers and Kuramoto–Sivashinsky (KS), with KS as the primary chaotic benchmark. On KS, QRC+PIC consistently improves over QRC alone in RMSE, NRMSE, and PDE residual, while Burgers highlights a regime in which simple baselines remain strong. These results suggest that QRC proposals with local physics-informed correction are a viable benchmark-dependent reduced-order forecasting strategy.

Predicting transcriptome-wide responses to unseen genetic perturbations remains a major computational challenge because accurate prediction requires recovering both perturbation-specific transcriptional shifts and heterogeneous cellular responses. Existing methods often entangle deterministic response structure with stochastic population-level variation, causing dominant shared patterns to mask weaker perturbation-specific signals and impair distributional modeling. To address these challenges, we propose DeMixPert, an approach for Decomposed response Modeling with Gaussian Mixtures for Out-Of-Distribution (OOD) single-cell Perturbation prediction. DeMixPert decomposes perturbation-induced changes into a basal-state-dependent systematic response, a perturbation-specific response, and population-level variation. The systematic component is derived from the basal state encoded from control-cell expression, whereas the perturbation-specific component is inferred from pretrained target embeddings for unseen-target generalization. DeMixPert models population-level variation using a Gaussian prototype Invertible Network and adaptively combines reusable Gaussian prototypes according to the basal state and perturbation condition. The resulting mixture is mapped to a condition-specific variation distribution. Sampled variations are integrated with the systematic and perturbation-specific components, followed by joint decoding with the basal state to reconstruct perturbed-cell gene expression. Experimental results show that DeMixPert effectively captures heterogeneous single-cell perturbation responses and achieves superior performance across unseen-perturbation settings. The source code is made publicly available upon publication.

Recent progress in the field of TinyML has demonstrated that low-power hardware based on microcontrollers can achieve bird species monitoring in real time based on acoustic sensor data for an entire breeding period on a single battery charge. However, the state of the art on low-power microcontrollers was so far limited to binary classification of a single species. In contrast, real fauna monitoring deployments often target multiple species simultaneously. To address this challenge we develop PolyChirp, an approach combining biological domain expertise, automated dataset curation, neural architecture optimization and novel hardware to achieve multiclass bird species detection in the wild. PolyChirp is based on newly designed tiny multiclass models that leverage recent microcontrollers and hardware acceleration with a neural processing unit (NPU). We evaluate the predictive performance of these models, and we measure their computational performance – flash footprint, latency, energy consumption – on common microcontroller hardware. Our results demonstrate that PolyChirp matches or exceeds the TinyChirp architectures retrained under our protocol on single-species detection, and further achieves robust classification of up to 10 species simultaneously (macro F2 up to 0.97), while still fitting the flash, latency and energy budget of a low-power microcontroller sensor. A data-driven front-end redesign additionally makes on-device mel feature extraction 7x to 11x cheaper.

One implicit DDIM inversion step is the cheapest probe of whether a pretrained diffusion model encodes local manifold geometry. It is the stationarity condition of an explicit potential, $x-G(x)=\nabla\Psi_t(x)$, strongly convex at the Bayes limit with modulus exactly $e^{-h_t}$ for the step's log-SNR gap $h_t$ $-$ for every data law, schedule and point, with no manifold, reach or unimodality hypothesis. Three consequences must be kept apart. (i) The solution is unique at the Bayes limit; a second one requires the trained score to violate the posterior-covariance bound by $1/(1-e^{-h_t})$, a hypothesis-free certificate of model error; the same bound makes contraction a schedule constant, $\rho_g^{\star}=1-e^{-h_t}<0.326$ throughout the standard DDPM schedule. (ii) The solver can still fail: Picard iteration is unit-step gradient descent on $\Psi_t$, unstable wherever $\lambda_{\max}(\nabla^2\Psi_t)>2$, so oscillation certifies nothing; damping below $2/\lambda_{\max}$ cures it. (iii) The geometry lives in the convergence domain: on the scale-free depth $w=r\kappa_{\max}$ the oscillation shell sits at $w=\tfrac12$, schedule-free, and the divergence shell at $w=1/(1+\rho_g^{\star})$, with a measured finite-noise correction in $\|\mathrm{II}\|^2$. Exact scores reproduce both to within $0.54\%$ on three classes; no trained score we probe shows a shell $-$ a derived limitation, not a null result: the Fermi window conflicts with the model's own training support by $3.6$-$5.6\times$, and the trained Hessian-Lipschitz constant is $2$-$12\%$ of the curvature the law reads, $0$ on a ReLU net. Finally the unconditional ceiling $\sigma_t\lambda_{\max}(\mathrm{sym}\,J)\le1$, from $\mathrm{Cov}(x_0\mid x_t)\succeq0$ alone, holds for the exact score to $3\times10^{-7}$ but is violated in all DDPM CIFAR-10/CelebA-HQ-256 settings, by $1.26$-$4.66\times$.

