Statistical Mechanics
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
All of arXiv
Tuesday 1 September
Information engines convert information into extractable work through measurements and feedback. We realize an engine that employs information to compress a gas of self-propelled bristle bots. Its work per measurement, $\bar{W}$, is controlled by the size of the detection region $\Delta x$ and the probability $p_1$ that this region is vacant. In thermal systems, $\bar{W}$ follows the universal form $-p_1 \ln p_1$, whereas the active engine shows a qualitatively modified relation. As particle density increases, the maximum of $\bar{W}(\Delta x)$ switches between two distinct operating regimes. Notably, this transition is also found in the solutions that maximize power output for finite-time cycles with dynamics affected by dry friction. We attribute this nonequilibrium feature to the accumulation of active particles near the boundaries.
Computational Complexity
On Top-Down and Local Lower Bounds for $\mathrm{AC^0}$ Circuits
Classical lower bounds for $\mathrm{AC^0}$ circuits proceed bottom-up by simplifying or approximating gates beginning at the input layer. We introduce a complementary top-down model called the Chopping Game, played by adversaries Spoiler and Duplicator on the sets of $0$- and $1$-inputs of a Boolean function. In each round, Spoiler keeps at least a $1/m$-fraction of one side, and Duplicator arbitrarily restricts the other; Spoiler seeks to minimize (and Duplicator to maximize) the number of rounds until some coordinate separates the two remaining sets. Every depth-$d$, fan-in-$m$ circuit induces a $d$-round winning strategy for Spoiler, while Duplicator strategies that survive $d$ rounds formalize top-down lower-bound arguments. Through the Chopping Game and using the polynomial-approximation method, we first obtain the classical lower bound for depth-$d$ $\mathrm{AC^0}$ circuits in a top-down fashion. We then consider a $k$-local variant of the Chopping Game, which relaxes Spoiler's win condition by requiring a separating coordinate within each Hamming ball of radius $k$, rather than a single coordinate globally. We put forward a conjecture that the $d$-round $k$-local Chopping Game for $\mathrm{PARITY}$ requires $m = n^{\omega(1)}$ in the regime $d \ll k \ll n$. We prove such a lower bound $m \ge n^{\Omega(k^{1/d}/d)}$ when Spoiler is restricted to so-called affine strategies, a class of strategies that achieves the best known upper bounds. Finally, we formulate a version of the $k$-local Chopping Game on $n$-regular graphs of girth $>2k$, and we conjecture a graph-theoretic analogue of “$\mathrm{PARITY} \notin \mathrm{AC^0}$”.
Read the full abstract
Cryptography and Security
Towards Behavior Tree-Guided Vulnerability Detection with Lightweight LLMs
Large Language Models (LLMs) are increasingly used for software vulnerability detection, but their performance depends on how source code is represented in the input. Most prompting approaches use source code in its original form, while some works propose the use of structured representations. Abstract Syntax Trees (ASTs) are one of the most popular approaches, but AST verbosity increases input size relative to source code, making them hard to fit within some LLMs context windows. This paper investigates Behavior Trees (BTs) as an alternative intermediate representation for LLM-based vulnerability detection. BTs encode control flow, conditions, and executable actions more compactly than ASTs, making them a natural candidate when token count is a constraint. First, we propose a preprocessing stage that parses Java source code into ASTs and then converts them into BT representations. We then compare vulnerability detection performance across 460 Java samples from the Juliet Java test suite, using three input representations: raw source code, AST, and BT. All experiments use a single quantized local LLM, Mistral Small 3.2 24B (Q4_K_M). Our results show that using BT representations improves recall on short code samples, while raw source code achieves higher precision. On longer samples, BTs improve overall performance over the original representation and fit within the context window, whereas many ASTs exceed the context limit. These findings suggest that BTs can provide a compact and useful structured representation for vulnerability detection with quantized, locally deployable LLMs.
