Peer Revenue Impact factor: volatile

Papers

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

All of arXiv

Monday 9 November

Computational Engineering

Thermodynamically consistent neural constitutive models for elastoplasticity via eigen-representations

Neural network material models can represent yield surfaces and hardening laws with far more flexibility than classical closed-form plasticity formulations, but training such a model end-to-end through an implicit return map introduces non-trivial challenges. This work presents a thermodynamically consistent neural constitutive model for rate-independent, isotropic elastoplasticity that jointly learns a flexible yield function together with its evolution laws. For non-associative flow, the yield function extends the convex flow potential by a jointly trained additive term, which renders the model thermodynamically consistent by construction. Isotropy and convexity are enforced simultaneously across all learnable components through a permutation-invariant, spectral (eigenvalue-based) representation. Training proceeds through a backward-Euler return map, with gradients propagated by the implicit function theorem together with an eigenvalue-degeneracy-aware scheme for safe differentiation. The robustness and generality of the approach are demonstrated by reproducing known constitutive models, including non-associative flow, learning material behavior from stress–strain data extracted from a full-field finite element simulation, and capturing unknown non-associative behavior from homogenized representative volume elements of a foam. Finally, the proposed neural plasticity model is evaluated within finite element simulations to solve boundary-value-problem under plane stress.
Read the full abstract

All authors Harikrishnan Vijayakumaran, Shunyu Yin, Martin Abendroth, Miguel A. Bessa, Karl A. Kalina

Artificial Intelligence

Cognitive Multi-Agent Search and Rescue: Language-Grounded Hierarchical Reinforcement Learning for Collaborative Robotics

Effective search and rescue (SAR) operations require autonomous systems capable of coordinating multiple heterogeneous agents, adapting to dynamic hazards, and aligning with human mission intent. Existing language-guided reinforcement learning (RL) methods primarily focus on single-agent instruction following and lack mechanisms for team-level coordination in complex, partially observable environments. This paper presents a Cognitive Multi-Agent Framework based on Language-Grounded Hierarchical Reinforcement Learning (LG-HRL) that integrates natural-language understanding, hierarchical policy decomposition, and a shared cognitive workspace to achieve cohesive multi-agent behavior in SAR scenarios. Human instructions are interpreted by a large language model (LLM) and transformed into structured semantic cues that guide high-level strategy selection and role assignment across UAV and UGV agents. A formal Language-Grounded Multi-Agent Hierarchical Decision Process (LG-MAHDP) is introduced to model the interplay between semantic grounding, hierarchical control, and distributed coordination. Comprehensive experiments in a simulated disaster environment show that LG-HRL improves learning speed by 2.3 times, increases coordination coherence by 37%, and enhances subgoal accuracy and hazard avoidance by 46% and 31%, respectively. The system achieves 20-30% higher victim rescue rates, demonstrates emergent role specialization, and adapts immediately to human verbal updates. Ablation studies confirm that both the shared cognitive workspace and language grounding are critical to overall system performance. Overall, this work establishes a new paradigm for language-guided multi-agent SAR robotics, demonstrating that semantic grounding combined with hierarchical reasoning enables safer, more efficient, and more interpretable autonomous rescue operations.
Read the full abstract

Robotics

How Much of Your VLA Is kNN

Vision-language-action (VLA) models now exceed 95% success on the LIBERO simulation benchmark, yet it remains unclear how much of this performance requires generalization beyond the training demonstrations. In this work, we show that retrieving nearby demonstrations and reusing their action chunks achieves a success rate of 91.7 % on LIBERO without policy training. Under LIBERO-Plus perturbations, retrieval shows the smallest drop of all evaluated policies. Across nine publicly available VLAs, fine-tuning mostly makes the vision-language backbone a better retrieval key, and the actions of the VLAs remain close to the demonstrations. In addition, we introduce Local Action Consistency (LAC), an offline metric that measures whether nearby observations in representation space are also associated with similar future actions. We show that LAC tracks the success rate of retrieval and can guide the training of a retrieval key. Our evaluation and source code are available at https://github.com/sereact/lerobot_knn.
Read the full abstract

Metric Geometry

Euclidean filling inequalities in CAT(0) spaces

We clarify and streamline OpenAI's proof of the sharp Euclidean filling inequalities in CAT(0) spaces. We extend the result to complete spaces without assuming properness or compact support, and we establish rigidity in the equality case.