We study parameter recovery in the Caldeira–Leggett (quantum Brownian) oscillator from partial moment traces. Our model is a moment-level PINN that predicts the five first/second moments and enforces the linear CL/HPZ ODEs by automatic differentiation. Physical structure is imposed through a PSD (Cholesky) covariance head, high-temperature CL assumptions with $D_{xp}\approx0$, and fluctuation–dissipation ties between $D_{pp}$ and $\gamma$. On synthetic CL data with channels ${\mu_x,\sigma_{xx},\sigma_{xp}}$, the constrained variant recovers $(\omega,\gamma)$ accurately, stabilizes $D_{pp}$, and achieves low rollout error compared to finite differences and Kalman–EM (expectation–maximization) with exact Van Loan discretization. Fisher-style checks confirm that diffusion needs at least one variance observable, and sparse $\sigma_{pp}$ “anchors” restore conditioning. We also show that the same PINN can learn time-varying HPZ coefficients.

Learned surrogates for expensive inner solver blocks are a widely pursued route to faster multiphysics simulation. We report a controlled, end-to-end negative result and identify three structural barriers, none of them a deficiency of the network we trained. The testbed is the implicit Newton solve coupling energy-dependent neutrino radiation to matter in a general-relativistic radiation-hydrodynamics code, its most expensive physics routine per call. First, per-call cost and share of runtime are different quantities, and only the second bounds acceleration. An exclusive self-time profile puts the target block at 16.9% of critical-rank wall clock, capping any surrogate at ~1.2x by Amdahl's law. A surrogate 5.8x cheaper per call merely ties the solver, and the configuration stable enough to run without fallback reaches only parity. Second, offline accuracy cannot rank surrogates for deployment: across fourteen networks the pooled Spearman error-versus-survival correlation (rho = +0.73) is a between-family confound that vanishes under control (rho = -0.04). Third, a correct out-of-distribution gate cannot accelerate a loop that leaves its training distribution. We give the break-even deferral fraction in closed form: because the visited states sit 73x off the data manifold, the gate defers 96.8 to 99.7% of cells, almost invariant to surrogate quality. Including its own cost, the gated loop is a 0.94 to 0.96x slowdown. We further separate stability from fidelity: a never-crashing gated run accumulates a linear -19.9% density bias over 6000 steps. The error is a directed, ballistically accumulating bias, not the variance-driven divergence the autoregressive literature targets.

Navigating large, photorealistic 3D apartments from raw pixels is widely considered infeasible for plain reinforcement learning. We build an agent that does it anyway, estimating its own pose from the camera alone. The agent has to reach several target objects in sequence, and their positions change between episodes, so it must explore to find them. It builds on our earlier object-centric topological controller, which still read the agent's true pose and its object detections from the simulator. Here we replace that true pose with an onboard, object-centric estimate. For each object we keep a bank of ORB features that, when the object is seen again, yield a rough pose measurement, which a minimal Extended Kalman Filter (EKF) fuses with a motion model. As on a real robot, the executed motions are noisy. The estimate drifts, but the agent and the nearby objects drift together, so a locally consistent pose is enough to follow each short edge and then home in visually on the target, which lets us replace full SLAM with a much smaller model, closer to how biological navigation appears to work. In the photorealistic Habitat simulator, the agent reaches its target objects from vision alone, with a pose that only needs to be locally consistent.

