Search arXiv⌕ Search

SEARCH · Search arXiv

Search Search arXiv

Search indexed arXiv papers on artificial intelligence, large language models, computer vision and robotics. Read source abstracts and follow links to arXiv.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 487 records · Page 27Linked to original sources
arXiv2026

Stability in stochastic hypergraph matching I: necessary and sufficient criteria

Stochastic matching on hypergraphs is an important topic for its versatility in capturing real-life systems, from living donor transplant to ride-hailing. Nevertheless, finding necessary and sufficient criteria for stability is a long-standing problem. One of the key difficulties is the fact that greedy policies, whilst maximally stable for stochastic matching on graphs, no longer achieve maximal stability region on hypergraphs. So far, no alternative families of policies with similar properties have been known. In this work, we introduce online assignment policies, in which each item is assigned to a matching hyperedge type upon arrival. We show that this is a good generalisation to greedy policies, by proving that they are maximally stable. Their natural amenability to analysis allow us to derive several necessary and sufficient criteria for stability, which generalise the known criteria for graphs. Furthermore, the constructive proof gives a maximally stable arrival-rate agnostic policy.

cs.DM↗
arXiv2026

ARCHER: Agentic Rule and Compliance Harness for Executable Regulations

Verifying building compliance requires validating thousands of rules against large Building Information Modeling (BIM) designs, which is laborious, capital-intensive, and unscalable. Existing Automated Compliance Checkers (ACCs) are often difficult to generalize across different scenarios, as they are typically developed for highly specific rule sets and use cases. In addition, many ACCs are proprietary, meaning the underlying verification code is not released to end users, so users cannot verify whether their regulatory intent can be accurately captured. We introduce ARCHER (Agentic Rule and Compliance Harness for Executable Regulations), a test-driven, deterministically orchestrated multi-agent program-synthesis harness that generates auditable verification code from regulatory Codes of Practice, enabling transparent, adaptable, and scalable compliance checking. To characterize what makes agentic synthesis work, we evaluate a taxonomy of six harnesses of increasing agentic sophistication across four backbone models, spanning realistic data-governance tiers (from frontier third-party APIs to a fully on-premise open-weights model) on a novel dataset derived from real-world compliance scenarios. ARCHER's deterministic multi-agent orchestration achieves the highest accuracy for every backbone, improving mean union accuracy by 82% over a naive single-pass prompting baseline. Our cost-accuracy analysis further shows that using the ARCHER harness, a self-hosted open-weights model can reach 97.8% of frontier-API accuracy at a quarter of the cost, making data-sovereign compliance checking practical.

cs.MA↗
arXiv2026

Some results on NIP groups and their Ellis groups

This paper has several parts. We begin by developing a theory of `piecewise (strong) f-genericity' in NIP groups, where we call a definable set piecewise (strong) f-generic if some union of finitely many translates of it is (strong) f-generic. We show that, in an NIP group, the definable sets that are not piecewise (strong) f-generic form an ideal. Our hope is that the corresponding piecewise (strong) f-generic types can provide a substitute in arbitrary NIP groups for the (strong) f-generic types of definably amenable NIP groups, and in the rest of the paper we give several applications. Two of the applications deal with the Ellis group of an NIP group. Let $T$ be an NIP theory, $G$ a definable group, and $M$ a model. In our first result we show that the size of the Ellis group of $G(M)$ is bounded above by $2^{|T|}$, independent of the choice of $M$, giving a substantial step towards the question of whether the isomorphism type is independent of $M$. In our second result, inspired by a theorem of Hrushovski, we show that, if $T$ and $M$ are countable and the formulas of $T$ have uniformly bounded VC-codensity, then the Ellis group of $G(M)$ has `finite Archimedean rank', ie its connected component is profinite-by-Lie. A crucial tool for us in both results is the recent result of Chernikov-Gannon-Krupiński and Basso-Zucker that the $τ$-topology on the Ellis group is Hausdorff. Finally, we use our techniques to obtain a `local' result valid in arbitrary NIP theories, without the assumption of uniformly bounded VC-codensity: for any `bi-invariant' formula $ϕ(x,y)$, the group $G/G^{00}_ϕ$ has finite Archimedean rank. More precisely, if the VC-codensity of $ϕ(x,y)$ is at most $δ$, then $G/G^{00}_ϕ$ is an inverse limit of compact Lie groups of dimension at most $(4δ)^2$. This connects to, though is different than, a question of Hrushovski.