Plasma Physics

Background-free and laser-ionization minimized single-shot fs-EFISH in luminous pulsed plasmas

This work introduces spectrally tuned electric field induced second harmonic (t-EFISH) generation for substantially reducing or eliminating background plasma emission in single-shot, time-resolved, electric field measurements. The technique is demonstrated using a femtosecond 1240 nm pump and 620 nm signal detection to avoid spectral and temporal overlap with the early, intense onset of N2(C-B) emission in nanosecond atmospheric-pressure air spark discharges. This enhances signal-to-emission ratios pre- and post-breakdown, enables lower pump energies, suppresses detector saturation, and reduces the coefficient of variation in the electric field measurements by 8.5x. This strategy also provides a pathway for background-free measurements in other discharge environments with varying spectro-temporal emission characteristics. Moreover, the longer 1240 nm pump wavelength reduces multiphoton intrusiveness, directly improving measurement fidelity.
Read the full abstract

All authors Vidyadhara B. V. Kumar, Karna Patel, Aman Satija, Sally Bane, Terrence R. Meyer

Superconductivity

Mimicking the zero-field Josephson diode effect with remanent magnetization and self-field

A Josephson junction whose critical current is larger in one direction than in the other is a diode. With a magnetized barrier, a junction can be a diode at zero applied field, and this is commonly attributed to an intrinsic Josephson diode effect, in which the magnetic barrier breaks time-reversal symmetry. Here we show that the same signatures arise from two ordinary ingredients: the magnetic field of the bias current in the electrodes and the remanent magnetization of the barrier. In cross-type Nb/Pd$_{89}$Ni$_{11}$/Nb junctions, the self-field displaces the interference patterns of the two current polarities in opposite directions. An in-plane remanent magnetization translates the pattern and brings the resulting asymmetry to zero field, where the junction mimics a nonvolatile diode with an asymmetry of 10 to 12% and a polarity set by the magnetization direction. Reversing the current or the field cannot distinguish this mimic from an intrinsic effect, but the wiring can: moving the current lead to the other arm of the electrode across the field reverses the self-field diode, in applied field and at zero field alike, whereas an intrinsic diode would be unchanged. This gives a simple dc test for Josephson diodes in the current-perpendicular-to-plane geometry.
Read the full abstract

All authors Nathan Satchell, Arjun Sapkota, Pukar Sedai, Kurt Lorenzen, Juan Gomez

Quantum Physics

Long-range correlated quantum errors from diffuson-correlated dephasing

Electrical solid-state qubits are generally initialized, manipulated and read out through diffusive conductors. We show that the disorder-induced low energy modes of such a conductor — diffusons — impose spatially correlated dephasing on any qubits that share it, a mechanism we term diffuson-correlated dephasing. Starting from a Keldysh nonlinear sigma model for the disordered conductive region, coupled capacitively to two-level systems, we obtain the noise kernel in closed form. It decays as the inverse square of the qubit separation and changes sign with the orientation of the pair, so that qubit pairs at equal separation may be correlated or anticorrelated, an unambiguous experimental signature. In two dimensions, the correlated noise a single qubit shares with the rest of the array, in the worst-case scenario, grows logarithmically with array size rather than converging. The sufficient condition for fault tolerance under correlated noise is therefore violated, so standard threshold theorems no longer guarantee protection against this channel for large qubit arrays. As this dephasing originates from the soft-mode content of the environment rather than on any specific gate detail, we expect correlated dephasing channels of this kind wherever solid-state qubits or sensors share a gapless conductive medium.
Read the full abstract

Number Theory

The $\mathbb Q$-gonality of $X_1(42)$ is $12$

We prove that the $\mathbb F_5$-gonality and the $\mathbb Q$-gonality of the modular curve $X_1(42)$ are both $12$, improving the previous lower bound of $10$. Since $J_1(42)(\mathbb Q)$ is finite, it follows that $12$ is the least degree in which $X_1(42)$ has infinitely many points. The lower bound comes from an exhaustive search over Riemann–Roch spaces, following Derickx and van Hoeij, using a new reduction by the orbits of a cyclic group of automorphisms of order $12$ on places of degree at most $5$ ($743\,231$ Riemann–Roch spaces). An earlier search with the reductions from van Hoeij's code (about $3$ million Riemann–Roch spaces, about $6$ times more CPU time) gives the same result and serves as supporting evidence. We prove the reductions and record the data needed to reproduce the computations.
Read the full abstract

