arXiv subjects
Search indexed arXiv papers on artificial intelligence, large language models, computer vision and robotics. Read source abstracts and follow links to arXiv.
We introduce the roots of supercritical Leung--Yau--Zaslow pairs and prove pointwise strong positivity of the associated Schur forms. The resulting LYZ cones contain the pliant cone on smooth projective threefolds, with strict inclusion on a Fujino--Sato threefold. We also establish Hard Lefschetz and Hodge--Riemann relations for the associated Chern classes on compact Kähler manifolds in bidegrees $1\le p+q\le\dim X$ with $\min\{p,q\}\le1$.
Topological deep learning (TDL) architectures are typically evaluated on synthetic complex pairs that differ in a target invariant. Yet, whether a model recovers that invariant or an artifact of the chosen discretization is seldom audited. To examine this, we design and test pairs of complexes that look identical to standard homological and local message-passing tests, have the same Hasse graphs, yet differ in the target invariant. We study two beyond-homology regimes: nonAbelian fundamental-group structure and homotopy-matched orientability. For the former, we construct genus-2 pairs that match homology, cell statistics, local incidence profiles, and rank-colored Hasse 1-WL behavior while differing in their fundamental groups. For the latter, we evaluate punctured torus-Klein bottle pairs and show that distance, boundary, cycle, and spectral signals can confound apparent orientability detection, motivating stronger matched controls. Under the tested settings, subcomplex-lifted architectures separate some pairs beyond the matched homological, local-incidence, and Hasse 1-WL controls, while a depth sweep reveals an empirical depth-dependent trend without establishing a general barrier or direct computation of the target invariants. These results show that claims of topological expressivity must be stated relative to the model representation, explicitly controlled information channels, and the precise scope of the benchmark.
Transformer-based models have shown strong performance across diverse tasks, but their reliance on self-attention mechanisms often leads to high computational costs. Token sparsification, which reduces the number of tokens processed in transformer layers, offers a promising solution to maintain performance while improving efficiency. However, although widely studied in the vision and language domains, token sparsification remains underexplored in audio, with existing work often limited to a single method, model size, or dataset. This work presents a comprehensive evaluation of token sparsification techniques for audio transformers. We assess a range of token pruning and merging strategies across multiple datasets and model sizes, under both post-training and fine-tuning settings. Furthermore, we introduce a novel sparsification approach that combines attention scores with spectral cues derived from the frequency content of token embeddings. Our experiments show that frequency-aware approaches, particularly those based on spectral flatness and low-frequency energy, consistently outperform attention-only methods. We also find that sparsification based on spectrogram patch intensity is especially effective for longer audio inputs. These results highlight the value of incorporating frequency-domain and input-specific information into sparsification strategies to enhance performance in audio tasks.
Air quality policy will cut aerosol and ozone precursor emissions, and studies disagree on how the tropical Pacific circulation and the El Nino Southern Oscillation (ENSO) will respond. Here we show, in 12 Earth system models with and without air quality controls, that the cuts strengthen the Pacific trade winds and offset half to two thirds of the weakening projected without them. The response is absent with prescribed sea surface temperatures and detectable only after 2040. It is consistent with the Indian Ocean warming relative to the Pacific, which also ranks the responses to cuts made region by region. ENSO amplitude changes by less than a tenth of its value, the smallest change this ensemble could detect. The change implied by the mean state shift is a sixth of that, and an ENSO model trained without the reduced emission runs predicts none. El Nino's regional impacts still change, because the background climate does.
AI weather models perform competitively with traditional physics-based models at a fraction of the computational cost, and they are thus already used operationally. However, studies reach conflicting conclusions about their ability to forecast extreme weather events, and comparisons across studies are complicated by differing evaluation protocols. We discuss the challenges that arise when comparing AI and physics-based weather forecasts for extreme events. Within a common experimental design, we find that model rankings are highly sensitive to the choice of evaluation metrics: the considered physics-based model tends to perform best when evaluation is restricted to observed extreme events, whereas AI models generally perform better when evaluated using weighted scoring functions. A contingency table analysis shows that AI models underpredict the frequency of extremes, so that their scores are generally dominated by misses, while the physics-based model forecasts extremes more frequently and its scores are dominated by false alarms. Different evaluation metrics penalise these errors differently. Further score decompositions confirm that the score differences arise primarily from miscalibration rather than discrimination ability. This sheds light on the contrasting results in the literature, and highlights the importance of the evaluation metrics used to evaluate AI models.
