Peer Revenue Now accepting positions

Papers

Fresh from arXiv, typeset properly. Highlighted words trade on the market.

All of arXiv

Monday 9 November

Robotics

Hierarchical Multi-Channel Tip-Over Prediction via Isolated Temporal Models for Planetary Rover Safety

Reliable tip-over prevention is critical for planetary rovers operating in unstructured environments where communication delays prevent real-time human intervention. Existing learning-based approaches typically optimize a single prediction objective, limiting their ability to capture both gradual instability evolution and abrupt terrain disturbances. This work presents a hierarchical multi-channel tip-over prediction framework based on isolated temporal learning. A multi-horizon Temporal Convolutional Network (TCN) learns progressive instability patterns from inertial measurements, while a Gated Recurrent Unit (GRU)-based forward dynamics model predicts nominal rover motion and detects unexpected dynamic deviations through prediction residuals. These complementary signals are integrated through decision-level fusion and priority-based arbitration to generate graded safety actions, including early warning and emergency braking. The proposed system is deployed as an Orocos Real-Time Toolkit (RTT)-based safety component within the Rock robotics framework, operating at 100 Hz on a planetary rover software stack. Experiments on Inertial Measurement Unit (IMU) samples collected across heterogeneous terrains demonstrate 97.58% event prediction coverage with a 1.82s median lead time, outperforming individual prediction channels. The results validate that isolated temporal models combined through hierarchical decision fusion provide a robust solution for real-time rover safety in uncertain terrain.
Read the full abstract

Chemical Physics

Variational free complement method with Gaussian expanded complement functions: extensions of general atomic framework including helium and lithium ground states with geminal induced expansion length

We expand the previous framework of hierarchical-decontraction Gaussian expanded complement functions approach (arXiv: 2603.16262v2) of the free complement (FC) method to $N$-electron atomic systems including general permutations. We consider the possibility that the Gaussian expansion length depends on the presence of the geminal $g$ function. To be more specific, fewer expansion length shall be used when the two-electron $g$ function is combined with the initial wavefunction. The present approach is applied to the ground states of the helium and lithium atoms. The complement function generations via 2s$\rightarrow$1s channel in the lithium ground state are discussed. Symmetries for the associated $N$-electron integrals are analyzed.
Read the full abstract

Computer Vision

A forensic evaluation framework for egocentric foundation models: an EFM3D case study for crime scene reconstruction

Foundation models increasingly combine geometry, semantics, and spatial context for three-dimensional scene understanding, but their suitability for forensic crime scene reconstruction remains unclear. This study presents a structured forensic evaluation framework for egocentric foundation models and demonstrates it using the Egocentric Foundation Model for 3D (EFM3D). Four simulated indoor crime scenes were recorded with Meta Project Aria Gen~1 at Robert Gordon University's East Lodge House forensic training facility. The framework evaluates room-scale reconstruction, semantic scene understanding, computational performance, and forensic suitability. Across the four scenes, EFM3D generated meshes averaging $12{,}668\pm2{,}430$ vertices and $24{,}838\pm4{,}903$ faces, with 39,529 recorded semantic predictions spanning 21–24 categories at a mean confidence of 0.40. Because reference meshes and expert-labelled 3D ground truth were unavailable, these statistics are treated as descriptive outputs rather than measures of geometric or detection accuracy. Final inference runs processed $1{,}388\pm203$ snippets per scene at $0.69\pm0.01$ samples/s and required $50.77\pm6.12$ minutes per recording. A seven-criterion capability profile indicated stronger support for room geometry, furniture localisation, scene context, and spatial relationships than for fine-scale forensic evidence. The results position EFM3D as a complementary tool for contextual documentation and scene review rather than a replacement for high-fidelity forensic capture. The framework provides a reproducible basis for evaluating future egocentric foundation models in specialised forensic settings.
Read the full abstract

Mathematical Physics

On transition probabilities for the ASEP on the half-line

We study the asymmetric simple exclusion process on the half-line with particle insertion and no particle removal at the boundary. We point out that there exist two distinct multiple contour integral formulas for the transition probability between two finite particle configurations containing the same number of particles. We recall one of these formulas from the literature and prove a second, new formula by showing that it satisfies the evolution equation using a coordinate Bethe ansatz calculation and verify that the initial condition is implemented by the contour integrals. As an application we sum this formula over all final configurations and obtain an integral formula for the probability that the number of particles is conserved up to a given time. The proofs use symmetric rational functions and their properties together with contour deformations.
Read the full abstract

All authors Alexandr Garbali, Cengiz Gazi, Jan de Gier, William Mead, Michael Wheeler

Geometric Topology

Real quadratic fields and finite quantum dilogarithms II: traces of quantum modular cocycles

