Computation and Language
AtomSearch: Evidence-Grounded Tree Search for Agent Skill Optimization
Jia Tong, Senyou Deng, Min Peng et al. (+1 more)
Agent skills adapt frozen models through editable external instructions. Revising several rules at once makes changes difficult to inspect, while judging each local edit only by its immediate score can discard useful intermediate states. We introduce AtomSearch, which keeps revision actions local while using subsequent skill states to guide search. Contrastive atomic editing links each primitive rule edit to failed and successful executions, a behavioral hypothesis, and an intended scope. Skill-State Tree Search (STS) evaluates short edit compositions and selectively retains them as executable states that can be revisited and extended. Outcome
memory supplies later proposals with observed repair and regression statistics. We evaluate AtomSearch across four tasks and five target models. With deepseek-v4-flash, it exceeds SkillOpt on three of four tasks, including a 9.68-percentage-point gain on LiveMathematicianBench, and SkillOpt's final skills contain 2.3-19.7 times as many characters. In 67% of completed searches, the best validation state passes through an ancestor that was not highest-scoring at its depth. These results show how local rule edits can support the exploration of useful revision paths and yield compact skills, with additional offline evaluations used to explore alternative states.
Read the full abstract ▾
All authors Jia Tong, Senyou Deng, Min Peng, Jianfeng Pan
Mesoscale and Nanoscale Physics
Observation of correlations beyond the single-particle picture in fermionic counting statistics
Johann Zöllner, Hendrik Mannel, Daniel Oppers et al. (+3 more)
The analysis of charge fluctuations in a mesoscopic device can shed light on the fundamental question of whether or not a single-particle picture of electron transport is appropriate, even in the presence of Coulomb interaction. It has been shown on general grounds that mapping the interacting system onto a fermionic single-particle description is subject to strict constraints on the charge-transfer statistics. These constraints can be formulated in terms of a sign rule for the factorial cumulants of the full counting statistics. We report here on the experimental observation of a sign-rule violation in a most elementary model system, i.e. a single-level
quantum dot, and provide theoretical modeling that can quantitatively reproduce the experimental data. This reveals the presence of correlations introduced by electron-electron interaction.
Read the full abstract ▾
All authors Johann Zöllner, Hendrik Mannel, Daniel Oppers, Martin Geller, Axel Lorke, Jürgen König
Neurons and Cognition
Reciprocity coordinates a three-way trade-off in recurrent neural circuits
Xiaoyu Zhang, Mingyi Huang, Yixuan Cui et al. (+2 more)
Biological neural
circuits contain both unidirectional and reciprocal connections, but how reciprocity affects
circuit computation remains unclear. We varied reciprocity h̊o} across 21 levels in recurrent-network simulations while holding network size, connection density, and directed-edge count fixed. Low-reciprocity networks produced more diverse stimulus-driven activity and performed better on four continuous-input tasks, whereas higher reciprocity generally improved performance on three delayed-
memory tasks. Increasing reciprocity reduced online
memory capacity and accelerated recovery from perturbations. A post-cue noise experiment showed that higher reciprocity suppressed internal noise propagation but made cue-evoked states less distinguishable, and component matching separated the contributions of cue separation and within-cue dispersion to readout robustness. Within this ensemble, reciprocity thus mediates a three-way trade-off among representational capacity, local stability, and task-state separability: higher reciprocity lowers capacity, strengthens stability, and compresses separability. Whether it improves behaviour depends on whether the stability gain outweighs the lost separation between task states. Connectome and microcircuit
datasets from five species showed intermediate reciprocity, while reciprocity varied substantially across adult Drosophila
brain regions, providing compatible structural context and motivating
circuit-specific functional tests.
Read the full abstract ▾
All authors Xiaoyu Zhang, Mingyi Huang, Yixuan Cui, Xin Lu, Yuguo Yu
Artificial Intelligence
When Protection Overrides Consequences: Understanding Protection Mode in Large Language Model Decision-Making
Yuanfan Li, Junyi Chen, Qi Zhou et al. (+3 more)
Frontier
language models excel at complex tasks yet may prioritize protection over larger physical consequences in simple, urgent decisions. GPT-5.6 Sol, for example, moves a girl playing in shallow water before retrieving a submerged camera. We call this Protection Mode and study it with ProtectionBench: 285
prompts, 34 physical families, and nine models. At lowest effort, at least one of four responses prioritizes the subject on 34.74% of Sol
prompts and over 83% of MiniMax M3 and Qwen3.5-27B
prompts; none of four independent human raters does so. Controls show that choices usually track physical risk. We elicit both risks before requesting an action. Some models assess risk incorrectly; others state the correct ordering but still choose the subject. Qwen3.5-27B with thinking shows 100% correct risk ordering but 57.81% subject choice among eligible conversations on eight selected scenes.
Attention audits find no stable reduction in
attention to relevant facts; removing the stated risk direction reduces object-first choice in dots3-note-prev but not detectably in Qwen3.5-27B. Adding a general "people or animals before property" rule increases subject choice by 19-38 percentage points, whereas factual reminders favor the object by 0.7-5.2 points. Removing naturally generated versions of that rule lowers subject choice by 10-15 points. We also examine whether post-training introduces this pattern. Base models already show Protection Mode; post-training changes justifications more than choices.