Quantum Physics

Rational quantum spectral filtering by repeat-until-success protocol

We introduce a finite-memory repeat-until-success (RUS) framework for implementing rational spectral filters on a quantum computer, extending spectral filtering beyond the polynomial transformations implemented by quantum signal processing (QSP) and quantum singular value transformation (QSVT). Our approach interleaves queries to a block encoding of a Hermitian operator $X$, whose spectrum lies in $[-1,1]$, and mid-circuit measurements that either accept, reject, or continue the computation without resetting the quantum state. Although the protocol may run for arbitrarily many rounds, its terminal statistics are determined by a finite-dimensional Lyapunov equation. Under uniform termination and phase-inversion symmetry, we characterize exactly the spectral filters $R(X)$ realizable upon acceptance: $R$ can be any real rational function with no poles on $[-1,1]$ and values in $[0,1]$. We provide a constructive compiler for any admissible $R$, with bounds on memory and expected query cost, and characterize the accepted quantum operation through a spectral coherence kernel. Applications to regularized resolvents and Lorentzian spectroscopy yield exact low-degree realizations and query-cost advantages over the fixed-depth QSVT polynomial-postselection baseline considered here, with all repetitions and independent restarts after rejection included in the cost.
Read the full abstract

Systems and Control

Fast Sampling or High Gain? iP versus PI Control

Intelligent proportional (iP) controllers, introduced in the framework of model-free control, are often presented as an alternative to proportional-integral (PI) controllers that requires no model of the plant. This paper studies this property for a first-order linear plant under sampled-data implementation with a zero-order hold. When the time derivative of the output is estimated by a backward difference, the discrete-time iP and PI controllers are known to obey the same control law under a one-to-one correspondence between their parameters. Building on this correspondence, the paper derives necessary and sufficient conditions for exponential stability of the iP closed loop, in closed form, in terms of the controller parameters KP and alpha, the sampling period h, and the plant parameters. It then shows that, when the iP parameters are held fixed and the sampling period tends to zero, the closed-loop decay rate tends to the iP gain KP, provided that KP is positive and that the ratio rho between the input gain of the plant and alpha satisfies 0 < rho < 2. One closed-loop pole then behaves as 1 - h KP, while the other tends to 1 - rho. With fixed PI gains, by contrast, the decay rate tends to that of the continuous-time PI loop, which depends on the plant parameters. The two parametrizations therefore suggest different tuning procedures, although they achieve the same closed loops at every fixed sampling period. High gain for the PI is, in some sense, replaced by fast sampling of the iP. The results also make explicit which plant information the iP controller requires: the sign of the input gain and a bound on its magnitude in the limit of fast sampling, and bounds on both plant parameters at a given sampling period. Numerical experiments illustrate these results.
Read the full abstract

High Energy Physics (Theory)

Nuclear geometry under external magnetic field in the Skyrme-Maxwell model

We find static solutions of the U(1) gauged Skyrme-Maxwell model in a constant external magnetic field, for baryon numbers $B \leq 6$. The field acts as an isospin-breaking source, stretching the skyrmion along its direction, while its size grows and then shrinks as the induction increases. We obtain the energy minimizers with a GPU implementation of the arrested Newton flow (ANF) algorithm. Gauged skyrmions have less energy and are more strongly bound than their ungauged counterpart. At weak magnetic field, the solutions are small deformations of the ungauged skyrmions, including regions of negative baryon density in sectors $B = 3, 5, 6$. For intermediate field, novel geometries appear, notably a $(B-1)$-toroidal chain. High field values lead to axially symmetric toroidal structures for all $B$.
Read the full abstract

All authors Christoph Adam, Andrés R. Anadón, Carlos Naya, Thomas Winyard

High Energy Physics (Theory)

General non-extremal rotating charged AdS black hole solutions with four independent charge parameters in $D=4$, STU gauged supergravity

We present an exact solution for generic non-extremal rotating, charged black holes with a negative cosmological constant in four-dimensional $\mathcal{N}=2, U(1)^4$ gauged supergravity. This general non-extremal solution contains seven independent parameters related to the mass, rotation, two electric and two magnetic charges, as well as a nonzero cosmological constant. It generalizes the four-dimensional Cvetič-Youm black hole solution to the gauged supergravity case. We anticipate that this is the only regular black hole solution in four-dimensional gauged STU model without any Misner string singularity and/or C-metric-like conical defect on the symmetry axes. We also investigate its thermodynamic property and give the Killing-Stäckel tensor responsible for the separability of Hamilton-Jacobi and massless Klein-Gordon equations. Various special cases related to the known solution are discussed, and its BPS limit is also analyzed. Our solution provides a new background to test the AdS$_4$/CFT$_3$ correspondence and to study the microscopic entropy via extremization of the BPS index.
Read the full abstract