We study traces of quantum modular cocycles at fixed points of hyperbolic elements $\gamma\in\operatorname{SL}_2(\mathbb{Z})$, formulate an algebraicity conjecture for these traces, and prove it for the cocycles associated to 1-dimensional state-integrals. In addition, we analyse a finite TQFT associated with $\gamma$ and conjecture that its invariants for 3-manifolds agree with the traces of their $\mathfrak{sl}_2$ quantum modular cocycles. The finite TQFTs predict a yet-to-be-defined motive with cohomology equipped with a $\mathbb{Z}/2\mathbb{Z}$-degree grading and a $\mathbb{Z}$-weight grading coming from a version of the Riemann hypothesis, which suggests a similar grading should exist for an extension of the Kauffman bracket skein module, for the framed $\operatorname{SL}_2(\mathbb{C})$ Floer cohomology of Abouzaid–Manolescu, and for the quantum modular cocycle. This leads to surprising additive identities among algebraic units, which we confirm for the complement of the simplest hyperbolic knot $4_1$, as well as for $S^1 \times \Sigma_g$ for a closed genus $g$ surface.
Read the full abstract

Physics and Society

Predictability of Collective Behavior in Networks of LLM Agents

As AI agents built on large language models (LLMs) increasingly interact at scale, anticipating their collective behavior becomes necessary. Unlike humans, individual LLM agents can be probed in {controlled scenarios}; however, the predictability of their collective outcomes hinges on what such measurements capture. We study information cascades on networks of LLM agents that decide whether to share a post given how many of their contacts have shared it. At the same fraction of sharing contacts, an agent's response depends on how many contacts it has and on the post. Responses that ignore the number of contacts mispredict the cascades, increasingly on heterogeneous networks, by an error we derive analytically and that network topology alone cannot reveal. Measured in context, the response predicts, through mean-field theory, the cascades of full LLM simulations, tipping points included. Anticipating the collective behavior of AI agents thus requires evaluating them in their local network context.
Read the full abstract

Econometrics

When the Norm is Wrong: Social Norm Misspecification and Targeting with an Application to Microfinance Adoption

Structural models of peer effects typically capture peer influence through the average, aggregate, or in some cases through extreme peer actions. Such specifications, however, are imposed ex ante and are typically ad hoc. We study the consequences of social norm misspecification by extending the constant elasticity of substitution (CES) norm from continuous to discrete choices. Considering a binary-choice game with incomplete information, we derive results on existence and uniqueness of a rational expectations equilibrium and establish conditions for identification of the structural parameters. We estimate the model by the Nested Pseudo-Likelihood estimator, implemented using the spectral algorithm. Monte Carlo simulations show that the CES estimates of the peer effects approach their true values as the sample grows, whereas imposing a linear norm biases them even in large samples. We apply the framework to microfinance adoption in India and find a strongly max-leaning norm, with significant heterogeneity in curvature and peer effects across villages grouped by caste composition and wealth. In a counterfactual exercise that allocates a fixed subsidy budget to maximise adoption, the estimated and linear norms target very different households, with the former favouring village "leaders", who are more likely to adopt and were the first to be informed about microfinance.
Read the full abstract

Mathematical Physics

Certified finite-precision readout for $p$-adic kernel aggregation

In non-archimedean numerical analysis, a verified $p$-adic approximation can determine an absolute value, and hence downstream probabilistic rankings and classification decisions, with mathematical exactness long before the underlying number is fully resolved. We turn this observation into rigorous finite-precision certificates for norm-based readouts of amplitude families. For a common error radius, a nonempty resolved set already fixes every exact maximizer, ties included, at the sharp integer absolute precision $v_*+1$, and conditioning on that set yields an exact total-variation identity separating unresolved $p$-adic mass from real rounding error. For aggregates of $p$-adic exponentials, we give a centered-moment certificate with an explicit remainder bound: on rational data it runs in exact rational arithmetic without individual exponential evaluations, attains its minimal bound at an active phase, and resolves every nonzero aggregate at high order; we also obtain an aggregate perturbation bound free of summand-count factors. For multiclass $p$-adic classifiers that decode by minimizing the norm of a class discriminant, we prove a necessary and sufficient certification condition and show that the sharp common precision threshold is $v^{(2)}+1$, governed by the second-largest discriminant valuation, so the winning discriminant need not itself be resolved. We apply these certificates to discrete $p$-adic path-integral actions on uniform affine chains, deriving explicit action centers and proving exact output-independence at zero effective quadratic coupling. Pseudocode and a reproducible reference implementation accompany the theory.
Read the full abstract

High Energy Physics (Phenomenology)