Read the full abstract ▾
All authors Yuanfan Li, Junyi Chen, Qi Zhou, Rui Yang, ZhangLiujie, Weihang Chen
Cryptography and Security
Quiet Triggers, Loud Footprints: A Tri-Space Measurement Study of Backdoor Stealthiness
Behrad Tajalli, Henk Berendsen, Xiaoyun Xu et al. (+1 more)
Backdoor attacks on deep neural networks are commonly evaluated by attack success rate and benign accuracy, yet their stealthiness remains inconsistently defined and measured. Existing work often studies only one observable aspect of stealthiness, such as trigger imperceptibility, feature-space separability, parameter-space concentration, or evasion of particular defenses. This creates a fragmented picture: an attack may appear stealthy under one metric while leaving conspicuous footprints elsewhere. This paper presents a tri-space measurement study of backdoor stealthiness in image classification. We distinguish latent stealthiness—the ideal but unobservable difficulty of distinguishing a backdoored system from a benign one—from measurable footprints in three spaces: input, feature, and parameter space. Under this framework, we evaluate representative backdoor attacks across four
datasets and three model architectures using a broad suite of footprint metrics. Our results show that stealthiness is not monolithic: attacks that are quiet in one space are often loud in another, and existing attacks rarely achieve consistently small footprints across all three spaces. We further analyze agreement and disagreement among stealthiness metrics and study how footprint measurements interact with attack success rate. Based on observed blind spots in existing measurements, we introduce two diagnostic metrics: Class-Specific Davies–Bouldin Index (CDBI), which captures class-conditional feature overlap, and TAC–UCLC Product (TUP), which identifies channels that are both trigger-responsive and downstream influential. These metrics are intended to augment, not replace, the broader measurement suite. Finally, we provide practical guidance for future stealthiness evaluations.
Read the full abstract ▾
All authors Behrad Tajalli, Henk Berendsen, Xiaoyun Xu, Stjepan Picek
Machine Learning
Learning Graph Energies for Finite-Time Generation
Amartya Banerjee, Xingyu Xu, Caroline Moosmüller et al. (+1 more)
Energy-based models can represent distributions over discrete structures such as molecular graphs, but an energy that is accurate at
equilibrium need not yield accurate samples when the sampler stops after a fixed number of steps. We introduce finite-horizon energy learning, which adapts the acceptance energy of a Metropolis–Hastings sampler to the distribution it returns after a prescribed number of steps, with initialization, proposals, and annealing schedule held fixed. For graph generation, a frozen pretrained model supplies initialization and proposals, and a learned energy governs acceptance. A training-only discriminator scores completed samples, and coupled continuations of the accepted and rejected branches of a proposal give an acceptance-gradient
estimator that is unbiased for a fixed discriminator under stated conditions, without equilibrated negative samples or differentiation through discrete states. Generation uses the learned energy in the original single-trajectory sampler, without discriminator evaluations or reranking. On twelve pairwise binary targets with exactly evaluated terminal laws, learned energies have lower median terminal total variation than the
equilibrium-correct energy in every tested short-horizon setting, and than its numerically tuned scalar multiple in most settings. On MOSES, a learned two-coefficient combination of pretrained Graph Energy Matching (GEM) potentials lowers FCD by 3.7–10.5% relative to scale-calibrated GEM at matched sampling steps, with validity gains at the shortest budget and a sub-point validity trade-off at higher budgets.
Read the full abstract ▾
All authors Amartya Banerjee, Xingyu Xu, Caroline Moosmüller, Harlin Lee
Computation and Language
Does Machine Translation Retain Media Bias?
Silvia V. Gargova, Tim Menzner and Jochen L. Leidner
Within the most recent two decades, machine translation (MT) has been advanced from a research area to a mature, deployed technology for prose text and most
language pairs. However,
language has different dimensions of meaning, and the alluded progress has primarily been demonstrated for propositional meaning. In this small exploratory study, we investigate whether one type of affective meaning, namely different types of linguistically expressed media bias, are retained under automatic translation for two
language pairs, English and German as the source
languages, and Bulgarian as the target
language. We evaluate four state-of-the-art translation methods to answer the question whether and to what degree a loss of non-neutral
language (specifically, 27 types of media bias) can occur as a result of applying MT. Our empirical findings are consistent with the claim that MT does not fully retain existing news bias of its input source text, but also may add extra bias of its own. To the best of our knowledge, this is the first experimental study of MT media bias retention.
Read the full abstract ▾
Optimization and Control
Recursive Gradient Retraction Methods for Stochastic Nonconvex Optimization on Stiefel Manifolds
Zhijie Yuan, Lei Sun and Ganzhao Yuan
Nonconvex optimization on Stiefel
manifolds underlies many problems in machine learning and deep learning. However, existing
stochastic Riemannian retraction methods often rely on stepsizes that require prior knowledge of the gradient Lipschitz constant, a prescribed iteration horizon, or restrictive batch-size choices. We develop RGR, a recursive gradient retraction method for
stochastic nonconvex optimization on Stiefel
manifolds. Instead of recursively updating gradient
estimators in changing tangent spaces, RGR constructs them in the Euclidean space and subsequently maps them to tangent space, thereby avoiding vector transport. The method comprises three variants: RGR-SPIDER for finite-sum optimization and RGR-SRM and RGR-EMA for online
stochastic optimization. RGR-SPIDER is Lipschitz-free and horizon-free, as its adaptive stepsize requires neither prior knowledge of the gradient Lipschitz constant nor a prescribed iteration horizon. RGR-SRM and RGR-EMA are mini-batch and horizon-free methods whose convergence guarantees do not rely on prescribed large batch sizes and whose parameters can be chosen without prior knowledge of the total iteration budget. We show that RGR-SPIDER reaches an $\varepsilon$-critical point using $\mathcal{O}(\varepsilon^{-2})$ iterations and $\mathcal{O}(N+\sqrt{N}\varepsilon^{-2})$ component-gradient evaluations, where $N$ denotes the number of component functions. For the online setting, RGR-SRM and RGR-EMA achieve iteration complexities of $\tilde{\mathcal{O}}(\varepsilon^{-3})$ and $\tilde{\mathcal{O}}(\varepsilon^{-4})$, respectively. Experiments on principal component analysis, low-rank
matrix completion, and orthogonal convolutional neural networks demonstrate faster empirical convergence than methods that construct recursive gradient
estimators directly in Riemannian tangent spaces.