SIMON and SIMECK belong to a family of Lightweight Cryptographic Algorithms (LCAs) based on the Feistel block cipher, designed for Internet of Things (IoT) devices. As with all Feistel ciphers, they are susceptible to differential cryptanalysis, necessitating rigorous resilience evaluations. While state-of-the-art techniques leverage heuristics and sampling to improve efficiency, little work has applied Machine Learning (ML) guided Graph Representation Learning (GRL) to efficiently identify and visualise high-probability differential clusters. We address this gap by introducing an efficient feature engineering strategy that extracts four differential attributes from a partial Difference Distribution Table (pDDT), revealing structural information concealed in raw differential data. Utilising the enriched features, we construct and compare three ML-guided directed graphs for SIMON$32$ and SIMECK$32$ using K-Nearest Neighbour (KNN), Decision Trees (DT), and Random Forests (RF). To the best of our knowledge, our framework produces the first graph-based visualisation of the differential clustering effect, in which high-probability single-bit differentials form geometrically close clusters in the learned embedding. All three models achieve a precision of $1.0$ in identifying high-probability differentials, confirming zero false positives. KNN achieves the strongest cluster separation, the highest F1 score and the lowest graph construction time of approximately $2.3$ seconds, while DT and RF produce optimal paths with near-perfect regression. The results are consistent across both LCAs, demonstrating the applicability of the framework to other AND-rotation LCA families.

Semi-structured $N$:$M$ sparsity has emerged as a practical direction for accelerating large language models (LLMs). However, existing learnable-mask approaches incur substantial parameter and memory overhead, limiting their scalability to large models and aggressive sparsity regimes. In this work, we revisit semi-structured pruning from a perspective that reconciles efficiency with scalability. We propose Reservoir of Importance (RoI), a lightweight semi-structured pruning framework that learns sparsity masks through differentiable subset sampling. Unlike prior methods that model full categorical distributions over all feasible $N$:$M$ patterns, RoI introduces a compact-logit parameterization for sparsity mask learning and performs sampling without replacement to select masks, thereby reducing trainable parameters from combinatorial complexity to $\mathcal{O}({M})$. As a result, RoI requires 1.5-8.75$\times$ fewer learnable parameters and significantly lower memory cost, while remaining fully aligned with hardware-friendly sparsity patterns. Extensive evaluations across multiple scales of the Qwen2.5 LLM family (0.5-7B parameters) demonstrate that RoI achieves competitive performance with strong memory efficiency, stability, and scalability to more aggressive $N$:$M$ sparsity patterns, offering a practical path toward efficient LLM deployment.

LLM agents remain unreliable on long-horizon tasks, where small local failures can compound over extended interactions and lead to overall task failure. Although external harnesses can substantially improve robustness, harness design remains a manual and expensive process that requires searching over a large space of prompts, tool configurations, and control logic. We propose AutoSaddler, an automatic harness optimization framework that formulates harness improvement as an offline learning problem and iteratively updates the harness using failure signals from mini-batches. AutoSaddler combines failure-trace diagnosis, structured patch generation that treats the harness as code, and validation-based update selection. Experiments on GAIA2, SWE-Bench Pro, and Terminal-Bench 2.0 show that AutoSaddler substantially improves agent performance over the corresponding base harnesses, achieving gains of 9.0, 9.6, and 10.0 percentage points, respectively. Ablation studies further suggest that effective harness optimization benefits from three ingredients: deep debugging rather than shallow reflection, targeted modifications rather than unconstrained editing, and generalization-aware selection rather than trajectory-specific repair. Together, these results suggest that automatic harness optimization is a promising path toward more performant and reliable agent systems.

Controlling restricted knowledge in large language models is essential for model alignment and safe deployment. Test-time unlearning avoids costly retraining and parameter updates by intervening only during inference. However, existing activation-editing methods apply isolated pointwise corrections, overlooking how autoregressive generation continually reconstructs hidden states from the prompt, cache, and generated prefix. Consequently, later states may return to restricted knowledge regions after a locally successful correction, causing restricted knowledge re-entry. In this work, we propose Stateful Test-Time Unlearning via restricted knowledge boundary control (ST$^2$U), which formulates test-time unlearning as trajectory-wide boundary control. ST$^2$U first models restricted knowledge boundaries in low-dimensional invertible coordinates while leaving orthogonal non-target components unchanged. During inference, ST$^2$U monitors risk along the trajectory, applies minimal boundary corrections with contextual anchoring, and propagates historical correction states across tokens to mitigate knowledge re-entry. This trajectory-wide control enables more persistent forgetting while preserving non-target capabilities and limiting inference overhead. Across three benchmarks and three model families, ST$^2$U delivers the strongest overall balance, combining best or second-best retention with competitive forgetting and substantially less restricted-knowledge re-entry than test-time baselines (13.76%-19.84% versus 46.50%-59.10%).