math.LO↗
arXiv2026

Intruder-driven mirror energy differences between $^{29}$Cl and $^{29}$Mg studied with antisymmetrized molecular dynamics

To clarify the mirror energy differences (MEDs) of the proton-unbound nucleus $^{29}$Cl and their microscopic origins, we investigate the low-lying states of the $^{29}$Cl-$^{29}$Mg mirror pair using antisymmetrized molecular dynamics. The calculation reasonably reproduces the normal and intruder states of $^{29}$Mg, while suggesting alternative spin-parity assignments for $^{29}$Cl. The $1/2^+$ and $3/2^+$ states are predicted to form a nearly degenerate ground-state doublet with a small MED because of their similar intrinsic structures. In contrast, the $3/2^-$ and $7/2^-$ intruder states exhibit large negative MEDs and are assigned to the observed resonances at approximately 500~keV and 1.1~MeV, respectively. Their large MEDs originate from the reduced Coulomb energies associated with the stronger deformation and spatially extended proton distributions in the intruder configurations.

nucl-th↗
arXiv2026

Analytical Series Expansion for Efficient Gradient Evaluation in Multi-Qubit Optimal Control

Gradient-based quantum optimal control and the theory of operator evolution are rarely discussed together. In this Letter, we bridge this gap by showing that the gradient of the time-evolved propagator involves the Heisenberg evolution of local operators. We present a unifying framework by deriving from first principles the formal solution for the gradient under an arbitrary pulse parameterization, and show that the same construction extends to derivatives of any order. We obtain a series of nested, time-independent commutators weighted by time-dependent scalar coefficients. The commutators are thus computed once, and each optimization step only updates the scalars. The method is particularly suited for simulating optimal control tasks in quantum systems with local interactions, which is a common situation in large multi-qubit platforms. In this setting, light-cone arguments and sparse-Pauli truncation heuristics may be used to keep the number of relevant terms in the series small. We compare the computational cost required for the series with the Gradient Optimization of Analytic conTrols (GOAT) method, and, focusing on the problem of preparation of a GHZ state, demonstrate more than an order of magnitude speedup for a qubit ladder and a chain geometry.

quant-ph↗
arXiv2026

Stability in stochastic hypergraph matching II: weights, batch arrivals, and continuous time

Many real-life systems can be found as examples of stochastic matching on hypergraphs, such as production lines or assemble-to-order systems. Two common features are the number of items required may vary between matchings, and there may intermediary items which exist as a combination of other items and not of external arrivals. Both of these phenomena can be modelled by considering the weighted variant of stochastic matching. In this work, we formalise the notion of stochastic weighted matching on hypergraphs. We also allow batch arrivals, meaning multiple items of multiple classes may arrive at the same time, and in particular, the arrivals can be correlated between classes. Unlike the classical setting where items arrive at discrete time $t \in \mathbb{N}$, we allow arrival processes to take place in continuous time $t \in \mathbb{R}_{\geq 0}$. We then extend the results from Nguyen and Bušić (2026) to overcome the intricacies brought up by this new setting. This allows us to derive necessary and sufficient criteria as direct generalisations of those in the unweighted setting, which depend only on the per-class arrival rates. As such, the correlation between classes bear no differences. The constructive proofs also give a maximally stable, periodic-review, size-based, arrival-rate agnostic policy.

math.PR↗
arXiv2026

AgentSnare: Learning to Delay, Divert, and Defuse Autonomous Penetration Agents

Large language model (LLM) agents automate penetration testing through an observation-action loop, selecting actions based on observations returned by tools. This dependence allows defenders to inject deceptive observations that can mislead the agent's decision-making process. However, existing defenses rely heavily on static, isolated artifacts planted in the environment prior to an attack. Advanced agents can progressively recognize and bypass these artifacts, ultimately refocusing their exploitation attempts on the real target. To address this issue, we introduce AgentSnare, a trajectory-adaptive deception system that dynamically unfolds a decoy environment to continually steer the penetration agent away from the real target. Specifically, AgentSnare employs an artifact-construction policy model that constructs candidate artifacts conditioned on the agent's interaction history and decoy state. AgentSnare then validates these candidates and incrementally incorporates valid artifacts into a factually consistent decoy environment, thereby delaying the attack by absorbing its tool calls, diverting its post-entry trajectory within the decoy, and defusing it by inducing completion reports grounded in decoy evidence. Across 15 CVE-Bench web applications and three attacker models, AgentSnare absorbs 46.8% of the agent's tool calls in the decoy and retains 55.9% of post-entry actions there, while 90.0% of completion attempts are grounded in decoy evidence; across all 45 attacker-CVE pairs, no real target is successfully exploited at pass@3.