Reinforcement learning has emerged as an effective approach for combinatorial optimization, but most existing methods focus on problems such as the Travelling Salesperson Problem (TSP) in which all elements must be served. Selective combinatorial optimization problems (SCOPs), including the Orienteering Problem (OP), Prize-Collecting TSP (PCTSP), Stochastic PCTSP and Knapsack Problem (KP), additionally require selecting which elements to include. This selection creates sampled solutions with substantially different objective values, yet standard preference optimization reduces such comparisons to binary rankings and discards the magnitude of these differences. We introduce Reward-Weighted Preference Optimization (RWPO), a pairwise training objective that incorporates reward-gap information into preference learning. For candidate trajectories sampled from the same instance, RWPO weights each preference loss according to the normalized difference in their rewards, assigning greater training emphasis to comparisons that represent larger improvements in solution quality. We further combine RWPO with scale-specific fine-tuning and reward- and likelihood-weighted tilted sampling to adapt trained policies to larger and shifted instance distributions without relying on local-search post-processing. Experiments on the OP, PCTSP, Stochastic PCTSP, and KP show that RWPO consistently improves over unweighted preference optimization and policy-gradient baselines across a wide range of instance sizes. On the OP, the resulting solver exceeds the mean reward of the Compass metaheuristic at scales from 100 to 2,000 nodes. The results demonstrate that preserving reward-gap information can substantially improve preference-based learning for selective combinatorial optimization and support generalization to large-scale instances.
Cerenkov Luminescence Microscopy (CLM) enables the visualization of radiotracer uptake in three-dimensional (3D) tumor spheroids but is limited by low photon statistics and spatial blurring caused by positron transport. Here, we present a Monte Carlo (MC)-guided deep learning framework for CLM image restoration. Using TOPAS/Geant4, we developed a physical model of heterogeneous 3D tumor spheroids to simulate positron transport, Cerenkov photon generation and optical propagation. A dataset of 15,360 paired degraded and idealized ground-truth (GT) images was generated from 80 simulated spheroid geometries and used to train a supervised Content-Aware Image Restoration (CARE) network. Compared with the self-supervised Noise2Void (N2V) denoising method, CARE achieved substantially lower reconstruction errors on synthetic validation data, reducing the mean squared error from 0.0492 to 0.0051. Training stability and noise robustness analyses further demonstrated high reproducibility of the restored images. The trained CARE model was subsequently applied to experimental CLM images acquired from 27 U87 tumor spheroids following 18F-FDG uptake. The restored images exhibited more localized intensity distributions and higher local contrast and signal-to-noise ratios compared with raw and N2V-processed images. These findings, aligned with the 3Rs principles of animal experimentation, demonstrate the potential of physics-based supervised restoration to mitigate both noise and positron-induced spatial broadening in CLM.
We study the long-time Fourier distribution of driven passive scalars transported by an incompressible alternating shear flow (the Pierrehumbert flow) subjected to IID random phase shifts, in the regime of strong shear-strength. With a novel and self-contained proof, we establish (i) a quantitative version of the cumulative Batchelor's law; (ii) quantitative negative regularity mixing estimates; and (iii) a strong form of `sparsity' -- unevenness in the distribution of Fourier mass in radial sectors in frequency space. Our method uses the strong shear-strength assumption to handle a purely deterministic scalar driving term.
The implication of entanglement in quantum battery charging remains poorly understood. In addition, intrinsic quantum randomness poses a fundamental challenge to charging uniformity across distinct trials. To address these issues, we study a three-qubit system within a cascaded collision frame, with two qubits acting as the charger and the third serving as the battery. After each collision, an adaptive measurement is performed on the charger, yielding four conditional post-measurement battery states with identical expected stored energy. For arbitrary fixed charger states, we derive analytical expressions for the recurrence relation of stored energy. We find that stored-energy fluctuations depend on whether the dynamics relies on the battery phase \(γ_n\). For two Bell-like families of entangled states, together with their computational basis states termed Bell-like separable states, the dynamics is \(γ_n\)-independent and fluctuations vanish, whereas for other entangled and separable states, fluctuations emerge. These insights establish a paradigm for engineering charger states to design uniform-charging quantum-battery architectures despite inherent quantum randomness.