Read the full abstract ▾
Logic in Computer Science
Finite Satisfiability of Two-Variable Logic with Arbitrarily Many Successors
Artur Jeż, Jakub Michaliszyn and Piotr Witkowski
We study finite satisfiability of two-variable first-order logic over structures equipped with an arbitrary finite number of independent successor relations of linear orders. The number of successor relations is part of the input, and the vocabulary may contain further unrestricted unary and binary relations. We prove decidability in deterministic doubly exponential space together with a triply exponential bound on the size of a smallest finite model. This settles the unrestricted finite case beyond two successors.
Read the full abstract ▾
Machine Learning
MorphingNet: Symbolic Regression Where the Network Is the Formula
Jincheng Gao, Jiahe Yi and Shuo Sun
We introduce MorphingNet, a network for symbolic regression (SR) whose units change their function continuously. Instead of searching over expressions built from a fixed operator library, MorphingNet trains the form of each function as a parameter, so SR becomes one continuous optimization problem. Its units are either pole-free rational functions or complex exponential–logarithm functions, which move between exponentials, sinusoids, damped oscillations, power laws and logarithms as their parameters change. The network is a readable formula at every step of training, and no operator library is needed to train it; a structure-recovery step afterwards rewrites the trained formula into short elementary expressions. Training is dense linear
algebra, so it is GPU-friendly and scales to large data. MorphingNet achieves state-of-the-art accuracy on SRBench, recovers most generating formulas on synthetic
benchmarks, and on large data is up to two orders of magnitude faster than the strongest search method. By turning the search for a formula into the training of a network, MorphingNet offers a new paradigm for SR and for scientific discovery from data.
Read the full abstract ▾
Quantum Physics
Joint-measurement advantage and finite-time transitions in accelerated detector registers
Zhiming Huang
The Unruh response thermalizes accelerated detectors, but the
entanglement remaining in a register does not by itself specify what distributed measurements can recover. We study this distinction for a four-
qubit cluster state whose two assisting nodes independently undergo generalized amplitude damping, while the target nodes remain isolated and the environments are inaccessible. We derive the exact joint-assisted concurrence and the exact one-way value for arbitrary local measurements. The joint optimum equals the bipartition negativity, and computational-basis measurements attain the one-way optimum. A polar-decomposition argument, together with compactness, establishes a necessary and sufficient condition for separable measurements to attain the joint optimum and proves a strict gap whenever this condition fails. For assisting detectors at equal Unruh temperature with relaxation-rate ratio $r>1$, joint advantage exists precisely when the thermal excitation factor satisfies $q<1/r$. At finite positive temperature it ends at a unique finite time, while the remaining
entanglement is nonzero. For $r=2$, both the transition time and remaining resource have closed expressions. These results identify when joint assistance is necessary within the independent weak-coupling thermalization model.
Read the full abstract ▾
Logic in Computer Science
A Non-Dichotomy for CSPs of Stable $\omega$-Categorical Structures
Manuel Bodirsky, David Bradley-Williams, Ehud Hrushovski et al. (+1 more)
We construct countable, stable, $\omega$-categorical structures with at most two binary relations whose CSPs exhibit three complexity regimes: NP-complete, decidable but not in NP, and undecidable. The templates arise from a controlled Hrushovski construction of stable $\omega$-categorical graphs $M_\alpha$ for suitable irrational parameters $\alpha>1$. Some of these graphs have NP-complete CSPs. Expanding the graphs $M_\alpha$ by the non-edge relation turns age membership into a restriction of the CSP. For every time-constructible bound $t$, there is such an expansion whose CSP is decidable but not in DTIME$(t)$. For an explicit parameter $\alpha$, the restriction to complete instances belongs to $\mathrm{coNP}$ and the full CSP belongs to $\Sigma_2^{\mathrm P}$. Separating the ages of the $M_\alpha$ using graphs with large girth and prescribed density yields $2^{\aleph_0}$ distinct CSPs of these expansions, all but countably many of which are undecidable.