cs.CR↗
arXiv2026

Arm2Air: Cross-Embodiment Skeleton Transfer for 3D Relay Formation

Unmanned aerial vehicle (UAV) relay networks can restore connectivity after communication infrastructure is damaged. Urban relay placement is difficult because line-of-sight blockage, communication range, altitude, and three-dimensional obstacles must be considered jointly. Arm2Air transfers obstacle-avoidance skeletons from robot arms to UAV relay placement through cross-embodiment transfer. Source-domain robot-arm motions from a pretrained Neural MP model are converted into ordered skeletons that pretrain a transformer-based transfer platform, which is then adapted to the UAV domain using limited target data and Low-Rank Adaptation. The transferred skeleton initializes a relay chain that is refined for connectivity, bottleneck capacity, delay, and movement cost. On nine held-out high-clutter 3D urban maps, Arm2Air reduced median end-to-end planning runtime by 64.9 percent relative to the fastest conventional planner. On the high-obstruction group of a separate 30-map dense urban holdout, it increased bottleneck capacity by 32.6 percent, reduced capacity variance by 74.7 percent, reduced maximum hop distance by 13.2 percent, reduced hop-distance variance by 75.2 percent, and reduced relay displacement by 16.9 percent relative to IMPC-MD. With only three target-domain training maps, Arm2Air reduced relay-position root mean square error by 53.6 percent relative to training from scratch while updating 0.134 million parameters, compared with 1.383 million for Scratch and Full Fine-tuning. These results demonstrate computationally and data-efficient UAV relay placement and suggest a broader principle for transferring ordered structural priors across heterogeneous embodied tasks.

cs.RO↗
arXiv2026

Beyond Rephrasing: Book-Level Organization Improves Synthetic Textbook Data for Mid-Training

Synthetic textbook data has improved language model pre-training, but prior work largely treats the benefit as a property of generated content or local rewriting style. We study a different factor: whether related content is organized into coherent book-level documents. We contribute both a scalable synthesis pipeline and controlled evidence that this organization matters. The pipeline retrieves source material from a pre-training corpus, clusters it into topical units, plans hierarchical tables of contents, and assembles source-grounded sections into complete books (our Full setting), yielding 686K textbooks (32B tokens) across 15,000+ disciplines. Replacing natural books in a mid-training mix with this corpus improves downstream performance by +1.09 on average. Controlled comparisons then disentangle the relevant design factors. A content-matched Split condition holds generated text and tokens fixed but treats each section as an independent document; Full's +1.02 mean gain isolates document packaging. A length-matched RandomConcat control that joins sections from different books remains below Full, ruling out document length alone. A retrieval-pool-matched Rephrase condition independently rewrites individual retrieved documents under the same audience-by-style scheme, without clustering, TOC planning, or book assembly; Full's +1.17 gain demonstrates the value of structured synthesis. On Llama3-8B, Full likewise outperforms both RandomConcat and Natural Books, supporting book-level organization as a useful axis for synthetic pre-training data design.

cs.AI↗
arXiv2026

On credit attribution and research software: A case study from lattice QCD

Questions of authorship, credit attribution, and the recognition of research software contributions have become increasingly prominent in continually expanding research areas in which software constitutes an essential part of the research process. While general guidelines and best practices exist, their application in concrete situations often raises nontrivial interpretative and procedural issues. This article presents a documented case study from lattice QCD research in which the author was directly involved, illustrating how the development of numerical strategies, long-term software infrastructure, and conceptual extensions of existing work can give rise to complex questions of authorship, priority, and credit attribution. Rather than discussing the underlying scientific results, the focus is on the sequence of events, the role of software as long-term research infrastructure, and the interaction with established mechanisms for credit attribution and research integrity assessment. The aim of this work is to contribute to transparency and discussion on how current practices and guidelines are applied in realistic collaborative environments, and to highlight structural tensions that may arise between open scientific collaboration, software sustainability, and traditional notions of authorship. The article is intended as a factual account and reflection on research practice, and aims to stimulate discussion on whether additional community standards for recognising long-term research software contributions may be beneficial within lattice QCD.

hep-lat↗
arXiv2026

Radon Measure Representations for Infinite-Width Neural Networks with Singular Activations