Thick, dispersion-engineered silicon nitride is a leading platform for integrated third-order nonlinear optics, supporting supercontinuum and frequency comb generation. However, it lacks the second-order nonlinearity required by many applications. Here we bond pre-fabricated x-cut thin-film lithium niobate membranes onto finished silicon nitride waveguides as a back-end step, compatible with commercially fabricated waveguides. A single silicon nitride waveguide generates an octave-spanning supercontinuum and, through the lithium niobate membrane, concurrent second-harmonic, sum- and difference-frequency generation. These results open new possibilities for thick silicon nitride photonics, from visible light generation to on-chip $f$-$2f$ comb self-referencing, a key step towards miniaturised timing and navigation quantum technologies.
Improving the efficiency of a stochastic search often requires to provide the searching agent with resources that modify its dynamics and help it explore space more quickly. Although such resources are often assumed to be unlimited and treated as a cost to mitigate, many realistic search processes operate under a finite resource budget, e.g. in processes with environmental memory in which the searcher deposits long-lasting cues along its path that it can interact with at later times to guide its further search. Here, we address the question of how to allocate a limited budget to maximize search efficiency by investigating a one-dimensional diffusing agent whose access to resources is constrained. We consider three resource models, namely an external force applied by an external controller to pull the searcher towards the target, directional flags deposited by the searcher along its path that redirect its motion towards the target, and self-generated chemorepulsive cues that act as a force field. For all these models, we find that resource usage should be concentrated in narrow regions of space when information about the searcher's current and past positions is available. In contrast, in blind strategies where a predefined time protocol is followed, resources should be used at a steady rate to ensure it to be used at most of its capacity by the time it reaches the target. We further show that blind strategies in non-Markovian processes with environmental memory are significantly less efficient due to the inherent fluctuations in the locations of deposited cues.
Structures with large global contributions to a flow field are not necessarily those that are most closely associated with a specified engineering observable. We propose a time-lagged cross-correlation (TCC) ranking criterion to extract target-related dynamical structures from a fixed modal expansion while preserving existing spatial modes and temporal coefficients. The criterion ranks mode groups by the mean squared norm of cross-responses between their flow contributions and the observable over a lag window. It is tested separately using dynamic mode decomposition (DMD) and proper orthogonal decomposition (POD). TCC-DMD analysis shows that TCC recovers three low-contribution target structures from a dominant background in a synthetic flow. The reconstruction error for the target subfield is 2.15\%. In the cylinder wake, different force observables select structures associated with the vortex-shedding fundamental, slow drag variations, and the second harmonic. In the cavity case, selected modes connect two main frequency bands of the target observable to distinct spatial amplitude and phase distributions. TCC-POD analysis shows that the criterion minimises the weighted truncation error of the correlation operator among equally sized subsets of the fixed orthogonal basis. A comparison with canonical correlation decomposition further reveals differences between fixed-direction selection and correlation-based basis construction in the spatial structure and retention of flow information. Lag aggregation can improve structure identification when zero-lag association is insufficient. The gains depend on the cross-response distribution and window selection. TCC separates target-related ranking from mode-set construction, enabling comparison of spatial structures associated with different observables within the same modal expansion.
Multi-exit networks improve inference efficiency by allowing inputs to terminate early at intermediate exits. We identify a privacy vulnerability in cloud deployments of these networks: a malicious Cloud Service Provider (CSP) can infer sensitive properties of individual user inputs solely from inference latency, without accessing inputs, intermediate states, or model outputs. The attack exploits the CSP's control over the model supply chain. Prior to deployment, the CSP supplies a specially crafted multi-exit network manipulated so that inputs carrying a target sensitive property are routed to a designated exit with high probability, while other inputs exit elsewhere. At serving time, the CSP infers the exit index from per-query latency, thereby determining whether each input possesses the target property without accessing inputs, outputs, or internal model states. We propose a training framework combining selective label smoothing and entropy-covariance optimization to control confidence distributions across exits and reduce premature termination. We conduct experiments on four datasets across ten target properties. The results demonstrate strong attack effectiveness and stealthiness. The attack achieves an average F1 score of 0.73, compared to 0.25 for the strongest baseline, while main task accuracy drops by less than 1.3\% and inference speedup degradation remains below 0.1 in most settings. Our code is available at https://github.com/kazefjj/Weaponizing_the_Exit.