Read the full abstract ▾
All authors Manuel Bodirsky, David Bradley-Williams, Ehud Hrushovski, Christoph Spiess
Astrophysics of Galaxies
The physical and chemical structure of TMC-1. The spatial distribution and excitation conditions of carbon, oxygen, and sulfur bearing species: Depletion of CO and CS
J. Cernicharo, B. Tercero, N. Marcelino et al. (+3 more)
We present a detailed study of the physical and chemical structure of TMC-1 from the observed spatial distribution of HC3N, NNH+, CN, and their isotopologues, the O-bearing species C18O, C17O, CH3OH, CH3CHO, and H2CO, and the S-bearing species CS, HCS+, H2CS, and H2CCS. The well known filamentary structure of TMC-1 can be interpreted as a single cloud with a velocity gradient perpendicular to the filament, or less likely as due to the
alignment of three independent clouds in the line sight with velocities of 5.3-5.7, 5.7-5.9, and 5.9-6.3 km/s. The main contribution to the emission of all studied C-bearing species is found at 5.83 kms/s. We conclude that the three velocity components are dynamically connected and that C-bearing molecules show a different behavior of their abundances compared to O-bearing species, which have emission peaks at 5.7 and 6.25 km/s, and appear at different spatial locations than cyanopolyynes and CnH and CnN radicals. The detection of 13C18O, with exactly the same line profile than that of C18O and C17O, excludes the possibility that the decrease of their emission at 5.83 km/s, could be due to self absorption or absorption by a foreground cloud component. Hence, as indicated by the spatial distribution of C18O and C17O, CO is significantly depleted at the velocities and positions where the emission of the carbon chains and radicals is found. This suggests that the emission of O-bearing species is coming from the external layers of the carbon-rich filaments. A comparison between S-bearing species and other molecules indicates that CS, CCS, HCS+, and H2CS are significantly depleted in the zone of CO depletion. A clear increase of the deuteration of several molecular species is found in the zones where CO and CS are depleted.
Read the full abstract ▾
All authors J. Cernicharo, B. Tercero, N. Marcelino, G. Esplugues, G. Molpeceres, M. Agúndez
General Relativity and Quantum Cosmology
Quantum noise and torque correlations in a torsion-bar Sagnac speed meter
Yuki Inoue, Mario Onglao, Daiki Tanabe et al. (+3 more)
Quantum fluctuations entering through weakly transmitting mirrors can limit a torsion-bar Sagnac speed meter for low-frequency gravitational-wave detection, even when its optical layout suppresses radiation-pressure back action. We show that, in the Cryogenic sub-Hz cROss torsion bar detector with
quantum NOn-demolition Speed meter (CHRONOS),
vacuum fields entering from behind the end and middle test masses can dominate the low-frequency
quantum noise despite cancellation of the torques on opposite ends of each bar. For a reference configuration with mirror power transmissions of only $10^{-7}$, these fields contribute $89.7\%$ of the total
quantum-noise power spectral density at 0.1 Hz. Omitting them would underestimate the amplitude spectral density by a factor of 3.11. Our input-output model connects the correlations produced by successive optical interactions to the forces on each bar and the measured signal. This connection shows why strong torque cancellation does not, by itself, ensure a corresponding improvement in detector sensitivity. Replacing mirror transmission with absorption at fixed reflectivity leaves the excess
quantum noise essentially unchanged for the two operating points studied. The predicted signal response and
quantum noise agree with independent FINESSE 3 calculations as recycling detuning and homodyne angle are varied. These results establish rear
vacuum fields as a limiting noise source in this torsion-bar speed meter and provide quantitative requirements for its mirror and readout design.
Read the full abstract ▾
All authors Yuki Inoue, Mario Onglao, Daiki Tanabe, Masashi Hazumi, Vivek Kumar, Apimook Watcharangkool
Astrophysics of Galaxies
The Role of Environment in Shaping the Assembly Timelines of Massive Quiescent Galaxies at Cosmic Noon
Tianmu Gao, J. Trevor Mendel, Letizia Bugiani et al. (+10 more)
JWST observations of massive quiescent
galaxies (MQGs) at $z\sim2$ have revealed diverse
star formation histories (SFHs). In this letter, we investigate whether these variations are linked to the environment in which MQGs reside, as quantified by the host dark matter
halo mass. We identify 12 MQGs from a representative sample of
galaxies at $1.6<z<3.5$ obtained by the Blue Jay survey, a JWST Cycle 1 program, and infer their SFHs via modelling of the deep NIRSpec medium-resolution spectra along with multi-band photometry. By correlating the formation and quenching epochs of individual MQGs with their respective host
halo masses from the 3D-HST survey, we find that after accounting for the
redshift range of the sample ($1.8<z<2.5$), MQGs in more massive
haloes formed at significantly earlier
cosmic epochs. Meanwhile, although these
galaxies also quenched earlier overall, the scatter in quenching epochs across similar
halo masses indicates that the quenching process likely involves other contributing factors and is not solely determined by
halo mass.
Read the full abstract ▾
All authors Tianmu Gao, J. Trevor Mendel, Letizia Bugiani, Minjung Park, Sirio Belli, Joel Leja, Claudia del P. Lagos, Amir H. Khoram, Rebecca L. Davies, Razieh Emami, Gabriel Maheson, Anita Zanella, Benjamin P. Collins
Quantum Physics
Zongxing Xiong, Chengshen Gao and Lvzhou Li
Entropy estimation of discrete probability distributions is a fundamental problem in statistical inference and
quantum property estimation. In this paper, we study the problem of estimating Rényi
entropy $H_\alpha(\mathbf{p})$ for any distribution $\mathbf{p} = (p_i)_{i=1}^n$ given its purified
quantum query-access. For non-integer $\alpha > 1$, our
algorithm estimates $H_\alpha(\mathbf{p})$ to within additive error $\epsilon$ using only $\widetilde{\mathcal{O}}\big(n^{\max\{1/2,1-1/\alpha\}}/\epsilon\big)$ queries, which improves the best known upper bound for dependence on both $n$ and $1/\epsilon$. For $0<\alpha<1$, our
algorithm has query complexity $\widetilde{\mathcal{O}}\big(n^{1/(2\alpha)}/\epsilon^{\max\{1,1/(2\alpha)\}}\big)$, which improves the dependence on $1/\epsilon$ of the current best upper bound without incresing its $n$ dependence. Our approach adapts the multi-level estimation framework to relative approximation of the power sums and leverage binary search to remove the requirement of prior knowledge of them. We also strengthen the precision-dependent lower bound to $\Omega\big(1/\epsilon^{\max\{1,1/(2\alpha)\}}\big)$ for $0<\alpha<1$, which identifies the optimal precision dependence of our
algorithm.