Computer Vision

VIGOR: Benchmarking Visual Rationale Correctness in Multimodal Large Language Models

Multimodal large language models (MLLMs) are typically evaluated by whether they produce correct answers, but correctness alone does not reveal whether their outputs are supported by the right visual evidence. We introduce VIGOR, a benchmark for evaluating VIsually GrOunded Rationales in MLLMs. VIGOR represents visual rationales as explicit concepts paired with pixel-level groundings, covering multiple abstraction levels including low-level visual properties, intermediate structures, objects, parts and high-level motion, relations. To construct VIGOR at scale, we develop an automatic annotation framework that reuses existing dense supervision when available and generates missing concept groundings through proposal-based annotation followed by conservative multimodal verification. Using VIGOR, we evaluate whether MLLMs "look at what they say" by comparing model-derived visual evidence for each expressed concept against annotated ground-truth rationales. Our results show that current MLLMs exhibit limited visual rationale correctness and reveal distinct grounding failures across concept levels. VIGOR provides a benchmark and scalable data construction pipeline for studying whether MLLM outputs are supported by appropriate visual evidence. Code is available at https://github.com/deep-real/VIGOR
Read the full abstract

Classical Physics

Energy Dissipation Modulates Sensitivity to Corner Advantage in Weighted Dice

Weighted dice have pedagogical value and a range of broader implications. The physics of such dice provides an entry point for understanding state probabilities, center-of-mass calculations, and inertia tensors. Related considerations arise in the physics of grains, seed dispersal, erosion, and materials science, all of which involve components that can be compared with the stochastic tumbling and settling of dice. Previous empirical results are consistent with a Boltzmann-like distribution. We reproduce this result at scale using 23 classes of corner-loaded dice and more than 500,000 physical dice rolls collected with an automated rolling and imaging system. Briefly, landing probability increases exponentially with corner advantage (lower center-of-mass). We test the additional hypothesis that landing statistics depend on the combined effects of corner advantage, inertia tensor, and energy dissipation. Corner advantage defines the direction of the gravitational preference, but the rate of energy dissipation determines how strongly that preference is expressed in the final landing distribution. Simulations suggest that the inertia tensor and its orientation affect roll statistics; indeed, we find a small but statistically detectable departure from uniform landing probabilities under our experimental conditions. Our approach has pedagogical value as an entry-level engineering and data science project. The dice are 3D-printed and automatically rolled at scale using affordable parts and open-source tools. Results are compared to rigid-body simulations and a six-state model of energy dissipation.
Read the full abstract

All authors Nick Kinney, Karissa Boyce, Hallie Nelson, Stephen Loftus, Lem Curran, Joanna Friesen

Other Condensed Matter

After equilibration

`Complexity' measures the exploration of state space by the dynamics at a finer level of detail than that required for the equilibration of standard extensive observables , and therefore continues to grow long after these observables have equilibrated. In a recent paper, we proposed that this exploration may be quantified by the large deviations attained over time, or, equivalently, by the time taken for `rare' observables – such as, for example the exponential of the magnetization – to equilibrate. It turns out that the evolution of complexity can then be naturally mapped onto a standard model of glassy dynamics, the "Trap Model" , which exhibits aging and effective temperatures. Under this mapping, the exploration of increasingly fine details of phase-space corresponds to the discovery of progressively deeper minima in the glassy model.
Read the full abstract

Solar and Stellar Astrophysics

Identification of chromospheric dynamical signatures in the 2022 December 27 solar flare with DKIST

Spectroscopic observations by the Daniel K. Inouye Solar Telescope (DKIST) with unprecedented capabilities provide new insights into the evolving fine structure in the chromosphere during solar flares. We analyzed Ca II 8542Å spectra, formed in the middle chromosphere, of a minor flare observed by DKIST's Visible Spectropolarimeter in a single raster lasting 1.5 hours. We characterized the dynamic features in the spectra, then identified the spatial scales and context of these features during the evolution of flare heating. To do so, we developed a multi-layer, hierarchical clustering algorithm, with which 17,422 line profiles of flaring spectra were characterized into 19 clusters of similar spectral features. We then examined the spatial structure of the spectral clusters and also the time when the spectra were observed with respect to the local heating history inferred from observations by the Atmosphere Imaging Assembly. The spectral clusters form a set of spatially coherent structures, comprised of a central kernel with an enveloping halo. The smallest kernel size that is identified has a width of $\le 400$ km. Furthermore, spectral clusters in the pore and plage exhibit distinct properties such as kernel size, line shift, and spatial variation, which may be a result of different background atmosphere, and may also reflect different stage of local heating. High-resolution observations with improved cadence can further our understanding of the evolution of chromospheric fine structure.
Read the full abstract