Federated Learning (FL) enables distributed clients to collaboratively train models without sharing raw data, making it promising for leveraging massive devices in communication networks. In distillation-based FL, each client applies its local model on an unlabeled public dataset, and shares only prediction results with the server. While heterogeneous local data introduces label distribution skew, thus biasing client models toward majority classes and leading to potentially inaccurate predictions. The lack of ground-truth labels in the public dataset hampers the server's ability to calibrate predictions, which ultimately degrades overall performance. To address this, we propose FedCC, a simple and effective algorithm for mitigating client misclassification. Instead of being forced to classify and risking error propagation, clients are allowed to tag ambiguous samples as 'unknown'. This additional class, together with calibrated pseudo-labels on the public data, balances confidence in majority classes against uncertainty in under-represented ones. Extensive experiments demonstrate that FedCC significantly outperforms existing methods, especially under severe label skew. In the extreme scenario where each client holds samples from only one of ten classes, FedCC achieves 67.3% accuracy, while baselines collapse to near-random results.

The dynamics of particles in the early universe are described by Boltzmann equations, which involve high-dimensional phase-space integrals. Classical approaches use quadrature integration and evolve the system on a fixed momentum grid, which scales poorly to complicated systems and parameter scans, severely limiting the complexity of processes that can be studied. We introduce Neural Boltzmann Equations (NBEs), which combine three coupled concepts to overcome these limitations. First, particle properties are encoded in physics-inspired neural distribution functions, with parameters that can be predicted using neural networks, enabling efficient parameter scans. Second, phase-space integrals are evaluated with Monte Carlo, using importance sampling tools from collider physics. Third, we use the natural gradient method to evolve the system. After demonstrating the individual benefits of NBEs, we use the framework to perform the most numerically precise calculation to date of the effective number of relativistic neutrino degrees of freedom, $N_\mathrm{eff}$, in the Standard Model.

Federated fine-tuning of on-device large language models (LLMs) faces a significant computing burden. To overcome this limitation, split learning (SL) has emerged as a promising solution, which offloads the primary training workload to a powerful server. However, SL requires exchanging high-dimensional activations and gradients between clients and the server, resulting in prohibitive communication costs. To overcome this challenge, we propose SplitLite, a communication-efficient split federated LoRA fine-tuning method that exploits the low effective rank structure of consecutive-epoch activation and gradient residuals. Our key finding is that, when LoRA uses rank $r$ updates in parameter space, the activation and gradient residuals of the same data sample between adjacent epochs also exhibit effective rank-$2r$ and rank-$4r$ structures, respectively. By revealing this property, SplitLite transmits only quantized truncated singular value decomposition (SVD) residual factors, thereby significantly reducing both activation uplink and gradient downlink traffic. Extensive experiments on the GLUE benchmark across a series of advanced on-device LLMs demonstrate that our method reduces activation uplink communication costs by up to 93.5% and total communication costs by up to 83.7%, without performance degradation.

Training large language models is costly. How low a loss the same compute can ultimately reach depends on how each step's gradient is converted into a weight update; the rule that performs this conversion is the optimizer. From SGD and AdamW to the recent Muon, effective update rules have mostly been shaped by engineering intuition and then selected on benchmarks. Muon semi-orthogonalizes the momentum matrix before applying the update and has kept breaking records on public training benchmarks; yet why the semi-orthogonalized direction works, and over how long a history the momentum should average, are two questions at present answered mainly by experience. Here we treat the weight matrix during training as a responsive medium with memory and build a physical model for it, in which both questions find answers: the semi-orthogonalized direction is the maximally dissipative response under an output-side safety budget, which explains why it works; momentum is the internal stress accumulated by the medium; how long it should average is set by the relaxation of this stress, and a real medium relaxes on more than one timescale, the simplest form being one fast and one slow. On this basis we propose the Bi-Maxwell optimizer. The framework further yields a testable consequence: gradient directions change fast early in training and more slowly later, so the optimal memory length should grow with training stage; step-by-step measurements of a proxy for it by a read-only probe across 8 independent training trajectories are consistent with this consequence. Replacing the memory kernel alone, from a single timescale to two, brings training to the target loss in noticeably fewer steps on a public large-language-model optimizer benchmark.