Read the full abstract
Computer Vision
AlphaRAD: Grounded Zero-Shot Classification in Chest Radiology via $\alpha$-Corrected Binary Cross Entropy and Factorized Latent Supervision
Vision-Language Pretrained Models (VLPMs) offer a scalable path to open-vocabulary chest radiology understanding, yet two aspects remain underexplored: how structured clinical semantics extracted from medical reports can reduce in-batch noise during contrastive learning, and how cross-modal fusion can be designed to produce more faithful spatial grounding without added complexity. We introduce AlphaRAD, addressing these opportunities through two contributions. First, we construct a large-scale structured medical concept space from medical reports parsed by a Large Language Model for training, thereby mitigating in-batch learning noise and removing heuristic pair matching in contrastive learning, and thus naturally positioning AlphaRAD as a medical concept discriminator trained via $\alpha$-Corrected Binary Cross-Entropy. Second, we propose FLaS (Factorized Latent Supervision), an extremely simple yet effective cross-modal feature fusion module that factorizes VLPM representations into independent subspaces, using dedicated alignment supervision to enhance the expressiveness of spatial grounding without introducing additional model parameters. Through extensive empirical validation, AlphaRAD shows strong zero-shot generalization across diverse chest radiology tasks. Notably, it establishes state-of-the-art average performance across 16 classification benchmarks, while achieving individual state-of-the-art results via distinct gains on 7 grounding/phrase grounding and 3 segmentation datasets.
Read the full abstract
Machine Learning
A Study of Conditional Diffusion Models for Open-Loop Control under Dry Friction and Stiction
Diffusion models have recently emerged as expressive generative priors for planning and control. This paper studies Action Diffusion, an action-sequence diffusion formulation used as an open-loop proposal distribution for a point-mass system with dry friction and stiction. In this benchmark, motion starts only when the applied input exceeds a static-friction threshold, so effective controls occupy a small and temporally structured subset of the action-sequence space. A compact conditional 1D U-Net generates bounded control sequences conditioned on initial and target states. We compare it with uniform random shooting, random shooting from the same structured dataset prior, and the Cross-Entropy Method (CEM). Results show that Action Diffusion reduces terminal error and stuck steps, especially in low-sample regimes. These results indicate that conditional diffusion provides an effective mechanism for generating temporally coherent control sequences that overcome stiction by conditioning and recombining structured control primitives from the training prior for state-to-state open-loop control.
Read the full abstract
Optics
Reciprocal asymmetric transmission in self-shadowed metallized gratings
A single oblique metallization step gives conventional gratings one-sided silver decoration, established by electron microscopy and elemental mapping. Opposite-face reversal measurements yield finite-aperture red/green/blue transmission ratios up to 1.74 for a 1-$\mu$m grating and 1.5-1.8 for a 30-$\mu$m grating, with reference-normalized transmittances for the latter. Flat-side coating and reduced silver dose give ratios near unity. Maxwell calculations on self-shadowed growth profiles reproduce the directional sign qualitatively and attribute the contrast to diffraction-channel redistribution. The structure remains reciprocal, including its normal-incidence zeroth-order transmission. These results establish directional metallization as a route to direction-selective optical transmission without multilayer alignment or overlay lithography.
Read the full abstract
Other Condensed Matter
Phase synchronization of fish schools through spatial gaps
We investigate how sensory coupling mediates phase coherence between spatially separated, confined active systems. Using two fish schools separated by a water-filled transparent gap, we isolate vision as the primary interaction channel. Under boundary confinement, each school's center-of-mass heading exhibits noisy bistable dynamics, spontaneously switching between two antiparallel directions along the partition wall. Turning events in one school bias the orientation of the other, driving distance-dependent phase synchronization. We model this behavior using two coupled stochastic bistable oscillators whose exact solution quantitatively reproduces the empirical results. Our work establishes a precise connection between nonequilibrium active systems and coupled stochastic oscillators, offering insight into boundary-spanning information transfer in living collectives.
Read the full abstract
Analysis of PDEs
The entropy production is not always monotone for hard spheres or for Maxwell molecules
In 1966, McKean asked whether the entropy production of the Boltzmann equation must be monotone decreasing in time. We show that this is not the case even in the space-homogeneous setting for the Boltzmann collision operator with a constant angular cross section and the kinetic parameter $\gamma \in [0,1]$. This recovers the classical case of hard spheres and the simplest case of Maxwell molecules. Our examples are radially symmetric mixtures of Maxwellians.