All authors Sarah Riley, Jiong Qiu, Graham Kerr, Dana Longcope, Gianna Cauzzi

Physics and Society

Everything in its right place: attention is a radio head

Attention, the core of language models, is a digital calculation over tokens: an exponential selects them, a division normalizes and a positional code sets the order. We show that attention can instead emerge from the synchronization of a receiver with its context tokens. Each past token is an oscillator whose state carries its content and whose frequency drifts with its age. The present token selects them by content, aligning an antenna with its query, and by position, setting the frequency of a receiver with its dial. Solving Kuramoto's problem for this receiver yields an attention layer in closed form with two phases: above a critical coupling set by the context, the receiver locks and its weights normalize; below it, it stays silent and its weights sum to less than one. Trained networks use both phases to select, abstain and count, fail where the theory predicts and keep reading far beyond the training length. Each part of attention finds its place in one device, a radio head: it tunes in to what to hear and when, and stays silent when nothing relevant is on air.
Read the full abstract

Numerical Analysis

Convergence of an adaptive multilevel stochastic Galerkin FEM driven by a residual-based error estimator

In this paper we analyze a residual-based a posteriori error estimation strategy and adaptive multilevel algorithms for a class of parametric elliptic PDEs, where the diffusion coefficient depends on countably many uncertainty parameters. Under a saturation assumption on the parametric approximation space, we propose a computable residual-based error estimator. We also introduce the notion of K-compatibility of finite element meshes, which ensures that only the triangulations with comparable mesh sizes are used, and incorporate this concept into the adaptive algorithm. We prove that the estimator is reliable and satisfies K-dependent stability and reduction properties. These results underpin our convergence analysis of the adaptive multilevel algorithm: we show that the algorithm drives both the a posteriori error estimator and the true approximation error to zero. Numerical experiments illustrate and confirm the theoretical findings.
Read the full abstract

All authors Alex Bespalov, Maximilian Brunner, Alexander Freiszlinger, Dirk Praetorius

Artificial Intelligence

Teaching the Harness: Student-Aware Skill Transfer for LLM Agents

Effective teaching requires reasoning about the learner as well as the task, a capacity that Theory of Mind describes as inferring another agent's latent state from its behavior. We ask whether a stronger LLM teaches a fixed student better when it applies this principle. The student is an LLM agent acting through a harness that fixes its tools, context, and action budget, so a procedure that helps one model can exhaust the budget of another. We introduce Student-Aware Teaching (SAT). The teacher builds a learner model that pairs an empirical profile with hypotheses about the student's limitations, tests these hypotheses with diagnostic probes, and writes reusable skills for the diagnosed failures. A skill enters the library only after it raises success on matched development tasks. With Claude-Sonnet-5 teaching DeepSeek-V4-Flash, SAT reaches 86.46% task success across four primary runs, leading both a profile-free teacher with the same framework (81.04%) and the untaught student (75.00%) in each run. Removing the analytical learner model lowers both prediction and success, and under misleading learner profiles none of the 48 candidate skills passes admission. When DeepSeek-V4-Flash and Gemini-3.8-Flash exchange libraries, each performs better with its own. Saved libraries keep average gains of up to 11.1 points after the teacher leaves. The benefit depends on a student that can follow instructions and a deployment route that delivers them.
Read the full abstract

Chemical Physics

Autonomous Optimization of Numerical Algorithms by an AI Agent: Application to Atomistic Simulation

Coding agents built on large language models, run in automated improvement loops, can discover more efficient algorithms. We present AIOA, a framework that brings this approach to the numerical algorithms of atomistic simulation: an agent improves an algorithm against a pre-specified metric in a loop guarded by verifiers it cannot alter. Applied to ASE, a widely used atomistic simulation library, AIOA targets two facets of an algorithm: its implementation, rewritten to run faster with unchanged results, and the method itself, redesigned to converge with fewer force evaluations. Agents made six kernels 22-24 times faster on average in Python with NumPy and SciPy. On unseen problems, optimizers from the best configuration needed 3.1-3.4 times fewer force evaluations than L-BFGS for structure relaxation and 4.6-17 times fewer than standard climbing-image nudged elastic band calculations for reaction paths. The agents' optimizers also contain two algorithmic components we did not find in the literature.
Read the full abstract