The theoretical foundation of infinite-width shallow neural networks relies heavily on continuous integral representations and Barron spaces. Recently, harmonic analysis-specifically the Radon and Ridgelet transforms-has emerged as a powerful tool to invert these representations and compute the optimal network weights. However, a major analytical bottleneck remains: standard neural network activation functions exhibit severe spectral singularities at the frequency origin. To bypass this divergence, existing frameworks either restrict the theory to specific activation families or mathematically quotient out the network's affine components, which inherently limits their practical scope. In this paper, we overcome these limitations by introducing a purely distributional framework for the generalized Radon transform R $σ$ that operates on a broad class of tempered distribution activation functions. By defining a regularized spectrum formulation g($ρ$) = (i$ρ$) $α$ $σ$($ρ$), we rigorously absorb the origin singularities without truncating the underlying functional space. Building upon this exact reconstruction, we show that under a definite parity assumption on the activation, there exists an exact linear isometry between Barron functions and their optimal weight measures.

math.FA↗
arXiv2026

HP-JEPA: Hierarchical Partitioning for Multi-Resolution Graph Joint-Embedding Predictive Learning

Graph self-supervised learning aims to learn transferable representations from large-scale unlabeled graph data. Joint-embedding predictive architectures (JEPAs) avoid explicit negative-pair construction and raw-input reconstruction by predicting masked targets directly in latent space. However, existing graph JEPAs typically rely on a single predefined graph partition, biasing the learned representations toward one structural granularity and limiting their ability to capture complementary patterns at different graph scales. To address this limitation, we propose HP-JEPA, a hierarchical partitioning framework for multi-resolution graph joint-embedding prediction. HP-JEPA organizes each graph into an ordered bank of coarse-to-fine partition resolutions and performs context-target latent prediction separately at each resolution using an online encoder, an exponential-moving-average target encoder, and a latent predictor. The resulting resolution-specific graph representations are subsequently integrated through concatenation or task-specific resolution weighting, allowing downstream models to combine complementary local, regional, and global structural information. Experiments on seven graph classification benchmarks and one graph regression benchmark show that HP-JEPA outperforms the fixed-resolution Graph-JEPA baseline on 6 of 8 tasks, improving upon Graph-JEPA on most evaluated benchmarks. Size-stratified analyses further show that HP-JEPA achieves higher accuracy than Graph-JEPA in most evaluated graph-size quartiles on three representative datasets. These results highlight the effectiveness of hierarchical multi-resolution partitioning for transferable graph representation learning.

cs.LG↗
arXiv2026

AGNI: A differentiable MHD stability solver & optimizer for magnetic confinement fusion devices

The existence of an ideal MagnetoHydroDynamic (MHD) equilibrium does not guarantee its stability. Finite toroidal mode number (n) instabilities degrade performance in both tokamaks and stellarators and differentiable stability optimization tools to date have operated only in the infinite-n limit. We present AGNI (Analysis of Global Normal modes in Ideal MHD), a GPU-accelerated, automatically differentiable finite-n ideal MHD stability solver and optimizer. AGNI discretizes the ideal MHD energy principle pseudospectrally in real space using differentiation matrices and geometric coefficients from a DESC equilibrium, giving a variational eigenvalue problem for the plasma displacement, and efficiently finds the most unstable modes. Built on jax, AGNI yields reverse-mode gradients of the growth rate with respect to boundary-shape and profile parameters without re-solving the equilibrium. We benchmark AGNI against the initial-value code NIMSTELL for a modified Landreman-Bulle-Drevlak quasi-helically symmetric stellarator and a DSHAPE tokamak,recovering the dominant mode with agreement in terms of growth rate and eigenfunction structure, and verify the automatic differentiation gradients against central finite differences. We quantify CPU and GPU cost for eigenvalue and gradient evaluation, establish the finite-precision limit on resolving near-marginal eigenvalues, and present a numerical scheme to impose incompressibility, compatible with gradient-based optimization. AGNI will allow us to optimize tokamaks, stellarators, and mirrors against ideal MHD instabilities.