Read the full abstract
Quantum Physics
$\mathcal{N}$-bein formalism for degenerate states in the parameter space of quantum geometry
Recently, we introduced a geometric object analogous to an orthonormal frame in the Cartan formalism to study the parameter space of quantum systems; we called it $\mathcal{N}$-bein, with $\mathcal{N}$ being the number of parameters that characterize the quantum system. Acting as the “square root” of the quantum geometric tensor (QGT), the $\mathcal{N}$-bein allows us to define new tensors to improve our understanding of the structure beneath the parameter space of quantum mechanics. In this work, we extend this mathematical framework surrounding the $\mathcal{N}$-bein to analyze the parameter space of quantum systems with degenerate spectra. As in the non-degenerate case, we define a non-Abelian two-state QGT to identify possible transitions between degenerate states after two consecutive parameter variations. Additionally, using the Wilczek-Zee connection, we introduce a torsion-like tensor as the covariant derivative of the $\mathcal{N}$-bein. This torsion captures the noncommutativity of successive parameter variations and coincides with the antisymmetric part of the two-state QGT. We also present a geometrical formulation using differential forms and discuss the physical implications of the newly defined tensors. Furthermore, we construct several gauge-invariant observables from the $\mathcal{N}$-bein and its derivatives to highlight the utility of the new tensors. Finally, to illustrate the convenience and applications of this formalism, we apply the theoretical framework to a system of coupled harmonic oscillators immersed in an electric field. The coupling between the oscillators results in a degenerate system. Thus, using the new formalism, we found correlations among the quantum states quantified by the new invariants.
Read the full abstract
Classical Analysis and ODEs
Orthogonal polynomials which are eigenfunctions of a partial differential operator
We study orthogonal polynomials of $d = d_1+d_2$ variables with respect to a wrapped product weight function $\Wb(\bm x,\bm y) = W_1(\bm x/\rho(\bm y)) W_2(\bm y)$ for $(\bm x, \bm y) \in \mathbb{R}^{d_1} \times \mathbb{R}^{d_2}$, where $\rho$ is either linear or the square root of a nonnegative quadratic polynomial, and identify all such polynomials that are eigenfunctions of a second-order linear differential operator. For $d =2$, it is known that there are primarily, up to affine transformations, five families of such polynomials, which are products or wrapped products of classical orthogonal polynomials of one variable; all five families have their counterparts in higher dimensions, but no characterization is known in dimension three or higher. Our study explores viable wrapped product families, finds explicit second-order differential operators for two new families of orthogonal polynomials in $d= d_1+d_2$ variables that have not been studied before when either $d_1>1$ or $d_2 > 1$, and provides, in particular, a complete list of such operators among all wrapped product orthogonal polynomials when $d = 3$. The list also includes four families that are eigenfunctions of a fourth-order differential operator, whereas no second-order operator is available. Moreover, orthogonal polynomials on the corresponding wrapped quadratic surfaces that are eigenfunctions of a second-order differential operator on the surface are also studied.
Read the full abstract
Plasma Physics
The Fusion Equilibrium Challenge: Inferring Magnetic Geometry Without Magnetic Diagnostics
Next-generation fusion reactor devices such as SPARC, ARC, and CFETR will operate in extreme neutron environments that compromise the magnetic sensors traditionally used to reconstruct plasma equilibria. However, reliable knowledge of the plasma equilibrium–including magnetic flux surfaces, safety factor profiles, and shaping parameters–is indispensable for real-time control, disruption avoidance, and physics interpretation. The Fusion Equilibrium Challenge invites the NeurIPS community to confront a deceptively simple but scientifically rigorous inverse problem: reconstruct the two-dimensional poloidal flux function psi(R,Z) and a suite of scalar equilibrium parameters from non-magnetic diagnostics alone, namely external poloidal-field coil currents and Thomson-scattering electron temperature/density profiles. The challenge provides the first open-access, harmonized multi-machine benchmark for fusion, releasing a curated dataset of 9,113 DIII-D shots and 2,416 MAST shots–filtered for Thomson-diagnostic availability, feature completeness, and EFIT-reconstruction quality. Each shot is packaged into a standard Parquet file containing approximately 260 (DIII-D) / approximately 80 (MAST) EFIT flux maps and rich high-rate diagnostics. Two complementary awards reward intra-machine reconstruction fidelity (S_model) on DIII-D and zero-shot cross-machine generalization (G_ratio) to the topologically distinct MAST spherical tokamak. We argue that the challenge functions as a benchmark for reactor-ready equilibrium inference and as a probe of how far machine learning can be pushed toward truly machine-agnostic plasma state estimation.