Dynamic spectrum access must coordinate secondary users under bursty primary-user activity while preserving packet reliability and delay. We present TACAN, a centralized policy that represents each channel as a token containing occupancy history and automatic-modulation-classification entropy; a context token supplies queue class, delay and user identity. A Transformer encoder is warm-started from an occupancy-greedy policy and refined with proximal policy optimization. The frozen policies were trained to maintain a channel assignment in every slot, including when queues were empty. We therefore replay them on held-out trajectories and distinguish standby assignment success from packet-present access and packet delivery. In a 20-channel network with 60 primary devices and 4 secondary users, TACAN achieves 92.53% +/- 0.47 packet-present access success, compared with 89.94% for Greedy and 83.53% for PPO+MLP. Its paired gain over Greedy is 2.59 points (parametric 95% CI 1.89-3.29), with wins in all five seeds; the exact two-sided sign-test value is 0.0625. The gain rises from 0.57 points at normal primary-user load to 7.67 points at extreme load. TACAN also reduces mean delivery delay from 1.208 to 1.123 slots and the conditional user-reliability gap from 9.69 to 3.15 points. Delivered packets per SU-slot remain arrival-limited (30.12% versus 30.11%), so no packet-throughput gain is claimed.

Knowledge graphs are often accompanied by ontological class hierarchies that encode valuable semantic information, yet many link prediction methods either ignore such hierarchies or incorporate them indirectly through additional graph edges. Recent work introduced hierarchy-aware graph neural networks (GNNs), which use semantic losses derived from box embeddings to encourage satisfaction of subclass relationships during GNN-based representation learning. While this approach has shown promise for biological regression tasks, its effectiveness for knowledge graph link prediction has not been investigated. In this paper we evaluate hierarchy-aware semantic losses on link prediction across three benchmark datasets: AIFB, CoDEx, and BioKG. We combine graph neural network encoders with box-embedding-based semantic losses that encourage learned representations to better satisfy ontology-derived class hierarchies, and compare this approach to both standard link prediction models and models incorporating subclass relations as graph edges. Across all datasets, hierarchy-aware semantic losses significantly improve mean reciprocal rank (MRR) and consistently outperform models that incorporate hierarchy information through additional subclass edges. Relative to the baseline GNN models, MRR improved by 7.6%, 2.4%, and 15.5% on AIFB, CoDEx, and BioKG, respectively. Furthermore, semantic losses consistently outperform the alternative of augmenting the graph with subclass edges. These results are consistent with ontology-derived class hierarchies providing complementary information to graph structure, and suggest that encouraging hierarchical consistency through semantic losses is an effective and comparatively parameter-efficient mechanism for improving knowledge graph link prediction.

Industrial monitoring models must detect operationally relevant deviations while satisfying target-specific data, calibration, and resource constraints. Time-series foundation models (TSFMs) promise reusable representations and zero-shot forecasts, yet evidence for their deployment value remains mixed when task definitions are heterogeneous and lightweight baselines are competitive. This work presents a protocol-aware empirical assessment across three settings: a C-MAPSS degradation-risk proxy, normal-only training for anomalous-sound detection on MIMII, and BDG2 forecasting-residual diagnostics with synthetic target perturbations. We assess classical one-class methods, compact neural autoencoders, residual forecasters, MOMENT-small, Chronos-T5, and TimesFM 2.5 in terms of anomaly-ranking performance, risk-horizon sensitivity, residual forecasting and perturbation sensitivity, and local implementation cost. Across 100 C-MAPSS engines evaluated out of fold, TCN-AE reaches fold-weighted AUROC/AUPRC 0.9570/0.8960, compared with 0.7310/0.3080 for MOMENT reconstruction; paired engine-cluster bootstrap confidence intervals exclude zero for both differences. Across five matched MIMII pump evaluations, OCSVM also exceeds MOMENT reconstruction in AUROC and AUPRC. On a fixed 12-meter BDG2 panel, TimesFM 2.5 has the lowest aligned forecast error and the highest synthetic AUROC point estimate, although synthetic AUPRC is similar across TSFM and fitted residual models. Same-device measurements show that MOMENT incurs higher latency, peak allocated VRAM, and serialized state-dictionary size than TCN-AE. Under the evaluated frozen and zero-shot settings, TSFMs are task-dependent deployment options rather than default replacements for fitted lightweight models.