Read the full abstract ▾
Chaotic Dynamics
Stability of inverse cascade in NLS Wave turbulence
Guillaume Costa, Sergey Nazarenko and Giorgio Krstulovic
We study the stability of the inverse cascade in the 3D Nonlinear Schrodinger Wave Kinetic Equation (WKE) in the presence and absence of a condensate. Using theoretical arguments supported by numerical simulation, we show that, the Kolmogorov-Zakharov solution is stable with respect to small isotropic perturbations in the absence of a condensate. Conversely, introducing a condensate component leads to a growth of the condensate occupation and ultimately destabilizes the inverse cascade. The system then evolves towards a new steady-state whose properties depend on the condensate dissipation rate.
Read the full abstract ▾
Computation and Language
To Reason but Not to Stop: Understanding and Recovering Pruned Large Reasoning Models
Minh-Anh Nguyen, Nguyen T. Ha Lan, Xiyan Fu et al. (+2 more)
Weight
pruning can degrade the accuracy of large
reasoning models (LRMs), yet it remains unclear whether it damages
reasoning itself. We find that much of this degradation stems from a failure to stop rather than to reason. As sparsity increases, pruned LRMs increasingly fail to terminate within the
token budget, often after the correct answer has appeared in their
reasoning, instead falling into repetitive loops, even though they still recognize where the dense model would stop. We further find that the hidden-state difference between terminating and looping continuations is consistent across pruned checkpoints, suggesting a shared direction associated with committing to an answer. Building on this, we propose Termination-Guided Activation Steering (TGAS), a training-free method that adds this direction to the residual stream at inference time. Across six LRMs, three
pruning pipelines, and math and code
benchmarks, TGAS improves termination and overall accuracy while reducing repetitive
reasoning and generation length, showing that this failure is actionable.
Read the full abstract ▾
All authors Minh-Anh Nguyen, Nguyen T. Ha Lan, Xiyan Fu, Dung D. Le, Wei Lu
Quantum Physics
Photon Demultiplexing Using Spatial Routing
Hashir Kuniyil, Asad Ali and Saif Al-Kuwari
Spatial sectioning of the far-field noncollinear field from spontaneous parametric down-conversion (SPDC) is a promising approach for multiparty
quantum communication. Although demonstrations with six pairs of users have been reported, the fundamental limits remain unknown. These include the maximum number of channels, the required degree of channel optimization, and the dependence on input parameters. In this paper, we provide a practical framework that answers these questions by connecting input parameters and invoking a design
theorem. For a representative configuration, we realize 12 user-pair channels with routing consistency exceeding $0.9999$ at a $12\%$ reduction in
photon flux. We further show that sectioning improves the accidental-coincidence budget linearly with the number of channels $N$, allowing the aggregate key rate to scale with the user count. These results provide a design strategy for spatial-division multiplexing of entangled
photons analogous to the channel-spacing standards used in wavelength-division networks.
Read the full abstract ▾
Computer Vision
Instruction Mining From Images: Constructing a Synthetic Dataset for Multimodal Learning ILSP DCVLR Competition Submission
Eleni E. Kasoura, Dimitrios Damianos, Georgios Paraskevopoulos et al. (+1 more)
This work presents a compact and reproducible pipeline for synthesizing and curating image-based instruction-response pairs for DCVLR. We draw from the Cauldron collection and use Magpie-style instruction and response synthesis with Qwen2.5-VL-7B Instruct to create concise, image-conditioned instruction pairs. The pipeline consists of instruction synthesis, conditioned on input images and controlled via a system
prompt, sanitization, tagging, and filtering of the instructions based on the score obtained by a reward model (DeBERTa-v3). We retain samples with positive scores to maximize the quality of the generated instructions. We observe that score distributions have a median and mean reward score of ~ 1.53 and ~ 1.75 respectively, indicating that most generated instructions are of good quality. The resulting
dataset contains diverse instructions in the math, spatial, and social
reasoning domains.
Read the full abstract ▾
All authors Eleni E. Kasoura, Dimitrios Damianos, Georgios Paraskevopoulos, Vassilios Katsouros
Computer Vision
SAM3RL: Memory Control via Reinforcement Learning for Visual Object Tracking
Tomáš Čížek, Matej Straka, Illia Volkov et al. (+5 more)
Segment Anything Model 3 (SAM 3) is a foundation model for image and
video segmentation that sets the state of the art in visual object tracking. Its tracker, inherited from SAM 2, stores information from past frames in a
memory bank, managed by a simple First-In-First-Out rule, allowing temporal consistency in complex
videos. Several recent methods improve SAM 2 with hand-crafted
memory update rules, each targeting a specific failure mode, such as distractors, occlusions, or object motion. However, their gains transfer only partially to the stronger SAM 3 backbone. We instead present a more general, data-driven approach that frames
memory control as a sequential decision-making task. SAM3RL is an extension of the SAM 3 tracker with a lightweight
memory controller, trained by reinforcement learning, that decides which frames enter the
memory bank in order to maximize tracking performance. SAM3RL sets a new state of the art on 10 of 12 tracking and
video object segmentation
benchmarks, outperforming all hand-crafted rules without any runtime overhead. The same approach applied to SAM 2 gives SAM2RL, which outperforms hand-crafted rules on 9 of the 12
benchmarks, showing the generality of our method across model generations.