Read the full abstract
Computer Vision
Swin Meets EfficientNet: Lightweight Architectures for GAN-Based Face Forensics
Modern generative models, such as GANs, diffusion architectures, and autoregressive systems, now produce facial images that are nearly indistinguishable from authentic photographs. This capability makes detecting forged images increasingly difficult, raising serious concerns about identity theft, fraud, and misinformation campaigns. Our research focuses specifically on GAN-generated synthetic faces, which underpin many face-centric deepfakes, and investigates efficient detection approaches using image analysis alone. Existing detection systems rely heavily on either convolutional neural networks (CNNs) or global vision transformers. While CNNs excel at identifying texture-based local features, they struggle with broader contextual understanding. Traditional Vision Transformer (ViT) models can capture long-range structures effectively, but demand substantial computational resources. Our work explores Swin-Transformer-based architectures across three implementations: a compact Swin Transformer trained from the ground up, ImageNet-1K pre-trained Swin-Tiny and Swin-Small models adapted for binary classification, and a novel hybrid combining EfficientNet-B0's convolutional processing with a Swin Transformer backend. We evaluated all models using the 140K Real and Fake Faces dataset, which includes StyleGAN-generated fake faces alongside authentic images from Flickr and DFDC, with balanced splits for training, validation, and testing. The EfficientNetB0+Swin hybrid achieved 99% accuracy and a 99.44% recall on 5,000 test images, outperforming both pure Swin variants and a previous CNN-only baseline on this dataset. Our results suggest that combining hierarchical CNN features with shifted-window self-attention provides an efficient and computationally lightweight method for detecting GAN-generated synthetic faces.
Read the full abstract
Numerical Analysis
A Double-Adaptivity Solver for Parabolic PDEs
We study minimal residual space-time finite element discretizations of linear parabolic initial value problems in canonical space-time variational form. To deal with the arising dual norm, we introduce the Riesz lift of the residual as an additional variable. Quasi-optimality of the primal variable of the mixed system follows from a uniform inf-sup condition. This condition is known to be satisfied for finite element spaces w.r.t. prismatic partitions of the space-time cylinder that allow for a decomposition into time-slabs. We prove that this condition cannot be expected to hold otherwise. To recover stability for general partitions and the data at hand, we derive an a posteriori condition on the error between the exact Riesz lift of the residual and its Galerkin approximation – being the secondary variable of our system – under which the primal variable is quasi-optimal. We derive a posteriori error estimators for both variables, and use them in a double-adaptive loop that alternates test-space with trial-space enrichment. We illustrate our findings with numerical experiments in $1+1$ and $2+1$ dimensions.
Read the full abstract
Number Theory
A lower bound for the Call-Silverman height in cyclotomic extensions
For quadratic postcritically finite polynomials defined over a number field $K$, we establish lower bounds for the Call-Silverman height of wandering algebraic integers lying in cyclotomic extensions of $K$ and provide partial results for the nonintegral case. Our proof combines potential theory and algebraic number theory with equidistribution techniques.
Read the full abstract
Machine Learning
CAT-Flow: Curvature-Adaptive sTeps for Flow Matching
Flow Matching has emerged as a leading framework for generative modeling, powering state-of-the-art systems such as FLUX and Stable Diffusion 3.5. However, the iterative nature of its ODE-based sampling process creates a fundamental efficiency bottleneck: the quality of generated samples is highly sensitive to the choice of step-sizes, and current models typically require 20 to 30 steps for good quality. In this work, we propose two lightweight, training-free algorithms, CAT-OV and CAT-OT that adapt step-sizes at inference time based on a novel connection between Flow Matching sampling and gradient flow. Our algorithms are computed efficiently by not requiring additional neural function evaluations. Specifically, CAT-OT estimates curvature over time via a finite-difference approximation of the time-derivative of the vector field, while CAT-OV approximates curvature over the state space via a gradient of the vector field. Under suitable conditions, both methods have truncation error bounds of constant order. Empirically, CAT-OV and CAT-OT outperform existing step-size heuristics in image quality metrics across four text- to-image Flow Matching models, reducing the number of generation steps required to reach comparable quality by up to 40%.
Read the full abstract
Combinatorics
The directional localization game on graphs
In the localization game on a graph $G$, a team of cops searches for an invisible, mobile robber on $G$ by "probing" vertices; each probe tells the cops the distance from the probed vertex to the robber. The cops win if they can uniquely determine the robber's location. In this paper, we introduce a related game: the directional localization game. In this game, instead of probes returning distances, they return directions: when the cops probe a vertex $v$, the robber must respond with one or more neighbors of $v$ that lie on a shortest path from $v$ to the robber's location. The minimum number of cops needed to win this game on $G$ is the directional localization number of $G$. We study the directional localization game on several classes of graphs, including chordal graphs, Cartesian products, and incidence graphs of projective planes. We also bound the directional localization number of a graph $G$ in terms of the degeneracy and the treewidth of $G$.