We analyze a variant of stochastic gradient descent with initial regularization (SGDIR) and derive dimension-free upper bounds on its expected excess risk for the squared loss. In the noiseless case, we obtain new bounds for both averaged and non-averaged SGDIR under moment, source, and capacity assumptions. For a particular value of the source parameter, these bounds are of order $m^{-2}\log^{2}m$, where the number of training samples is of order $m$. For another value of the source parameter, we obtain, for any $\epsilon>0$, bounds of order $m^{-3+\epsilon}$, provided that the capacity parameter exceeds $\epsilon^{-1}$. We also establish a lower bound that matches our upper bounds in certain regimes up to a polylogarithmic factor. In the noisy case, we provide an instance-based comparison between SGDIR and ridge regression. Under general assumptions and a mild lower bound on the regularization parameter, we show that the expected excess risk of SGDIR is no larger than that of ridge regression, up to a polylogarithmic factor. Numerical experiments on synthetic and real data are consistent with our theoretical findings.

Federated learning has traditionally been formulated as a single-objective optimization problem, primarily focused on maximizing model utility. In real-world applications, however, machine learning models often need to optimize multiple and potentially conflicting objectives simultaneously. This motivates federated multiobjective optimization (FMOO), which provides a natural framework for jointly handling multiple task-specific objectives in federated learning. In this paper, we propose a momentum-based variance-reduced algorithm for federated multiobjective optimization. The method incorporates a momentum-driven gradient estimator into the local updates to reduce the variance of stochastic updates, leading to an improved convergence rate. We establish theoretical guarantees showing that the expected Pareto stationarity measure of a randomly selected output iterate decays at a rate of $\mathcal{O}(T^{-2/3})$, improving upon the $\mathcal{O}(T^{-1/2})$ rates established for existing methods such as FSMGDA and FedCMOO. Numerical experiments on federated multiobjective optimization benchmarks demonstrate the effectiveness and competitive performance of the proposed algorithm.

Multi-behavior recommendation (MBR) leverages auxiliary behavioral signals, such as clicks and add-to-cart, to enhance target behavior prediction like purchases. While recent graph neural network-based approaches have achieved strong performance by systematically propagating auxiliary behavior signals, they still suffer from two fundamental challenges inherent to auxiliary behaviors: (1) missing auxiliary signals, which hinder generalization to items without auxiliary observations, and (2) unreliable auxiliary signals, which amplify noise misaligned with the target behavior. To address these challenges in a unified manner, we propose BOAR, an environment-conditioned MBR framework that addresses missing and unreliable auxiliary signals through two complementary modules conditioned on auxiliary observability. Extensive experiments demonstrate that BOAR consistently outperforms state-of-the-art baselines, achieving up to 7.82% gains in HR@10 overall and up to 44.2% gains for target items without auxiliary observations, highlighting its ability to capture hidden preferences beyond observed auxiliary relations. Our code is available at: https://github.com/LSH0411/BOAR.

Inference-time pipelines often sample multiple outputs, filter them with a learned safety model, and return the proxy-feasible output with the highest learned reward. We show that this composition creates a two-stage failure: an imperfect safety proxy first contaminates the feasible set with unsafe outputs, and reward maximization can then amplify this residual contamination. We define safety hacking as selecting an output that passes the learned constraint but violates the true safety criterion. For constrained Best-of-$N$ sampling, we derive finite-$N$ bounds governed by the joint upper reward tails of safe and unsafe outputs within the proxy-feasible set. If unsafe-but-feasible outputs have the heavier tail, safety hacking becomes asymptotically certain as $N$ grows, even when false-positive mass and average safety- and reward-proxy errors are arbitrarily small. We also show that policies within a bounded $\chi^2$ divergence from the proxy-feasible reference distribution admit an $N$-independent safety-hacking bound, and instantiate this general coverage-control principle with constrained pessimistic sampling. Coverage control limits amplification but cannot repair a contaminated feasible set: admitted unsafe outputs may still be favored, and regularized selection is not necessarily safer than constrained Best-of-$N$ for every reward proxy. Toy and language-model experiments characterize both contamination and its reward-tail amplification, which exposes an inherent difficulty in inference-time scaling with learned safety models.