Read the full abstract ▾
All authors Tomáš Čížek, Matej Straka, Illia Volkov, Matěj Kripner, Ervin Macić, Jiri Matas, Klara Janouskova, Martin Schmid
Classical Analysis and ODEs
Elementary ratio and zero limits for varying recurrences
K. Castillo and S. Yakubovich
Two
theorems of Loureiro and Van Assche on ratio and zero limits provided part of the motivation for this note. Their elementary nature permits direct proofs; Appendix B gives complete proofs of both for the canonical initial conditions, using earlier continued-fraction identities. We prove ratio limits and weak limits of normalized zero measures for varying higher-order three-term recurrences with arbitrary coefficient periods and independent limiting functions for their phases. Joint local coefficient limits give ratio convergence without uniform coefficient bounds. For zero measures, convergence in measure to finite measurable profiles suffices, without boundedness or integrability; the limiting measures may have noncompact support. Coefficients may vanish, and real diagonal terms are allowed for ordinary three-term recurrences. The proofs use interlacing, partial fractions with nonnegative residues, local computations of moments, and stability of zero counts under changes of recurrence coefficients. Truncation removes the coefficient bounds from the zero
theorem, whose proof is independent of ratio asymptotics. Uniform bounds give the version with compact support. Power-law limits and their explicit transform formula follow as a special case.
Read the full abstract ▾
Space Physics
The Gannon Storm: Citizen science observations during the geomagnetic superstorm of 10 May 2024
Maxime Grandin, Emma Bruus, Vincent E. Ledvina et al. (+12 more)
The 10 May 2024 geomagnetic storm, referred to as the Gannon Storm in this paper, was one of the most extreme to have occurred in over 20 years. In the era of smartphones and social media, millions of people from all around the world were alerted to the possibility of exceptional auroral displays. Hence, many people not only witnessed but also photographed the aurora during this event. These citizen science observations, although not from scientific instruments operated by observatories or research groups, can prove invaluable in obtaining data to characterise this extraordinary event. In particular, many observers saw and photographed the aurora at mid-latitudes, where ground-based instruments targeting auroral studies are sparse or absent. We created an online survey and circulated it within networks of aurora photographers to collect observations of the aurora and disruptions in technological systems that were experienced during this superstorm. We obtained 696 citizen science reports from over 30 countries, containing information such as the time and location of aurora sightings, observed colours and auroral forms during the geomagnetic storm. We supplemented the obtained
dataset with 186 auroral observations logged in the Skywarden catalogue by citizen scientists. The main findings enabled by the data collected through these reports are that the aurora was widely seen from locations at geomagnetic latitudes ranging between 30{\deg} and 60{\deg}, with a few reports from even lower latitudes. This was significantly further equatorward than predicted by auroral oval models. The reported auroral emission colours, predominantly red and pink and intense enough to reach naked-eye visibility, suggest that the auroral electron precipitation contained large fluxes of low-energy (< 1 keV) particles. (...)
Read the full abstract ▾
All authors Maxime Grandin, Emma Bruus, Vincent E. Ledvina, Noora Partamies, Mathieu Barthelemy, Carlos Martinis, Rowan Dayton-Oxland, Bea Gallardo-Lacourt, Yukitoshi Nishimura, Katie Herlingshaw, Neethal Thomas, Eero Karvinen, Donna Lach, Marjan Spijkers, Calle Bergstrand
Computation and Language
MemDuet: Learning Dual-Process Memory for Long-Context Reasoning
Jiangcheng Song, Haodong Jing, Yulin Mu et al. (+7 more)
Reasoning over infinitely long documents is becoming a reality with recurrent
memory agents. However, increasing context length exacerbates the trade-off between
reasoning performance and inference efficiency. Inspired by Kahneman's account of fast and slow thinking, we propose MemDuet, a dual-process
memory framework that learns to coordinate a small
language model for fast, parallel evidence selection from raw context chunks (System 1) with a large
language model for
memory-update verification and
memory updates in document order (System 2). We train MemDuet in three stages: supervised fine-tuning of System 1 for evidence selection, Group Relative Policy Optimization (GRPO) of System 2 for verification and
memory rewriting, and joint optimization of both policies to coordinate selection and updating for long-context
reasoning. Experiments on diverse long-context
reasoning benchmarks (RULER-HQA, LongBench v2, and 2WikiMultiHopQA) demonstrate a favorable accuracy-efficiency trade-off as context length grows. Averaged across evaluation settings on the three
datasets, MemDuet achieves an end-to-end inference speedup of 9.74x and an output-
token reduction of 73.91% relative to RL-MemAgent.