NLO QCD corrections to $W^{+}$-boson decays into doubly heavy baryons

The next-to-leading-order (NLO) QCD corrections have been investigated for the production of doubly heavy baryons via the inclusive $W^{+}$-boson decays $W^{+}\to\Xi_{cc}+\bar{c}+\bar{s}+X$ and $W^{+}\to\Xi_{bc}+\bar{b}+\bar{s}+X$ within the nonrelativistic QCD framework. The decay widths, the kinematic distributions, and the uncertainties induced by the heavy-quark masses and the renormalization scale have been analyzed for both the color-antitriplet ($\bar{\mathbf{3}}$) and the color-sextet ($\mathbf{6}$) $S$-wave states. With the central input parameters $m_{c}=1.8~\mathrm{GeV}$ and $m_{b}=5.1~\mathrm{GeV}$, the $K$ factors are $2.27$ for $\Xi_{cc}$ and $2.95$ for $\Xi_{bc}$, i.e., the NLO corrections increase the corresponding LO widths by about $127\%$ and $195\%$, which shows that the NLO corrections are indispensable for quantitative predictions. The individual-channel $K$ factors range from $1.94$ to $5.53$, with the largest one occurring for $\Xi_{bc}$ in the $[{}^{3}S_{1}]_{\mathbf{6}}$ channel. The distributions show that the NLO corrections enhance the low invariant-mass region and shift the baryon energy spectrum toward smaller energy fractions. With $\Gamma_{W}=2.085~\mathrm{GeV}$, the inclusive branching ratios are $8.86\times10^{-5}$ and $3.63\times10^{-6}$ for $\Xi_{cc}$ and $\Xi_{bc}$, respectively, which provide an NLO benchmark for future experimental searches.
Read the full abstract

All authors Hong-Hao Ma, Xu-Chang Zheng, Juan-Juan Niu, Xing-Gang Wu

Cryptography and Security

Lessons Not Learned: Why Cross-Chain Bridges Keep Losing Billions the Same Way

Bridge hacks have cost billions, often without exploiting a code bug. In the cross-chain bridge hacks we catalogue from 2021 to 2026, failures of security posture, meaning how a bridge's people, code, controls and off-chain infrastructure are set up, account for 88% of reported losses since 2023, and nearly all repeat earlier failures. Code audits cannot catch these failures, so we propose a bridge security assessment rubric. It scores each bridge deployment from on-chain evidence, and a hard cap drops the score to the lowest value when a fatal risk is present. A token is only as safe as its weakest deployment if all its deployments share one backing. For attacks nobody anticipated, we show that rate limits and an emergency stop still bound the loss. We build an analyzer that applies the rubric to deployed bridges and monitors them for changes, and an agentic analyzer that runs the same checks on any bridge. Operators learn what to fix, and users can compare bridges. Our rubric's controls cut bridge hack losses 105-fold, from \$3.52B to \$33M.
Read the full abstract

Machine Learning

RAIL: Representation-Aligned Imitation Learning for Student-Compatible Teacher Policies

Reinforcement learning (RL) from raw sensory inputs such as images or onboard sensors can be challenging due to high-dimensional observations and sparse rewards. A common strategy is to train a teacher policy with access to privileged state information and then distill it into a student policy that acts from raw inputs alone. However, when the teacher relies on information unavailable to the student, exact imitation may be impossible, creating an irreducible imitation gap. Existing approaches typically mitigate this mismatch through careful reward shaping or additional reinforcement learning on the student policy after the teacher has been trained. Instead, we introduce Representation-Aligned Imitation Learning (RAIL), which addresses the mismatch during teacher training itself. RAIL learns a latent representation shared across teacher and student observations using contrastive learning, and trains the teacher policy directly in this space. Because the teacher policy is constrained to operate on this shared space, it learns behaviors that are reproducible from student observations. This substantially reduces the imitation gap while preserving task performance. Across multiple environments, RAIL outperforms strong baselines without reward modification or post-hoc student fine-tuning, and surpasses direct reinforcement learning from raw observations. The learned representations further enable zero-shot transfer to new tasks through teacher-only training. Project site: https://merajmammadov.com/rail/.
Read the full abstract

Signal Processing

Movable Antenna-Assisted Covert Communications: Robust Beamforming and Antenna Position Optimization

In this paper, we study a movable antenna (MA)-enhanced covert communication system, where a legitimate transmitter (named Alice) serves multiple legitimate users in the presence of multiple aerial wardens that attempt to detect the transmission from Alice. To ensure data covertness with imperfect knowledge of warden channel state information (CSI), we develop a unified robust covert transmission framework that incorporates bounded and statistical CSI error models. In particular, we aim to jointly optimize the transmit beamforming, the artificial noise (AN) covariance matrix, and MA positions at Alice to maximize the minimum weighted signal-to-interference-plus-noise ratio (SINR) among the legitimate users subject to data covertness constraints. The resulting optimization problem is intractable to solve optimally due to the strong coupling among the optimization variables and the random channel uncertainty. To deal with this issue, we propose a block coordinate descent (BCD)-based algorithm. Specifically, for the norm-bounded CSI error model, we first employ the S-procedure to convert the covertness constraints into a sequence of linear matrix inequalities (LMIs). Then, a successive convex approximation (SCA) procedure incorporating semidefinite relaxation (SDR) and bisection search techniques is proposed to solve the resulting problem. For the statistical CSI error model, the Bernstein-type inequality is employed to derive deterministic approximations of the probabilistic constraint in terms of LMIs and second-order cone (SOC) constraints that can be efficiently tackled. Numerical results are provided to demonstrate the effectiveness and robustness of the proposed algorithm in improving the performance of MA-enhanced covert communication systems.
Read the full abstract