We develop a finite-width geometric framework describing how learned feature geometries are organized, transported, and selectively aligned in deep neural networks. Incompatibility among weight-generated covariance, gates, and backward sensitivities is quantified through three families of commutators: between gates and covariance, between sensitivities and covariance, and between average gradient outer products (AGOPs) and neural feature matrices (NFMs). An exact layerwise identity decomposes the sensitivity-covariance commutator into four sources: downstream transport, adjacent-layer imbalance, pointwise sensitivity fluctuations, and nonlinear gate-covariance interactions. The AGOP-NFM commutator is a singular-value-weighted transport of the internal commutator, explaining why observed feature-side alignment alone does not determine the internal geometry from which it emerges. Buffered localized energies resolve mixing between separated covariance subspaces. We establish spectral-gap, projector-evolution, and stabilization estimates, and formulate conditional Lyapunov principles that yield decay under explicit geometric error-bound or intrinsic-damping assumptions. These criteria do not follow from gradient flow alone and clarify why risk reduction need not imply commutator collapse. Analytic examples and numerical experiments exhibit factorization of spectral and activation geometry, transient growth, and cancellation among nonzero sources. In tested finite-time regimes, cancellation dominated by a negative transport-imbalance interaction persists across depths, widths, and two regression benchmarks. Spectral alignment therefore appears as a layer- and scale-dependent compatibility phenomenon governed by transport, interaction, cancellation, and possible damping, rather than a universal consequence of training.

Hybrid sequence models must satisfy prefix invariance: representations at position t must not depend on future inputs, yet this is rarely verified. We formalize prefix invariance and give a lightweight audit, two forward passes, no training or gradients, yielding a per-layer score localizing where causality breaks. Attention-mask inspection, the field's default check, is incomplete: causality is a graph-level property, and leaks can occur via scans, aggregations, or normalization despite correct masks. Across 192 injected-fault trials on eight checkpoints, mask inspection detected none, while our audit localized all 192/192 to the exact layer. Static/dynamic analysis of chunked-scan code in transformers found the same defect in Zamba2 and Nemotron-H, an inter-chunk axis error fixed via the reference implementation. The method fits on one page and runs in seconds.

Neurobiological studies and representation learning have observed that representations of objects belonging to the same category in high-dimensional neural spaces exhibit low-dimensional object manifold characteristics, and different object manifolds are linearly separable in these neural spaces. However, these experimentally observed phenomena lack rigorous theoretical validation to date. This paper proposes a new stochastic separability theorem for embedding manifolds of two different object categories. First, we establish a projection measure concentration theorem for embedding manifolds under general conditions. We develop a new two-layer measure concentration analysis technique, which unifies two estimation bounds via the law of total expectation to derive measure concentration inequalities. Based on the measure concentration theorem, we further prove a stochastic separability theorem for embedding manifolds of two different object categories. If two datasets have distinct means and bounded total variances, their samples become linearly separable with high probability, provided that the projection direction satisfies a non-singularity condition. The main contributions of this paper are twofold: 1. We prove the projection concentration properties of embedding manifolds in high-dimensional spaces by using two-lawyer tail-bound inequalities. 2. We identify a non-singularity condition for the stochastic separability between embedding manifolds, and rigorously prove the stochastic projection separability theorem. The theorem not only uncovers geometric and statistical properties of the object embedding manifolds, but also provides a novel mechanism for representation learning in deep networks.

Existing data pruning methods predominantly rely on high-dimensional feature embeddings to measure sample importance. However, these compressed vectors often obscure fine-grained semantic interactions, leading to suboptimal coverage of rare semantic concepts in the pruned subsets. In this paper, we propose Mapping the Concept Landscape (MCL), a novel structural perception framework for transparent data pruning. Instead of abstract embeddings, we represent each image-caption pair as an explicit sample-level graph comprising entities, events, and attributes. By integrating these individual graphs into a comprehensive dataset-level graph, we characterize the global distribution of semantic concepts and quantify their rarity across the entire corpus. Based on this structured perception, we develop a greedy concept-coverage maximization algorithm that iteratively selects samples to maximize the marginal gain of high-value, under-represented concepts. Experimental results on various benchmarks demonstrate that our method not only achieves superior pruning efficiency compared to state-of-the-art methods but also provides a transparent and interpretable audit trail for the selection process.