Read the full abstract ▾
All authors Jiangcheng Song, Haodong Jing, Yulin Mu, Lesong Tao, Jixin Wang, Jin Cui, Tianjing Ren, Runxuan Xie, Yongqiang Ma, Nanning Zheng
Quantum Physics
Deterministic and Programmable Fusion for Scalable Generation of Photonic Graph States
Yan Li, Yunfan Yang, Zenghui Bao et al. (+6 more)
Entanglement in multiphoton graph states is an important resource with various applications, such as measurement-based
quantum computing and
memory-less
quantum repeaters. While some primitive instances have been successfully prepared, scalable generation of large-scale graph states with programmable connections remains an outstanding challenge. Fusion provides a promising "divide-and-conquer" approach by linking smaller resource states into larger, more complex graphs. However, its practical scalability has been limited by the probabilistic nature of conventional fusion protocols, which necessitates extensive hardware overhead and feed-forward resources to achieve deterministic outcomes. Here, we introduce and experimentally demonstrate a deterministic fusion operation with built-in error mitigation, capable of connecting small time-bin-encoded cluster states–originally generated on-demand from a
quantum emitter–into larger, reconfigurable photonic graph states of increased sizes. The fusion operation is realized by nondestructively projecting pairs of flying microwave photonic
qubits into Bell states using superconducting
circuits with in-situ programmability. Crucially, the protocol features active reset and reuse of matter
qubits during graph state generation, thereby circumventing the coherence constraints of the
quantum emitter and enabling the generation of genuine multipartite
entanglement across 13 photonic
qubits. Our results establish a scalable framework for photonic graph-state synthesis, paving the way for fault-tolerant measurement-based
quantum computation and distributed
quantum networks.
Read the full abstract ▾
All authors Yan Li, Yunfan Yang, Zenghui Bao, Jize Yang, Luyan Sun, Yipu Song, Yukai Wu, Hongyi Zhang, Luming Duan
Astrophysics of Galaxies
Sleeping or Dead? Probing the Quenching Mechanism of JWST Dormant Galaxies at z > 3 with ALMA
Mengyuan Xiao, Songnan Qi, Pascal A. Oesch et al. (+16 more)
JWST has revealed a population of "dormant”
galaxies at z>3 with strong Balmer breaks and weak emission lines, whose
star formation appears to have declined only recently. Whether these
galaxies are undergoing a lasting transition toward quiescence or merely experiencing a temporary suppression of
star formation remains unclear. We investigate their molecular gas content using archival ALMA observations of [CII]158um and CO rotational transitions. Starting from all publicly available JWST/NIRSpec prism observations and using joint spectrophotometric fitting, we construct a parent sample at z>3 and log(Mstar/Msun)>9.5 in the ALMA-accessible COSMOS, UDS, and GOODS-South fields. Our final sample comprises 7 dormant
galaxies and 15 quiescent
galaxies with ALMA constraints on their molecular gas content (mostly 3 sigma upper limits). We find that dormant
galaxies exhibit systematically lower
star formation rates averaged over the past 10Myr (SFR10Myr) than over the past 100Myr (SFR100Myr), supporting a recent rapid decline in their
star formation. For the ~70% of the sample with sufficiently constraining ALMA observations, their molecular gas fractions lie at least ~0.3-2dex below those expected for main-sequence
galaxies. Moreover, the molecular gas fractions of dormant
galaxies show no systematic difference from those of quiescent
galaxies in our sample. This indicates that dormant
galaxies have already reached a gas-poor state shortly after quenching and favors a scenario in which the rapid shutdown of
star formation is associated with depletion or removal of the molecular gas reservoir, rather than primarily with suppressed
star formation efficiency while retaining a substantial molecular gas reservoir. Our results suggest that, if no substantial gas replenishment occurs in the future, dormant
galaxies are more consistent with becoming "dead” than with being temporarily "sleeping”.
Read the full abstract ▾
All authors Mengyuan Xiao, Songnan Qi, Pascal A. Oesch, Longji Bing, Miroslava Dessauges-Zavadsky, Andreas L. Faisst, Mahsa Sanati, Rashmi Gottumukkala, Gabriel Brammer, David Elbaz, Robert Feldmann, Yoshinobu Fudamoto, Lucas C. Kimmig, Benjamin Magnelli, Rui Marques-Chaves, Anurag Amol Sawarkar, Sune Toft, Francesco Valentino, Christina C. Williams
Machine Learning (Statistics)
Convergence Rates of Gaussian Process Thompson Sampling for Efficient Global Optimization
Haowei Wang and Yilin Zheng
Efficient global optimization aims to identify a global optimum of an expensive black-box function within a limited evaluation budget $N$. We investigate the convergence rates of Gaussian process Thompson sampling (GP-TS) under noiseless observations, assuming a fixed objective function in a reproducing kernel
Hilbert space (RKHS). For Matérn kernels with smoothness parameter $\nu$ in dimension $d$, we establish a worst-case expected simple-regret upper bound of $O(N^{-\nu/d}\sqrt{\log N})$ for both discrete GP-TS on sufficiently fine grids and its continuous counterpart. For the continuous implementation, we develop a nested-grid argument to control posterior sample paths over the entire domain. We further establish matching lower bounds of $\Omega(N^{-5/2}\sqrt{\log N})$ for both implementations under a fixed one-dimensional Matérn-$5/2$ kernel, demonstrating that the logarithmic penalty is intrinsic to standard GP-TS in this setting. Motivated by this limitation, we propose Gaussian Process Capped Thompson Sampling (GP-CTS), which caps posterior sample paths using a valid RKHS norm upper bound while preserving optimism. This modification successfully eliminates the logarithmic penalty and achieves the minimax-optimal rate $\Theta(N^{-\nu/d})$ for both discrete and continuous implementations.