Read the full abstract
Quantum Physics
Reliable Sample-Level Quantum Error Mitigation via Dominance-Aware Clustering
Many quantum algorithms for classically difficult optimization tasks must return high-quality bitstrings from finitely many circuit executions, whereas most quantum error-mitigation methods target expectation values. We study sample-level recovery when measured probability mass is distributed around multiple latent bitstrings, called centers. Each component of the measured probability mass is called a source and we assume that each center is associated with one source. We identify dominance-at every coordinate, more than half of a retained region's probability mass comes from one source and agrees with its center-as a sufficient condition under which majority voting recovers that center with exponentially decreasing error probability. We show that nearest-center assignment, as used in clustering algorithms such as the $k$-modes algorithm, can fail to produce dominated regions even when the true centers are known. This failure motivates responsibility thresholding and a local dominance screen, whose combination we call dominance-aware (DA) refinement. Synthetic and simulated MaxCut-QAOA experiments show that DA refinement favors precision, while $k$-modes with DA refinement improves overall center recovery. All procedures are classical post-processing and require no additional quantum-circuit executions.
Read the full abstract
Computer Vision
SCULPT: Training Edge Vision Models for Post-Training Quantization Readiness
Edge vision models are difficult to deploy on resource-constrained hardware, making low-bit post-training quantization (PTQ) attractive. In practice, standard FP32 training often produces heavy-tailed activation distributions whose outliers destabilize activation quantization: preserving the full range wastes quantization bins on rare extremes, while aggressive clipping causes information loss. Existing solutions typically rely on quantization-aware training (QAT), which adds training complexity and bit-width coupling, or advanced PTQ procedures that repair the model after training. We present SCULPT (Statistical Clipping and Uniform Loss for Post-Training), a training-time method that improves PTQ readiness during ordinary FP32 fine-tuning. SCULPT combines a topology-aware activation regularizer that suppresses quantization-hostile skewness and kurtosis with a stable percentile-based clipping mechanism that learns deployment-ready activation bounds. Unlike QAT, SCULPT does not simulate quantization during optimization; unlike post hoc outlier-repair PTQ methods, it does not require runtime activation transformations. The learned clipping bounds can be exported directly into a standard PTQ workflow for low-bit deployment, including INT8 and lower-bit settings such as W4A8.
Read the full abstract
Computer Vision
Evidential Deep Learning for Multi-Modal Anti-UAV Detection
Anti-UAV systems increasingly fuse multiple sensors, yet their detection heads provide no per-modality reliability signal. This study evaluates whether evidential deep learning (EDL) heads, Dempster-Shafer (DS) evidence fusion, and uncertainty-driven temporal sensor gating improve anti-UAV detection through a controlled ablation on three benchmarks: thermal tracking (AntiUAV600), RGB-audio-RF classification (TRIDENT), and RGB-IR tracking (MM-UAV). The EDL training objective improves accuracy over retrained sigmoid baselines (+5.9 percentage points in accuracy and a tripled tracker-on-absent rate in E1; +4.8 percentage points in classification accuracy in E2, surviving a clip-clustered bootstrap, p = 0.011) and ranks classification errors substantially better (entropy UAUC approximately 0.94 vs. 0.51). The remaining components do not support their respective hypotheses. DS fusion does not outperform simple probability averaging. Dirichlet vacuity adds no ranking power beyond predictive entropy and inverts at the detection level, where extreme background imbalance causes it to encode class membership rather than error likelihood, a failure also observed for entropy and sigmoid confidence. Temporal gating preserves accuracy only when nearly inactive and yields no realised latency saving on shared-backbone hardware. The benefit of evidential learning therefore arises primarily from its training objective rather than its uncertainty estimate; a crop-level control further localises the detection-level breakdown to anchor-level evaluation rather than the learned representation.
Read the full abstract
Artificial Intelligence
When Can a Machine Trust a Statute? A Survival Certificate for Machine-Extracted Legal Logic
Statutes are increasingly parsed by machines before people read them, and the parsers disagree: on Missouri's statutes, two independently written extractors diverge on numeric-threshold presence at a false-negative rate of 0.43. We ask what formal logic survives such noise. We build a passive survival certificate for the Duquenne-Guigues implication basis of machine-extracted statutory contexts: per-attribute inter-extractor disagreement is measured, replayed against the basis in 1,000 Monte Carlo trials, and an implication is certified only when a one-sided Wilson 95% lower bound on survival reaches 0.95; every certified implication carries premise spans and a minimal counterexample. On 29,365 Missouri sections and 502 Indian central-Act sections, the preregistered held-out gate passes (10 statute families across 7 Titles exact; 16 across 11 with 5% tolerance), yet under one globally deployed error model 93.2% of held-out chapters fall below the informativeness floor, and a 2x2 factorial assigns that to calibration-rate transfer, not selection. The certificate is usable but fragile: deploy it per-chapter-calibrated or error-tolerant. Code, data products, and the audit trail, including one retracted claim, are released.