Practical deployment of large language models (LLMs) requires families of post-trained variants—instruction-tuned, reasoning-tuned, and chat-style models—each at multiple sizes to meet diverse latency and memory budgets. Producing each (variant, size) pair independently is prohibitive, so model families typically span only a handful of coarse-grained sizes per post-trained variant. Boomerang distillation (Kangaslahti et al., 2026) reduces this cost along the size axis for base models. Through model size interpolation, it constructs models of intermediate sizes from a single teacher-student pair without additional training. However, it still treats each post-trained variant as a separate object of optimization. We introduce ADAPT—Amortized Distillation Across Post-Trained LLMs—a framework for amortizing distillation across both axes of a model family: size and post-training variant, producing $L \times K$ models for $L$ interpolated sizes across $K$ post-trained variants with a single distillation run. ADAPT combines two components. First, a two-phase distillation procedure constructs post-trained students through pre-training alignment and supervised fine-tuning distillation, enabling smooth size–performance interpolation on generation and reasoning tasks. Second, weight-delta initialization approximates this construction across post-trained variants by transferring the distillation-induced weight change from the base model to students initialized from different post-trained variants. The resulting continuum of interpolated models also enables adaptive model-size selection at inference time, improving the compute–accuracy trade-off for long-form reasoning tasks.

Full-length RNAs, particularly messenger RNAs, often exceed the context lengths used to pretrain existing RNA foundation models, limiting complete-transcript modeling at single-nucleotide resolution. We present RIBOSPAN, a 1.61-billion-parameter bidirectional RNA foundation model natively pretrained with context lengths up to 10,240 nt. RIBOSPAN combines dense bidirectional self-attention, single-nucleotide tokenization, and attention-isolated sequence packing to enable high-resolution modeling of complete long RNAs. Native 10K pretraining preserves strong reconstruction at 10,240 tokens and, in a controlled long-context benchmark, maintains strong contextual responsiveness and context-specific representation separation while keeping perturbation-induced changes highly localized. Inference-time YaRN scaling recovers much of the contextual organization lost by direct short-context extrapolation, but induces substantially greater distal representation diffusion. Frozen RNA-type evaluations show that RIBOSPAN learns state-of-the-art RNA representations, with a particularly clear advantage on long RNAs. Across downstream biological benchmarks, RIBOSPAN emerges as the strongest encoder-only RNA foundation model, achieving state-of-the-art performance in both full-transcript biological property prediction and zero-shot mutation-fitness modeling. Building on the same backbone, we develop a multidimensionally conditioned discrete-diffusion framework for full-length mRNA generation and redesign, including synonymous-codon diffusion for protein-preserving CDS optimization. Together, RIBOSPAN establishes a powerful long-context foundation for transferable RNA representation learning, biological prediction, and full-transcript mRNA design.

Large language models (LLMs) are increasingly used for scientific decision support, yet reliable confidence estimation remains difficult in black-box settings. We study uncertainty estimation for hierarchical taxonomic reasoning generated by a black-box LLM in a long-tailed biodiversity monitoring pipeline. Using proxy features extracted by an open-source tool LLM, we train lightweight supervised estimators with hierarchy-aware supervision to predict rank-wise correctness. Across three tool LLMs, the supervised estimators consistently outperform a token-likelihood baseline for micro discrimination and selective prediction under a single global rejection threshold, improving micro AUROC from 0.57 to 0.75–0.80. The best results are achieved by a rank-specific multi-head design (H3), suggesting that accounting for hierarchical output structure is important when a unified abstention rule is required. Our code is publicly available at https://github.com/uoguelph-mlrg/hierarchy-aware-llm-uq

We derive convergence guarantees for mirror descent and proximal mirror descent with a logarithmic barrier as the distance-generating function. When the solution lies on the boundary, the Bregman divergence blows up and classical guarantees become vacuous. While existing remedies recover $O(\log k/k)$ rates by adapting classical bounds, we propose a more direct route in which the Bregman distance to the solution never appears, and which achieves the same rate with constant stepsize. We also construct an example showing that this rate cannot be improved in general. Finally, we compare the proposed approach with interior-point methods.