All authors Xiangyu Dong, Ran Yang, Weidong Mei, Zheng Dong, Yue Xiu, Daniel Benevides da Costa, Zhongpei Zhang, Ning We

Atmospheric and Oceanic Physics

One2N: Recursive refinement breaks the resolution barrier for Earth-system AI

Earth-system AI models approach kilometre-scale resolution, but halving grid spacing quadruples the grid-point count and end-to-end training cost. We introduce One2N, a recursive incremental framework extending a coarse model through $N$ levels of cross-resolution residual correction. At each level the parent is frozen and excluded from backpropagation, so its activations and optimizer states are never retained; only a tier-specific module is trained. A single refinement level on the NeuralGCM-0.7$^\circ$ baseline yields the two-level One2N-0.07$^\circ$, trained on one 80 GB GPU. Against HRES-0.07$^\circ$ and ERA5-0.25$^\circ$ analyses it matches its parent's tropospheric skill while reducing upper-stratospheric errors and resolving finer structure. Tropical-cyclone track errors nearly match the parent's and are lower than Pangu-Weather-0.25$^\circ$ at medium range; RMSE is lower than that model's on 93.46% of common targets. One2N thus turns an existing forecasting system, or an inexpensively trained coarse baseline, into a reusable foundation for high-resolution development, breaking the resolution barrier.
Read the full abstract

All authors Yanyan Tan, Xiangjun Tian, Can Wang, Hongqin Zhang, Yinhai Luo, Min Zhao, Zhe Jin, Xingchao Yang, Zaifei Wang, Banglin Zhang, Aiwen Wang, Jing Zhou, Tao Wang, Baosheng An, Xuliang Zhuang, Shilong Piao, Tandong Yao

Machine Learning

A method to automatically discover symbolic local learning rules

We present a general method for discovering symbolic local learning rules that can replace backpropagation, applicable, in principle, to any deep learning architecture and task. The method automatically discovers local loss functions specific to each layer by (1) constructing a dictionary of monomials of local tensor variables, inspired by Penrose diagram notation, (2) regressing over coefficients weighting these monomials to build layerwise polynomial losses whose gradients together minimize a target global loss. We test the method on classification, reconstruction, and self-supervised learning tasks with small multilayer networks and minimal datasets. Across all tasks, the discovered rules induce non-trivial interactions between layers, which lead to significant reductions in the corresponding global losses. Furthermore, the method selects sparse combinations of monomials, finding local rules that are more interpretable than those generated by MLP-based alternatives. The potential implications of our method range from energy-efficient deep learning to the discovery of biologically plausible learning mechanisms.
Read the full abstract

Information Theory

Fundamental Outage and Diversity Limits of Fluid Antenna Multiple Access Under Correlated Nakagami-$m$ Fading

Existing analytical studies of fluid antenna multiple access (FAMA) are largely confined to Rayleigh fading, independent-port assumptions, or asymptotic selection metrics, leaving the fundamental outage behavior under realistic fading and spatial correlation conditions uncharacterized. This paper fills this gap by presenting the first exact and asymptotically tight outage analysis of interference-limited FAMA systems under spatially correlated Nakagami-$m$ fading. By explicitly modeling the joint statistics of the desired and interference powers across fluid-antenna ports, \textcolor{black}{we first formulate the outage probability in terms of the joint desired-signal and interference distributions and then derive an explicit integral expression under a tractable single-reference-port spatial correlation model.} Building on this result, we establish a rigorous diversity-order characterization revealing that the achievable diversity is jointly dictated by the fading severity and the spatial correlation rank imposed by the finite antenna aperture, rather than by the nominal number of ports. To enable accurate and tractable performance evaluation beyond the asymptotic regime, we further introduce a Gamma-matched outage approximation that preserves both the exact diversity order and the array-gain constant. The resulting closed-form expression provides an accurate characterization over a wide range of signal-to-interference ratio (SIR) thresholds and offers clear physical insight into the role of spatial correlation in FAMA systems. Extensive Monte Carlo (MC) simulations confirm the tightness of the exact analysis, the sharpness of the asymptotic results, and the accuracy of the proposed approximation.
Read the full abstract

Audio and Speech Processing

Nemotron 3 Diarization: Advancing the Accuracy-Efficiency Frontier in Streaming Speaker Diarization