Read the full abstract
Computer Vision
ZipTok3D: High-Fidelity 3D Tokenization with Compact Token Prefixes
Compact token sequences are essential for efficient 3D generation. However, existing 3D tokenizers typically organize latent representations either over spatial regions or as fixed-size sets of global tokens, both suffering sharp reconstruction degradation when compressed to extremely low token budgets. In this paper, we present ZipTok3D, a 3D tokenizer designed for high-fidelity reconstruction from extremely short token sequences. Its key idea is to organize object geometry into progressively informative global-token prefixes and unfold these compact representations through iterative decoding. Specifically, nested dropout randomly truncates the latent sequence after encoding during training and requires each retained prefix to reconstruct the complete object, thereby prioritizing essential geometric information in the leading tokens. The decoder then repeatedly applies a parameter-shared Transformer block to recover fine-grained geometry from each prefix without a separate generative sampling stage. With the same token dimension, ZipTok3D achieves reconstruction quality comparable to the 32-token COD-VAE baseline using only one token on ShapeNet and four on TRELLIS, yielding $32\times$ and $8\times$ shorter token sequences, respectively.
Read the full abstract
Astrophysics of Galaxies
Virgo Filaments VIII: Characterizing the Structural Parameters of Virgo Galaxies with Machine Learning to Probe Environmental Quenching
A persistent challenge in galaxy evolution involves disentangling the many correlated properties in order to isolate the effects of a galaxy's environment on its star formation history. To address this multidimensionality problem, we apply k-means clustering to 2831 galaxies in the extended regions around the Virgo cluster to define objective, reproducible subsets of structurally similar galaxies. Using measurements of size, light distribution, and stellar mass, k-means partitions these galaxies into three feature classes (FCs): dwarfs, spheroids, and large disks. In addition to being structurally different, these FCs show distinct offsets from the star-forming main sequence to $>3\sigma$ significance, with the spheroid population systematically shifted to lower star formation rates. Examining environmental dependence within each FC, we find that denser environments are associated with progressively stronger quenching. However, star formation for the dwarf and large disk galaxies is not strongly affected until the rich group and cluster environments. For the spheroid galaxies, star formation instead smoothly decreases as environment density increases. We verify that these trends are not driven by differences in Sersic index within each environment, suggesting that the effectiveness of environmental quenching depends on the galaxy's structural class. Our results speak to the utility of a simple machine learning model to create broad classes of structurally similar galaxies based on a small set of parameters, which has important implications for navigating this data rich era of astronomy.
Read the full abstract
Computer Vision
UAV Thermal Imagery for Inert Ordnance Screening: Multi Campaign Dataset Development,Object Detection, and Practical Recommendations
Unexploded ordnance (UXO) continues to restrict civilian access, agricultural activity, infrastructure recovery, and environmental remediation in contaminated areas around the world. This study created a multi campaign UAV thermal image data set of inert ordnance, developed a labeled image set from collected imagery, tested object detection models, and identified practical considerations for humanitarian mine action and demining applications. Data were collected during four field campaigns in Tennessee under summer and winter conditions using inert mines, munitions, and other ordnance placed in short grass, tall vegetation, gravel, mulch, rock, compost, and compacted surfaces. Thermal imagery was collected under flight altitutes of 33 m and 15 m. The final source inventory contained 5,855 thermal image label pairs, including 918 positive images and 4,937 background images. After retaining all positive images and downsampling background images, the 33 m dataset contained 420 training and 106 validation images, while the 15 m dataset contained 629 training and 157 validation images. YOLOV11l and RT-DETR-R50 algorithms were trained and evaluated to develop an automated candidate detection model. Practical recommendations include collecting thermal and RGB imagery together, incorporating varied surfaces and background only imagery, considering periods following changes in solar exposure, balancing survey coverage against target pixel representation, calibrating models with representative local data, and retaining qualified human review. The intended use is screening and prioritization for follow on technical survey or EOD assessment, and not a standalone clearance.