Robotics

DOBB: Dynamic Oriented Bounding Boxes for Multi-Agent Collision Avoidance under State Uncertainty and Heterogeneous Dynamics

Learning-based decentralized collision avoidance must remain reliable under state-estimation uncertainty and heterogeneous agent dynamics. Existing methods typically address these factors through training-domain variation or by exposing uncertainty and capability estimates to the policy at runtime. However, both leave the policy to infer how these quantities should alter the physical separation required for safe interaction, which can couple collision-avoidance behavior to relationships learned from the training distribution. We introduce the Dynamic Oriented Bounding Box (DOBB), an online geometric representation that maps the current state estimate, uncertainty, and agent-specific response characteristics into a dynamic, direction-dependent safety margin. We derive geometric features that jointly characterize the relative configuration and severity of interactions between agents' DOBBs, along with a reward function that preserves the physical interpretation of those interactions. Across four evaluation environments, DOBB achieves the lowest mean total failure rate of 2.6%. This performance persists under out-of-distribution shifts in actuation dynamics and state-estimation uncertainty, with collision-based failure remaining at or below 2.3% in every tested condition. We attribute this robustness to the geometric abstraction of the safety margin, which is less coupled to the specific uncertainty and actuation regimes encountered during training.
Read the full abstract

Statistical Mechanics

Dynamics of Rényi entanglement entropies and Eigenstate Thermalization

How is entanglement dynamics encoded in the eigenstates of the time-evolution operator? We address this question by focusing on Rényi entropy dynamics in deterministic chaotic Floquet models in arbitrary dimensions. Inspired by the eigenstate thermalization hypothesis (ETH), we employ spectral representations of reduced density matrix moments (purity and its higher-order generalizations) whose decay through spectral dephasing captures nonequilibrium entanglement growth. These representations closely resemble those governing local-observable dynamics, but involve eigenstates in a replicated Hilbert space and observables are replaced by subsystem permutation operators, such as the swap operator for Rényi-2. Focusing on Rényi-2, we combine exact random matrix theory calculations, established ETH results and numerical tests in 1D to formulate ETH-like ansätze for subsystem swap matrix elements in the replicated eigenbasis and their cross-correlations with initial-state overlap statistics. In the thermodynamic limit, purity dynamics is entirely determined by smooth slowly varying functions that emerge from such statistics, paralleling generalized ETH formulations in a single replica. In the limit of large subsystems, these functions exhibit a striking structural simplification: locality of the dynamics together with bulk spectral dephasing reduce their functional dependence to a single fundamental frequency. Our formalism provides a setting where ETH-like statements naturally coexist for both arbitrarily small and large subsystem sizes. We discuss extensions of our results to higher Rényi entropies. Beyond clarifying the spectral foundations of entanglement dynamics, these results lay some of the groundwork for future applications of spectral filter methods to the computation of Rényi evolution in deterministic systems.
Read the full abstract

Artificial Intelligence

Agreeing on the Answer but Not the Reason: Measurement Limits of LLM Reasoning Failure Taxonomies

Failure taxonomies for LLM reasoning are increasingly reported as measurements: papers state what fraction of a model's errors fall into named categories, then compare those fractions across models. We show that for one such category the reported fraction is largely a property of the measurement procedure. We defined Compositional Collapse (a model notices a required sub-part of a problem, then drops or misapplies it) and measured it across twelve models and eleven languages on a public multilingual math benchmark (2,316 graded records). First, 50.1% of failures carry evidence for more than one category, so the reported plurality depends on the order in which the classifier tests them: reversing two tests moves it in eleven of twelve models, and across all six orderings Collapse's share spans 26.4% to 76.4%. Second, agreement collapses for the category but not for the task. Native speakers of all eleven languages judged 179 transcripts, agreeing with our inspector at kappa = 0.932 on answer correctness but only kappa = 0.093 on Collapse. A second native speaker per language re-rated 142 items; on 121 determinate pairs the two humans agree with each other at kappa = 0.659, while each agrees with the inspector at kappa <= 0.094. Re-running the same inspector model with the human instructions verbatim raises kappa against Human 2 from 0.147 to 0.636, so the gap is largely attributable to the prompt, not the model. Third, the inspector labels Collapse 1.81 times as often as native speakers, but only 1.17 times once a defective benchmark item (a deleted clause in the Bengali and Chinese translations) is excluded. We recommend four cheap checks before comparing failure-mode rates across models: report label co-occurrence, permute the tie-break, validate on a shared independent sample, and audit systematic disagreements.
Read the full abstract