Streaming Sortformer introduced an arrival-order speaker cache to maintain speaker identities across chunks in end-to-end streaming speaker diarization, with a single model supporting configurable latency profiles spanning offline operation to ultra-low-latency streaming. Nemotron 3 Diarization addresses the original model's four-speaker limitation by extending support to eight speakers while improving both accuracy and computational efficiency. The previous FastConformer-Transformer stack is streamlined into a pure Transformer encoder. A learned upsampling layer then maps the encoder's coarse 80-ms representations to 10-ms resolution, enabling fine-grained predictions. Training leverages large-scale simulated mixtures and adds activity and phantom-speaker losses to the standard Sortformer objectives. Despite doubling speaker capacity over its predecessor, the new architecture reduces training memory by a factor of 2.8, makes training $2.2\times$ faster, and delivers up to $6.4\times$ higher inference throughput. Nemotron 3 Diarization establishes a new state of the art in DER on DIHARD III and CALLHOME. When used to condition DiCoW, it also achieves state-of-the-art tcpWER on the single-microphone conditions of AMI and NOTSOFAR-1. Training code is open source, and model weights are released under the permissive OpenMDW-1.1 license.
Read the full abstract

All authors Ivan Medennikov, Taejin Park, Tatiana Timofeeva, He Huang, Alexander Polok, Sofia Kostandian, Weiqing Wang, Dongji Gao, Jagadeesh Balam, Boris Ginsburg

Combinatorics

Odd realisation numbers of rigid graphs

Determining the number of realisations of a rigid graph for a specific choice of edge lengths is a fundamental problem in discrete geometry. Gortler, Healy and Thurston showed that the number of real realisations of a generic framework is always even or equal to one. Despite being rare, there do exist graphs whose number of complex realisations is odd and not equal to one. In this article, we exhibit two necessary conditions for a rigid graph to have an odd realisation number other than one. We also provide an infinite family of rigid graphs with an odd realisation number other than one.
Read the full abstract

Systems and Control

Prospective Action-Resource Graph Multi-Agent Reinforcement Learning for Electric HGV Fleets under Shared Charging Infrastructure

Electric heavy goods vehicle (eHGV) fleet coordination requires each vehicle to choose driving, charging and resting actions whose consequences depend on shared infrastructure. Existing multi-agent reinforcement learning (MARL) methods mainly represent current states, leaving action-specific futures and resource interactions to be learned indirectly. We develop Horizon-Preview and Graph-Coordinated Proximal Policy Optimisation (HGC-PPO), which integrates horizon preview and candidate-resource graph coordination into the PPO decision architecture. Horizon preview represents each legal action through its short-horizon operational consequences. Graph message passing then updates these candidates through station, charger-type and time-slot resources shared across the fleet. On the 794km Folkestone-Glasgow task, HGC-PPO achieved the lowest mean charging cost among five RL methods at {\pounds}504.32 per vehicle, with a fleet completion time of 24.920h. It also completed the task 1.763h faster and cost {\pounds}36.51 less per vehicle than a replay-verified finite-budget MILP benchmark. Across six unseen routes and fleet sizes from 5 to 20 vehicles, HGC-PPO delivered the lowest macro-average completion time, charging cost and charging frequency among the RL methods. A 2 by 2 ablation showed that the full preview-graph architecture performed best across all four operational metrics. These results show that prospective action information and fleet-wide resource coordination improve eHGV decisions under charging and driver-hours constraints.
Read the full abstract

All authors Yiming Xu, Fulong Yao, Maurizio Albano, Rigel Gjomemo, Liana Cipcigan

Computer Vision

CADMA: Capacity-Aware Coverage Decomposition for Generative Model Assessment

Evaluating generative models requires measuring not only sample fidelity but also how well a generated set covers the real data distribution. Recent work has therefore introduced recall-oriented metrics for distribution-level evaluation. However, existing recall metrics often reduce distributional coverage to geometry-based coverage. This simplifying assumption weakens their diagnostic value and interpretability in two ways: these metrics provide only a single aggregate score that cannot distinguish essential diversity failures such as mode dropping and mode collapse, and geometry-based coverage can provide unreliable estimates by assigning high scores to poorly covered generative distributions. We therefore introduce CADMA, a capacity-aware framework for recall-oriented evaluation and diagnosis. CADMA evaluates distributional coverage through local capacity matching between real and generated samples, separating geometry-based coverage from effective coverage. This yields multi-dimensional coverage diagnostics that quantify mode dropping and mode collapse while reducing the overly optimistic estimates produced by geometry-based metrics. Through extensive experiments on toy data, real-world datasets, and generative models, we show that CADMA provides more reliable coverage estimates and more fine-grained, interpretable diagnostic signals of diversity failures than existing metrics. Our code is available at https://github.com/Open-Debin/CADMA_NeurIPS2026.
Read the full abstract

All authors Debin Meng, Zheng Gao, Ziquan Liu, Yanran Li, Ioannis Patras, Georgios Tzimiropoulos

Machine Learning

Controlling Temporal Pseudo-Label Marginals for Stable Online Test-Time Adaptation