Read the full abstract
Computation and Language
SpeakPay: Domain-Adaptive LoRA Fine-Tuning of Whisper for Low-Resource Nepali Financial Speech Recognition
Mobile payment applications in Nepal are graphically mediated and largely inaccessible to visually impaired users. This paper presents SpeakPay, a voice-first digital wallet, and documents the central technical contribution: a controlled study of domain adaptation for low-resource financial speech recognition. We introduce NepFinSpeech-403, a 403-utterance dataset of Nepali financial voice commands (send, load, and balance operations spanning 237 unique numerals), and fine-tune Whisper large-v2 with LoRA. On the held-out test set, the domain-adapted model reduces Word Error Rate from 129.95% (zero-shot baseline) to 42.58% — a 67.2% relative reduction — and improves Devanagari numeral recognition accuracy from 0.0% to 73.9%. We find that word-level metrics understate the practical task-level impact: domain adaptation improves the Transaction Success Rate from 1.67% to 33.33%, a roughly 20x gain. The improvement is consistent at the individual-utterance level (sign test, $p < 10^{-17}$) and across all command types. A data efficiency analysis shows that as few as 100 domain-specific utterances are sufficient to halve the zero-shot WER, with performance plateauing around 300 examples. Error analysis reveals systematic numeral confusion patterns (zero insertion/deletion, prefix hallucination) that account for the majority of remaining transaction failures. The trained system is deployed as a publicly accessible voice-first web application. All code, dataset, model weights, and this paper are released at https://github.com/subedibiraj/speakpay.
Read the full abstract
Software Engineering
Harness Engineering in LLM Tool Use via Agent-Native Reusable Tool Primitives
Large language models (LLMs) augmented with external tools have demonstrated remarkable capability in solving complex real-world tasks. However, existing approaches suffer from two key challenges: brittle multi-step and multi-turn reasoning caused by incompatible tool output types and API schemas, and performance degradation under large tool catalogues. To address these, we introduce Tool Primitives, a design that replaces rigid API schema-based invocation with natural language as the interface for tool calling, where each tool is wrapped with an LLM interface that handles schema resolution and execution internally, enabling natural inter-tool communication for nested and multi-turn tool calling. Building on Tool Primitives, we host ToolFace, a centralized repository of 25,519 functions from which LLMs dynamically retrieve only the relevant tools at inference time, eliminating the need to enumerate raw API schemas in context. To orchestrate Tool Primitives and ToolFace reliably in complex settings, we further propose HEART, a Harness Engineering framework via Agent-native, Reusable Tool Primitives, comprising a Planner, Router, and Verifier that jointly support dynamic tool invocation planning, multi-step execution, and feedback-driven recovery. Experiments on five benchmarks demonstrate that HEART outperforms SFT-based models by $10\%$ on average and surpasses GPT-5.4, Claude-4.6-Sonnet, and Gemini-3.1-Pro by $6\%$ on average while reducing API cost by up to $85\%$. On 50 real-world tasks, HEART achieves $84\%$ task completion, $3.8\times$ the average of three frontier commercial models ($22\%$).
Read the full abstract
Neural and Evolutionary Computing
CircuitsDNA: Discovering Unconventional Multi-Accuracy Arithmetic Circuits via Evolutionary Synthesis
Emerging edge AI workloads increasingly require arithmetic units that can trade computational accuracy for efficiency on demand. However, existing approximate arithmetic circuits are typically fixed-accuracy or rely on predefined structures for runtime configurability. This work introduces CircuitsDNA, an evolutionary framework that automatically evolves accuracy-configurable arithmetic circuits supporting multiple accuracy modes within a single circuit. It integrates three key features: 1) multi-threshold verifiability miter to enforce mode-specific accuracy requirements, 2) resource-limited verifiability-driven search to reduce verification overhead without sacrificing correctness, enabling efficient exploration of large circuit design, and 3) feedback-driven adaptive mutation to prioritize effective structural modifications and accelerate search convergence. Experimental results show that the 8-bit multiplier variants synthesized in 28-nm CMOS reduce the area-power product by up to 56% on INT8 DNN workload and 93% under exhaustive activity, compared with an exact 8-bit multiplier. Across CNNs and DeiTs, the accuracy loss relative to FP32 remains below 2% after fine-tuning under worst-case error (WCE) budgets of at most 1%. CircuitsDNA eliminates all search stalls observed in conventional methods across 8/12/16-bit multipliers, while adaptive mutation provides up to 1.33 times faster convergence than its non-adaptive counterpart.