physics.plasm-ph↗
arXiv2026

One QK Channel, Many Sources: Tracing Low-Precision Attention Collapse

A bfloat16 transformer can train normally, then collapse abruptly. Prior work links collapse to structured attention errors and shows QK normalization disrupts their compounding. Distinct low-precision errors trigger the same collapse, leaving unclear whether each needs a fix at its source or one shared route can be blocked instead. We isolated the fault behind a reproduced GPT-2-class collapse to the streaming-softmax accumulator, where an fp32 streaming core repairs it, and turned it into an assay for moving a controlled error across sources. Using it, we found that errors placed outside attention still drove the same QK spectral runaway, and that correcting only QK kept training stable while the fault stayed active. This is a source-channel dissociation: fault source is not failure channel. It held across tested architectures and scales, and reproduced on a second GPU architecture. As a causal probe, projecting each update off the current QK weights' leading three singular directions held the query projection's largest singular value to 11.1, whereas removing equal energy elsewhere left it at 237: the QK channel causally drives the early runaway. What lets the injected error in is temporal sign-coherence, its per-head sign persisting across steps, not aggregate deviation; once inside, the runaway shows as attention-logit saturation. QK-Guard, a dormant controller, tests this by switching on parameter-free QK normalization at the first monitored threshold crossing. On the runs designated for this test, the QK-local action prevented the failure of each matched or same-configuration unguarded run; on plain GPT-2, all 12 final train and validation losses were within 0.03 nat of same-configuration always-on QK-norm, and both methods ran 60k steps without collapse. Intervention at the QK locus therefore suffices in place of a fix at each source.

cs.LG↗
arXiv2026

RamanPFN: learning from Raman spectral structure with a tabular foundation model

Raman spectroscopy enables label-free molecular characterization across materials science, analytical chemistry, biomedicine, and industrial process monitoring. However, machine learning for high-dimensional spectroscopy remains constrained by limited labelled data and a mismatch between the physical organization of spectra and feature-agnostic models. Channel coverage alone does not ensure that related bands share a common inference context. Here we present RamanPFN, a general-purpose spectral foundation framework that enables unified in-context inference through physics-guided spectral learning. It captures full-spectrum compositional covariation via Global Compositional Unmixing (GCU), which decomposes distributed, multi-band mixture signatures into shared non-negative latent bases. Simultaneously, it resolves local vibrational structure through Local Vibrational Subspace Encoding (LVSE), which preserves fine-grained peak morphology, intensity fluctuations, and peak shifts within contiguous spectral neighborhoods. Extensive evaluation across 74 diverse public Raman datasets covered 129 regression targets and was further extended to 21 classification tasks. RamanPFN achieved state-of-the-art performance across all reported aggregate metrics against 28 independently reproduced methods spanning chemometrics, spectral neural networks, deep tabular learners and tabular foundation models. RamanPFN establishes a physics-guided paradigm for scientific spectroscopy, enabling data-efficient predictive learning across diverse chemical systems.

cs.LG↗
arXiv2026

Elliptic complements of cubic hypersurfaces

Let $D\subset\mathbb{P}^n$, $n\geqslant2$, be an arbitrary cubic hypersurface, and let $D_{\mathrm{red}}$ denote its reduced support. We prove that $\mathbb{P}^n\setminus D$ is holomorphically elliptic, and hence Oka, unless $D_{\mathrm{red}}$ is the union of three distinct hyperplanes containing a common codimension-two linear subspace. In the exceptional case, $\mathbb{P}^n\setminus D\cong(\mathbb{C}\setminus\{0,1\})\times\mathbb{C}^{n-1}$, so the complement is not Oka. As applications, we prove that, for every elliptic curve $E$, the space of degree-three holomorphic maps $E\to\mathbb{P}^1$, and the space of degree-three holomorphic self-maps of $\mathbb{P}^1$, are both holomorphically elliptic, and hence Oka. The second application is connected with the classification through an irreducible cubic hypersurface in $\mathbb{P}^4$.

math.AG↗
arXiv2026

AccelNet: Exact backward-compatible acceleration of polynomial angular descriptors through Cartesian moment factorization

We present AccelNet, an exact, backward-compatible method for accelerating existing trained aenet and n2p2 neural-network potentials without retraining. For angular terms with separable one-neighbor weights and a finite polynomial dependence on $\cos θ$, the method exploits their hidden finite-rank structure to replace explicit neighbor-pair loops by one-neighbor Cartesian moments. AccelNet reads models trained with either package and reproduces their descriptors, energies, and analytic forces to floating-point roundoff. We verified this equivalence for H$_2$O and TiO$_2$ models and tested the resulting potentials in LAMMPS molecular-dynamics simulations. The implementation, model-conversion tools, and LAMMPS interfaces are released as open-source software. AccelNet also enables GPU-accelerated molecular dynamics with existing aenet and n2p2 potentials. Moment evaluation achieves speedups of up to 10.9 on a CPU and 15.8 on a GPU relative to direct evaluation within AccelNet. The implementation, model-conversion tools, and LAMMPS interfaces are released as open-source software.

cond-mat.mtrl-sci↗