Continual Test-Time Adaptation (CTTA) methods update models online using a stream of unlabeled data to overcome distribution shifts encountered at test time. However, CTTA is often unstable, as it can reinforce biases in early predictions and increasingly predict only a few classes, resulting in drift or even collapse. In this paper, we identify narrowing pseudo-label class usage as an indicator of CTTA instability and quantify this behavior. Most existing methods have attempted to address this issue by either selecting reliable adaptation samples or re-weighting the loss to encourage diverse pseudo-labels, but they do not explicitly enforce a class marginal over a temporal window. Instead, we propose Sinkhorn Class-Marginal Regularization (SCALAR), a simple plug-and-play module that controls the pseudo-label distribution over a short temporal window. SCALAR works by maintaining recent predictions in a buffer, balancing their class distribution via the Sinkhorn algorithm to match a specified target distribution, and using the balanced soft targets for adaptation. Without changing the model architecture, SCALAR maintains the per-step runtime and increases the peak allocated GPU memory by only 10 MB on ImageNet-C. Our method is particularly effective on skewed, non-stationary data with severe distribution shifts, where SCALAR significantly alleviates drift and collapse when combined with existing stabilization methods based on entropy or diversity.
Read the full abstract

Cryptography and Security

SNAPRAG: Cost-Efficient Knowledge Extraction Attack on RAG Systems via Adversarial Perturbations

Retrieval-Augmented Generation (RAG) systems improve Large Language Model (LLM) responses by incorporating information from external knowledge bases. However, recent studies have shown that attackers can extract substantial portions of private knowledge bases from RAG systems through repeated querying. Yet, state-of-the-art extraction attacks either rely heavily on external LLMs to generate adversarial queries or require a large number of queries. In this paper, we propose SNAPRAG, a novel and cost-efficient attack for extracting RAG knowledge bases through adversarial perturbations. Our approach algorithmically generates adversarial queries without relying on external LLMs and allows for a strictly configurable query budget. In particular, SNAPRAG uses a discrete token-mutation algorithm to generate novel queries in a strict black-box setting. Our adaptive dissimilarity threshold calibration systematically forces queries into unexplored semantic regions of the target database, enabling extraction with fewer queries. Specifically, to extract half of the target knowledge base, our method requires up to 6.77$\times$ fewer queries than existing approaches. Furthermore, within a fixed query budget, our method recovers up to 96.9%, outperforming the closest approach by 7.7-19.9%. We further evaluate SNAPRAG against active defenses, where it remains the leading attack on four of five defended targets, achieving 50.7-96.3% extraction rate.
Read the full abstract

All authors Melih Coşğun, Mehmet Emre Gürsoy, Alptekin Küpçü, Sinem Sav

Computation and Language

Prune, Don't Merge: Expert Compression of Mixture-of-Experts LLMs on Multilingual Production Tasks

Mixture-of-Experts language models are cheap to run per token but expensive to hold in GPU memory, which drives the cost of serving them in production. One-shot expert compression shrinks them after a single calibration pass. Its two current state-of-the-art methods are REAP for expert pruning, which deletes the least salient experts, and REAM for expert merging, which merges them into the most salient ones. Both have been evaluated only on academic benchmarks. We compare the two on five multilingual production tasks of a web-hosting company, from classification to the generation of marketing HTML email templates, with two 30B-parameter models (Qwen3-30B-A3B-Instruct-2507 and the hybrid Mamba-attention Nemotron-3.5-Lightning) at 25% and 50% of experts removed and the same calibration data. Pruning wins: on the same rows, REAP is significantly better in 16 of 20 comparisons and REAM in 2. At 25%, REAP keeps 93.6% of Qwen3-30B's generative quality and REAM 62.5%, although on benchmarks the two are within a point for the same model. The damage concentrates in the output contract rather than the content: 89.8-95.2% of each significant loss on email templates comes from hard-gate failures such as invalid JSON, broken mobile layouts and runaway generation, and most failed non-English domain descriptions copy the prompt's English opening phrase. Cross-perplexity and logit divergence, two diagnostics from the REAP paper, also track production quality, as does the rate of repetition loops; n-gram diversity does not. We close with practical guidance for compressing MoE models for production.
Read the full abstract

Combinatorics

Cube tilings, partition counters, and the Calkin–Wilf and Kepler trees

A tiling of the integer lattice $\mathbb{Z}^d$ by translates of the discrete cube $K_d=\{0,1\}^d$ cuts every translate $K_d+x$ into subcubes. We study the function $\ell(x)$ counting these parts. Its sum over every translate of $K_d$ equals $3^d$, for an arbitrary translation set. For lattice tilings with a triangular basis we obtain a dimension-reduction formula, and we describe every cocyclic lattice by a single normalized congruence. Along a cyclic generator, the counter of a signed bidiagonal lattice is a row of Stern's diatomic sequence; pairs of counter values realize the Calkin–Wilf tree, and complementary binary words realize the Kepler tree. For the binary lattice, the translated counter labelings of the cube carry generating face partitions, supplied by the tiling itself, which force the levels of the Kepler tree, and every nontrivial face partition has a unique lamination direction. Using finite Fourier analysis we prove that the ordered cyclic profile determines a standard cocyclic cube tiling lattice up to coordinate reflections. For lattices tiling by discrete boxes we determine exactly when the analogous statement holds.
Read the full abstract