Read the full abstract ▾
Computational Engineering
Sounds Cheesy - Bridging the Texture Gap Between Dairy and Plant-Based cheese with Non-Contact Ultrasound and AI
Wenjing Feng, Thomas A. Hazlehurst and Nik Watson
Plant-based cheeses often fail to reproduce the texture of dairy cheese. Closing this gap is slow because the reference methods, texture profile analysis (TPA) and sensory panels, destroy the sample and are poorly suited to screening many formulations in a timely and cost-effective manner. Here we show that non-contact, air-coupled ultrasonic imaging combined with machine learning can characterise cheese texture rapidly and without damaging the sample. Thirteen commercial products (nine dairy, four plant-based; 119 slices, 357 scans) were imaged in through-transmission at 100 kHz to give two-channel amplitude and time-of-flight C-scans, and each slice was then measured by TPA. The C-scans revealed millimetre-scale acoustic domains in ripened dairy cheeses, consistent with their fused-curd microstructure, that were absent from the homogeneous starch and oil
matrix of the plant-based products. TPA showed that the texture gap is one of range and failure behaviour rather than average firmness. Mean hardness was similar for the two categories, but dairy products spanned a 7.5-fold range against 1.9-fold for plant-based products, and the plant-based products failed in a brittle, strain-softening manner rather than by ductile yielding. No single ultrasonic parameter tracked texture across products. However, a compact convolutional neural network predicted full true stress-strain loading and unloading curves from the C-scan images of unseen slices across all thirteen products without product-specific calibration. Clustering the predicted curves produced a texture map on which three plant-based slices sat with the Cheddars and one with Emmental, while soft, mozzarella-like textures remained unoccupied. The approach offers a one-minute, non-contact measurement suited to formulation screening and at-line quality control of cheese.
Read the full abstract ▾
Quantum Physics
Majorana Edge Modes Reshape Kibble-Zurek Scaling in Kitaev Honeycomb Ribbons
Hyesung Choi, Antonio Mandarino and Mahn-Soo Choi
Recent progress in programmable
quantum simulators motivates exact
benchmarks for nonequilibrium dynamics in topological
quantum matter. We study the Kitaev honeycomb model driven through its critical regions by a linear ramp of the bond coupling $J_z$ over a ramp time $\tau$, comparing zigzag and armchair ribbons with fully periodic tubes. The production of excitations exhibits a geometry-dependent finite-size crossover controlled by $N_2/\sqrt{\tau}$, the ratio of the ribbon width to the Kibble–Zurek length. Periodic tubes obey the expected critical-line Kibble–Zurek scaling and serve as the boundary-free
benchmark. Zigzag ribbons follow the same scaling until defect production is exponentially suppressed by a finite-width algebraic gap floor associated with the formation of Majorana edge modes, whose occupation is pinned by inversion parity. Armchair ribbons instead support transient Majorana edge modes with exponentially small gaps, producing width-parity-dependent defect plateaus: exactly one defect per unit length for odd widths and an approximate, transient plateau near one half for even widths. The quench-generated
entanglement entropy and fermionic logarithmic negativity inherit the same geometry-dependent structure, and the negativity further distinguishes the coherence-borne
entanglement of the Kibble–Zurek regime from the population-borne
entanglement of the edge-mode plateaus. These results identify boundary Majorana modes as the mechanism controlling departures from ideal Kibble–Zurek dynamics and provide geometry-resolved
benchmarks for programmable simulations of topological matter.
Read the full abstract ▾
Quantum Physics
Quantum Folding in Chebyshev-Based Quantum Subspace Diagonalization
Seung Park, Dongni Chen, Stefano Chesi et al. (+2 more)
Quantum subspace diagonalization (QSD) with a Krylov space spanned by Chebyshev
polynomials of the
Hamiltonian is a natural ground-state method for fault-tolerant
quantum computers, but every guarantee it offers—convergence rate, noise floor, and the cost of turning its classical output back into a physical state—is controlled by the overlap of the reference state with the true ground state, which decays exponentially with system size for any product state and non-trivial
Hamiltonian. We propose
quantum folding: for a system governed by a local
Hamiltonian, seed the QSD with the product of the ground states of its subsystems, which differs from the target only through the interface terms. We prove a nonasymptotic error bound for the Chebyshev QSD that is linear in the measurement noise, by transplanting Kirby's effective-
Hamiltonian analysis to the Chebyshev–Krylov basis, and identify the block-encoding rescaling of the
Hamiltonian as a parameter of the guarantee. Exact simulations of antiferromagnetic Heisenberg chains show that the folded reference state reduces the Krylov dimension for a given accuracy severalfold and lowers the error under noise by two to five orders of magnitude relative to the Néel state. We show how the Ritz vector is prepared as a physical state and that the ground-state generation probability obeys a tight ceiling set by the overlap, so that recursive folding prepares ground states at
polynomial cost. Taken together, these results allow for a scalable QSD framework that avoids the exponentially small ground-state overlaps associated with conventional product-state initialization. Finally we compute the fold fidelity itself: a slow power law for the gapless
spin-1/2 chain, a size-independent constant for the gapped
spin-$1$ chain, and an exact closed form converging to 1/4 for the AKLT chain.
Read the full abstract ▾
All authors Seung Park, Dongni Chen, Stefano Chesi, Kyunghyun Baek, Mahn-Soo Choi