High Energy Physics (Phenomenology)

Theory overview on LLP searches at the LHC

In these proceedings, we discuss heavy neutral leptons (HNLs) as well-motivated long-lived particle (LLP) candidates to illustrate the rich phenomenology that LLP searches can probe at the LHC. We briefly discuss the minimal HNL scenario, in which the phenomenology is governed by active-sterile neutrino mixing, and its characteristic displaced signatures. We then consider more general BSM scenarios with additional HNL interactions, showing how these can modify the main production and decay mechanisms, opening up new signatures beyond those of the minimal benchmark.
Read the full abstract

Databases

Answering Conjunctive Queries over Linear Existential Rules with Transitivity

We study Conjunctive Query Entailment Problem over ontologies con- sisting of Linear Tuple Generating Dependencies and transitivity asser- tions. A special case of this problem has been shown to be decidable in [1]. In particular, it covers the case of linear single head TGDs with bi- nary signature. In This paper we show that The problem for linear multi head TGDs with binary signature is Ackermanian hard
Read the full abstract

Machine Learning

Byzantine-Robust Muon: Resilient Momentum Orthogonalization for Distributed Learning

Optimization methods, including the Muon optimizer based on the Linear Minimization Oracle (LMO) and momentum orthogonalization, have shown strong empirical performance in large-scale and distributed learning, yet their robustness to adversarial behavior and data heterogeneity is not well understood. We study these methods within a unified first-order trust-region framework and analyze Muon-based updates with robust aggregation for Byzantine-robust distributed and federated learning. Our analysis does not rely on the i.i.d. assumption and introduces a non-Euclidean notion of heterogeneity consistent with the trust-region geometry. Under this framework, we establish convergence guarantees for both general non-convex and star-convex objectives. Our results provide a theoretical characterization of Byzantine-robust Muon-based optimization, with an additive heterogeneity error floor consistent with known impossibility results, under explicit aggregation assumptions. Under an explicit additive oracle-error condition, the guarantees also cover inexact orthogonalization, with the non-convex Byzantine terms in the bound's numerator unchanged but a smaller progress denominator. In simulated federated training with ResNet-18 and ViT-Tiny on CIFAR-10, additional image benchmarks, and a 38M-parameter byte-level language model on TinyStories, the tested Muon configuration attains higher attacked accuracy (lower perplexity) than the tested SGD and normalized-SGD configurations, and than sign updates on CIFAR-10. Fixed-honest controls further show a substantial RFA advantage over averaging under a bounded RFA-query attack, higher attacked accuracy for Muon-NS than tensorwise Euclidean normalization from shared checkpoints, and lower held-out TinyStories perplexity than a tuned AdamW baseline.
Read the full abstract

Solar and Stellar Astrophysics

Near-Ultraviolet and Optical Observations of the White Dwarf ZTF J0328$-$1219 Reveal Variable Periodicity in its Transiting Debris System

We observed the transiting debris disk system around the white dwarf ZTF J0328$-$1219 using the Hubble Space Telescope (HST) in 2025 October, obtaining near-ultraviolet (NUV) spectroscopy with STIS over five consecutive HST orbits. Our abundance analysis supports the accretion of a chondritic body. We complemented our HST program with 16 days of optical time-series photometry beginning in 2025 October, which reveal variable periodicity and rapidly evolving transit morphology. Previous monitoring by TESS during Sectors 4 and 31 confirmed a predominant, stable transit orbital period, $P_{\rm A} = 9.937$ hr, along with a second, transient period, $P_{\rm B} = 11.2$ hr. Our 2025 autumn light curves show transits are dominated by $P_{\rm B} = 11.17$ hr with $P_{\rm A}$ initially appearing as a weaker, secondary feature and eventually returning to predominance. We also identify a candidate transit period of 7.64 hr during the final days of the 2025 autumn campaign, although its significance and persistence remain uncertain. We add simultaneous $ugr$ light curves collected with ULTRACAM in 2024 November that are dominated by $P_{\rm A}$ and show no significant color term. We further compare the transit depths in optical photometry concurrent with our NUV observations, again measuring a lack of transit-dependent extinction except during the fifth HST orbit where a reddening term may arise. The observed variable periodicity and potentially changing color behavior illuminate ongoing dynamical processes, namely collisions, and demonstrate the opportunity for studying dust physics in real time at white dwarf debris disks.
Read the full abstract