Optics

Off-focus detection for SNR enhancement in coherent single-pixel imaging

Coherent single-pixel imaging (CSPI) enables amplitude- and phase-resolved computational imaging with a single coherent detector, offering particular advantages in spectral regimes where high-performance detector arrays remain unavailable. In conventional Fourier-plane CSPI, the detector is typically placed at the focal point to maximize the collected power. However, this focal position does not necessarily maximize the reconstruction fidelity for image reconstruction in the presence of system nonidealities. Here, we introduce detector position as an additional degree of freedom for CSPI optimization. We show that appropriately selected off-focus Fourier-plane positions can improve reconstruction performance by suppressing fluctuation-sensitive coherent backgrounds while preserving spatially encoded signals. Experimentally, using a programmable liquid-crystal terahertz spatial light modulator and coherent heterodyne detection, we demonstrate improved amplitude and phase reconstruction from selected off-focus channels compared with focal detection. A symmetric off-focus pair with complementary transverse-wave-vector weighting further enhances image fidelity, contrast, and compressive-sampling robustness through coherent channel fusion. These results establish detector-position engineering as a new strategy for enhancing the robustness and information efficiency of CSPI.
Read the full abstract

All authors Wei Zhu, Hui Yuan, Hangbing Guo, Sheng Wang, Da Tian, Benwen Chen, Kebin Fan, Caihong Zhang, Huabing Wang, Biaobing Jin, Jian Chen, Peiheng Wu, Hartmut G. Roskos, Jingbo Wu

Data Structures and Algorithms

A 1.764-competitive algorithm for online fractional vertex cover on forests

We study online fractional vertex cover on forests under edge arrival. We give a deterministic algorithm with competitive ratio at most $4 - \sqrt{5} \approx 1.764$, which simultaneously produces an online fractional matching with competitive ratio of at least $1/(4 - \sqrt{5}) \approx 0.566$. This improves on the previous best competitive ratios of $11/6 \approx 1.833$ and $5/9 \approx 0.555$, respectively.
Read the full abstract

Quantum Physics

Commuting-check gauging for fault-tolerant logical measurements

Gauging turns a logical operator into a product of local checks, providing a route to logical measurement and lattice surgery. These checks can fail to commute outside the code space, so measuring one can randomize the outcome of another. In this work, we introduce completion qubits that make the checks commute for Clifford $\mathbb{Z}_2$ symmetries of qubit stabilizer codes. Adding one completion qubit per noncommuting pair preserves the gauged code space and its distance while keeping the checks Clifford. Completion also preserves the low-density parity-check property with a constant-factor qubit overhead. We generalize completion to qudits and establish bounds on the Clifford-hierarchy level of the completed checks in prime dimensions. Global commutation ensures that a fault-free round fixes the syndrome until the next fault and allows us to bound fault propagation independently of the number of measurement rounds. For qubit codes, we show that gauging and ungauging a distance-$d$ parent code implements a logical measurement with circuit fault distance $\Omega(d)$ in $O(d)$ rounds. We demonstrate the construction by gauging two toric codes into the $D(D_4)$ quantum double and back, with an explicit circuit whose circuit fault distance equals the parent code distance. Our results provide a framework for fault-tolerant logical measurements using commuting checks, with applications to magic-state preparation by lattice surgery.
Read the full abstract

Medical Physics

Bayesian Kinetic Model Selection in Dynamic PET via Consistency Models

Accurate voxel-wise parametric imaging in dynamic positron emission tomography (PET), particu- larly with total-body PET (TB-PET) scanners, requires se- lecting the appropriate kinetic model for diverse organs and lesions. Conventional nonlinear least squares fitting paired with the Akaike or Bayesian Information Criterion (NLS+AIC/BIC) is computationally slow, sensitive to lo- cal minima, and prone to selecting simple models under high voxel-level noise, whereas classical Bayesian methods such as Markov chain Monte Carlo (MCMC) and Approx- imate Bayesian Computation (PET-ABC) are too slow for whole-body imaging. In this work, we propose a consis- tency model (CM) that performs simultaneous Bayesian kinetic model selection and parameter posterior estimation. By combining the continuous kinetic parameters and dis- crete model index into one target vector conditioned on the measured time-activity curve (TAC) and input function, the CM samples from the joint posterior distribution p({\theta}, m|y) in seconds per slice with a GPU workstation of 4 NVIDIA A100. We validated the framework on two simulated and clinical applications: whole-body [18F]FDG TB-PET across seven compartment models, and brain [11C]raclopride PET for detecting smoking-induced dopamine release between two neurotransmitter models. In simulations, the CM achieved 97.97% model selection accuracy in the [18F]FDG applica- tion (vs. 58.67% for NLS+AIC) and 91.4% accuracy in the [11C]raclopride application (vs. 73.5% for NLS+BIC), while closely matching reference MCMC parameter posteriors. In clinical applications, CM generated whole-body FDG model selection maps, in which each regions was assigned litera- ture established sensible kinetic models; CM also success- fully mapped voxels in the brain where dopamine release was induced by smoking.
Read the full abstract