Read the full abstract
Instrumentation and Methods for Astrophysics
Learning and Predicting the Nonlinear Variability of X-ray Binaries with the Koopman Operator
X-ray variability in compact-object binaries encodes the nonlinear dynamics of corona-jet interactions and accretion disk instabilities. Standard timing techniques characterize periodic and quasi-periodic variability well, but do not model underlying nonlinear dynamics or forecast their evolution. We apply Koopman operator theory and a data-driven approximation, extended dynamic mode decomposition (EDMD), to X-ray light curves for the first time. Koopman theory represents nonlinear evolution as an infinite-dimensional linear operator $\mathcal{K}$, whose eigendecomposition separates a complex system into independently evolving linear modes. We derive that each Koopman eigenfunction contributes a Lorentzian peak to the power spectrum, giving quasi-periodic oscillations a dynamical interpretation in which process noise damps modes and broadens their peaks. In both chaotic Duffing oscillator simulations and $\sim$30 yr of RXTE ASM and MAXI monitoring of the X-ray binary 4U 1705-44, the slowest-varying eigenfunction partitions state space into low- and high-flux regimes, changing sign days to weeks before transitions become visible in the light curve. Iterating $\mathcal{K}$ additionally yields flux forecasts on days-to-weeks horizons. The strengths of this framework are its generality across linear and nonlinear systems, its intrinsic interpretability through decomposed modes carrying explicit dynamical meaning, and its predictive power from propagating learned dynamics forward. These results establish Koopman operator theory as a new frontier of astrophysical timing and help advance interpretable machine learning for scientific discovery and understanding.
Read the full abstract
General Relativity and Quantum Cosmology
Exploring the construction of field equations for regular black holes with dynamical regularization scales
The study of the dynamics of regular black holes is an emergent area of research. Recent works have discussed how to deform the Einstein field equations to regularize spherically symmetric black hole solutions due to the introduction of a new physics scale acting as regulator. In this exploratory paper, we consider for the first time the problem of constructing field equations in which the regularization scale is a dynamical field. Such field equations may provide an effective description of the evolution of the cores of regular black holes due to the backreaction of matter, and a novel avenue for constructing alternative dynamical solutions of theoretical and phenomenological interest. We discuss a general family of field equations describing spherically symmetric gravity coupled to a scalar field, and study whether the sector with a constant scalar field contains solutions describing static regular black holes with no additional physics scales aside from the ones provided by the scalar field. We show that this question can be reduced to solving a linear partial differential equation for the coupling functions between the scalar and the metric degrees of freedom.
Read the full abstract
High Energy Physics (Theory)
Truncated Polyakov bootstrap
We set up a truncated numerical approach in the Polyakov bootstrap (PB) framework. We employ a gradient descent optimization to solve PB sum rules numerically, and hence solve crossing for arbitrary deformations of the generalized free field (GFF) spectrum in an iterative way starting from a perturbative solution. For unitary deformations the solutions coincide with extremal CFT spectra. But more interestingly, our findings indicate that a general crossing solution, not necessarily unitary (i.e. without positivity), can be uniquely identified by a smooth connection to GFF. The truncated Polyakov bootstrap approach is established through a number of examples in both single and mixed correlator settings.
Read the full abstract
Robotics
TriSAR: Task Coordination and Collision Avoidance for Aerial Robot Teams in Disaster Response
Multi-Unmanned Aerial Vehicle (UAV) disaster-response systems require coordinated task assignment and local trajectory control, yet the individual and combined contributions of these coordination layers to mission efficiency and operational safety remain insufficiently characterised under controlled experimental conditions. TriSAR is evaluated as a five-UAV coordination system operating in a physics-based Gazebo simulation of an earthquake-damaged urban environment. A 2 x 2 factorial design compares two task-allocation strategies (Genetic Algorithm and greedy fitness-based allocation) with reactive collision avoidance enabled or disabled. Each of the four configurations was evaluated over 30 stochastic episodes in a common scenario of five UAVs and eight targets. Under greedy allocation, enabling repulsion eliminated recorded collision-threshold violations, confirmed by a Mann-Whitney test (U = 885, p = 4.03 x 10^-12, rank-biserial r = 0.97). Under GA allocation, the same protective effect was confirmed (U = 675, p = 1.26 x 10^-5, rank-biserial r = 0.50). For mission-efficiency metrics, GA-based allocation showed no statistically detectable advantage over greedy allocation when repulsion was enabled, but a significant advantage in steps, path length, and energy when repulsion was disabled (Welch's t-tests, |g| between 0.92 and 1.76). These results show that reactive repulsion provides a substantial, allocation-dependent safety benefit, while the additional computational complexity of GA-based task allocation yields a detectable mission-efficiency benefit only when repulsion is disabled.