When clinical speech corpora are pooled for evaluation, corpus membership can itself become predictive if class prevalences differ across sources. We derive the AUC of an oracle predictor that uses only corpus identity and empirical prevalence, a diagnostic computable directly from corpus composition. Across five depression corpora in three languages (590 speakers), the corpus-only AUC is 0.697, and corpus identity is predicted from six acoustic and self-supervised representations with 97.5-99.8% accuracy. Permuting depression labels within each corpus removes label structure while preserving prevalences; the residual pooled-minus-LOCO discrepancy then increases across three nested configurations in the same order as the diagnostic. Pooled 5-fold cross-validation also outscores corpus-held-out evaluation across all six representations. Prevalence equalisation substantially reduces the discrepancy for eGeMAPS and WavLM-Base, whereas a sample-size-matched control does not. Corpus-specific base rates therefore create a measurable shortcut opportunity in pooled clinical-speech evaluation, motivating corpus-aware diagnostics alongside pooled performance.
A strong direct tree model may remain the best standalone photovoltaic forecaster, yet a weighted combination can reduce its residual errors. We examine whether complementary temporal predictors can lower the tree's mean absolute error (MAE) over 15--240 minutes without weaker ensemble members degrading performance. The candidate bank contains LightGBM and seven neural implementations with different temporal representations. A combiner fitted to MAE has nonnegative weights that sum to one. Unhelpful candidates can receive zero weight, while the tree anchors the combination and provides a development-stage fallback. The main design shares weights across horizons; horizon-specific weights and alternative losses test whether further complexity is needed. On 56,078 windows from two co-located public PVDAQ systems, the historical shared-MAE policy reduces scaled MAE by 3.21\% relative to the tree. The latest bounded neural correction does not reduce MAE further than this static combination. The results support selective combination of complementary predictions in this bank rather than equal averaging or indiscriminate addition of modules. Both analyses reuse a previously inspected evaluation period. They do not establish independent replication, weather-specific causation or universal gains from eight experts.
Quantum metrology is ultimately shaped not only by measurements, but also by the dissipative dynamics that prepare and maintain the probe. Nevertheless, a general mechanism for enhancing both steady-state and transient precision in open quantum systems has been lacking. Here we show that adding a simple reset channel provides such a mechanism: reset acts as a minimal dissipative control that can enhance metrological precision without modifying the intrinsic Hamiltonian or applying continuous coherent driving. We derive a unified quantum-Fisher-information (QFI) reset-rate response identity, showing how reset enhances steady-state and transient estimation while revealing a sharp transient distinction. For static-parameter estimation in diagonal dynamics, reset leaves the leading short-time QFI unchanged; for the separate task of elapsed-time estimation, it generates an immediate drift contribution that can scale quadratically with the reset rate in the short-time limit. For frequency estimation, the enhancement is instead dominated by a genuinely quantum eigenvector contribution. These results establish reset as a simple dissipative resource for open-system quantum metrology.
We give a construction of the previously announced family of three-dimensional noncommutative Fano minifolds of Euler degree $72$. We prove that their defining $\mathbb{Z}$-algebras are Koszul and Artin--Schelter regular, and compute the Hilbert series of their section algebras.
Neural architecture search (NAS) on a new dataset often affords only tens of short training runs, so what the search knows before its first evaluation decides its efficiency. Yet most NAS methods start from scratch, guided at best by zero-cost proxies that are unreliable across tasks. QuaNAS treats a library of earlier searches in the same search space as explicit knowledge of uncertain reliability. Main effects of design decisions learned from this library and, secondarily, zero-cost proxies become priors over candidate architectures without any label of the new dataset. A Gaussian process starts at the prior optimum and re-estimates after every evaluation how far each prior is trusted. Under a protocol frozen before any confirmation data were seen, QuaNAS was evaluated once on 12 new image, time-series, text and tabular datasets. It lowered anytime regret on the short-training objective 2.7- to 3.9-fold relative to five pre-specified NAS baselines that start from scratch (Holm-adjusted p = 0.0012 each). Its final models matched (non-inferiority margin 0.5 pp) or exceeded matched-recipe reference models on nine datasets, as did the prior optimum, selected without a single evaluation. Post-hoc comparisons with established transfer methods given the same library trace the gain to the library: without it, the same Gaussian process had over twice the regret. They also show where each transfer mechanism works best: the pooled main effects of QuaNAS for weakly transferable tabular tasks, per-task surrogates for image, time-series and text tasks and for small or mismatched libraries. Finally, the main effects make the transferred knowledge inspectable.