All authors Yun Zhao, Qinlin Gu, Andrew J. Reader, Yanan Fan, Georgios I. Angelis, Katya Amadita, Gaelle Emvalomenos, Evan D. Morris, Guobao Wang, Steven R. Meikle

High Energy Physics (Phenomenology)

Constraints on the Real Singlet Scalar Dark Matter Model from Higgs Self-Couplings

Taking the real singlet scalar dark matter model as a representative test case, we examine the constraints imposed on the model's parameter space by Higgs trilinear coupling measurements and perturbative unitarity requirements. For the low dark matter mass regime, the constraining strength delivered by the current highest-precision $\kappa_3$ measurement is notably inferior to the exclusion limit derived from perturbative unitarity requirements, while the perturbative unitarity bound itself remains considerably less restrictive than the constraints extracted from the extant experimental upper limit on the Higgs invisible decay branching ratio $\mathrm{BR}_{\mathrm{inv}} \leq 0.107$. Both the coupling modifiers $\kappa_3$ and $\kappa_4$ exhibit a well-defined decoupling behavior, with their respective magnitudes monotonically decreasing as the dark matter mass $m_S$ increases. As a direct consequence, the corrections to both couplings asymptotically approach zero for dark matter masses in excess of 60 TeV, when the parameter $a_2$ remains within its perturbative validity domain. Neither existing collider measurements of the Higgs trilinear coupling nor measurements targeted for planned future collider facilities possess the requisite sensitivity to impose more stringent constraints on the model's viable parameter space than the well-established bounds derived from experimental probes of dark matter relic density, invisible Higgs decay signatures, dark matter direct detection, and dark matter indirect detection. Nevertheless, under the scenario that the experimental measurement uncertainty of $\kappa_3$ reaches the $\mathcal{O}(0.1\%)$ scale or lower, non-negligible constraints acting on a subset of the large mass parameter space associated with the WIMP dark matter scenario can be expected.
Read the full abstract

All authors Hao-Yu Luo, Kun-Peng Chen, Han Hu, Yun Wei, Ming Zhong

Optics

X-ray parametric down-conversion probes radiation-induced dynamics in hybrid perovskite MAPbBr3

We report minute-timescale time-resolved measurements of X-ray parametric down-conversion (PDC) of X-rays into ultraviolet photons in a MAPbBr3 single crystal under continuous irradiation. Despite the known radiation sensitivity of hybrid perovskites and the high photon flux required to detect the intrinsically weak PDC signal, the nonlinear response remains measurable throughout the irradiation sequence. Both PDC and Bragg diffraction exhibit a slow decay consistent with diffusion-driven damage dynamics on timescales reported for bromide-vacancy and proton migration. However, despite their similar decay rates, the PDC and Bragg signals evolve toward markedly different steady states, while an additional fast decay component is observed in the Bragg response but absent in the PDC signal. Density functional theory calculations identify N 2s states of the methylammonium cation as the most likely contributors to the measured PDC response and show that their Fourier component is only weakly affected by bromide-vacancy formation. These results establish nonlinear X-ray diffraction as a channel-selective probe of radiation-induced dynamics and demonstrate that radiation damage can couple differently to nonlinear and linear X-ray responses.
Read the full abstract

All authors S. Sofer, O. Sefi, G. J. Man, H.N. Barad, N.P. Jasti, D. Cahen, S. Shwartz

Artificial Intelligence

ECOS: Environment-Calibrated On-Policy Self-Distillation for Language Agents

On-policy self-distillation lets a language agent act as its own Teacher by conditioning on privileged context such as skills written from its own trajectories. Its guidance, however, is weighted by how much this context shifts token likelihoods rather than by whether it improves behavior, so errors in self-written skills return through later updates. We propose environment-calibrated on-policy self-distillation, which uses environmental outcomes to decide how much to trust the Teacher's current guidance and which self-written skills to reinforce. Command-level advantages, estimated by comparing commands executed at similar observations, scale each Teacher preference and reward reflections and skills whose induced preferences favor higher-advantage commands. Without skills at inference, 3B and 7B agents reach 90.5% and 93.3% macro success on seen ALFWorld tasks and 75.5% and 80.5% success on WebShop. On unseen ALFWorld tasks, they outperform GRPO under the same protocol by 11.0 and 7.6 percentage points, respectively. In an ablation study, supervising only executable commands rather than the Actor's full response raises 3B WebShop success from 52.3% to 75.0%.
Read the full abstract

All authors Chenhao Zhang, Zhongan Wang, Mingxuan Li, Hehe Fan, Zheng Zhu, Ming Li