All authors Joseph A. Guidry, J. J. Hermes, Zachary P. Vanderbosch, Boris Gaensicke, Simon Blouin, Laura K. Rogers, Jay Farihi, Stefan Arseneau, Lou Baya Ould Rouis, Siyi Xu, Bruce Gary, Thomas G. Kaye, Yarin-Meir Shani, Hannah Worters, Kaylee E. Grace, Steven Z. Savery, J. L. Provencal, Chun-Hai Bai, Keaton J. Bell, Barbara G. Castanheira, Paul Chote, Millie A. Dill, Inajara Endl, Torben Endl, Lester Fox Machado, Elizabeth M. Green, Kim Green, Easton J. Honaker, Nikoo Hosseininezhad, Xiao-Jun Jiang, Momin Y. Khan, Morgan A. Mitchell, Michael H. Montgomery, Rudy Morales, Tali Natan, Hui-Juan Wang, Jia-Qi Wang, Tu-Hong Zhong

Machine Learning

Which Successes Make Good References? Action-Sequence Alignment for Multi-Turn On-Policy Distillation

On-policy distillation (OPD) reduces train-inference distribution mismatch by supervising a student on its own trajectories. For multi-turn tool-use agents, conditioning the teacher on a validated same-task successful trajectory as a reference can provide useful hindsight, but different successes may reach incompatible interaction states and offer misleading supervision. We cast successful-reference selection as a free-target optimal transport problem, which reduces to assigning each failed trajectory its most compatible successful reference. Because latent state paths are hidden, we use action-sequence alignment as an observable proxy for compatibility and instantiate it with longest common subsequence (LCS), which identifies a shared ordered backbone of actions while allowing recoverable detours. This perspective makes trajectory compatibility the central criterion for reference selection in multi-turn OPD. Experiments on ALFWorld and the airline domain of $\tau^3$-bench show that LCS is the only evaluated reference selector to improve over reference-free OPD on both ALFWorld overall scores and all three airline metrics. Without further optimization, the airline-trained models also transfer to the retail domain, where LCS improves all three metrics.
Read the full abstract

All authors Mufei Li, Nikki Lijing Kuang, Zhuoer Wang, Samson Koelle, Jie Yuan, Manqing Mao, Hong Wang, Yanjun Lin, James Feng, Daniel Edmiston, Pan Li, Wei Niu

Data Structures and Algorithms

A Fast Algorithm for Subspace Convolution

A cornerstone result in Boolean lattice theory is the subset-zeta transform, which computes the $f$-weighted subset summation problem $(Z f)(B)=\sum_{A\subseteq B}f(A)$ for all $ B\subseteq[n]$ in $O(N \log N)$ additions instead of the naïve $O(3^n) = O(N^{\log_2 3})$, where $N=2^n$. This paper studies the algebraic analogue of this problem–the subspace summation problem–where subsets are replaced by linear subspaces in $\mathbb{F}_q^n$. Surprisingly, we show that this problem admits a similar $\Theta_q(N_q \log N_q)$-time algorithm, where $N_q$ is the number of subspaces. This improves substantially (in both theory and practice) over lattice-based algorithms ([BHK+16, KKW16]), which are known to be optimal for addition-only algorithms over $\mathbb R$. We prove similar runtimes for inversion, rank-constrained sums and span convolution. We describe two main applications: Fast linear algebra on Grassmann graphs: Matrix–vector multiplication for arbitrary distance-invariant kernels on $N$-vertex Grassmann graphs in $O_q(N\log N)$ arithmetic operations, together with fast spectral projections and exact solutions of the corresponding linear systems. Graph coloring of low binary cut-rank graphs: An $n^{2^{r^2/4+O(1)}}$-time algorithm for computing the chromatic number and chromatic polynomial of an $n$-vertex graph, given a rank-decomposition of width $r$. We also give quantitative reductions from satisfiability and monotone constraint problems. Experiments show speedups in measured running time of up to $18$ over the covering-pair algorithm for the zeta transform, $190$ over Poisson summation for rank-constrained sums, and $22$ over the distance-matrix recurrence for Grassmann kernel multiplication. Lean 4 formalizations of the transform identities, arithmetic operation bounds, and graph-coloring constructions accompany the paper [PW26].
Read the full abstract

Group Theory

Profinite groups with finite conjugacy classes of commutators

It was shown in a recent work by Senise and Shumyatsky that if $G$ is a finite group such that $|x^G|\leq m$ whenever $x$ is a commutator of prime power order, then $G''$ has $m$-bounded order. In this paper, we examine the following question: given a profinite group $G$ such that $|x^G|$ is finite whenever $x$ is a commutator of prime power order, does it follow that $G''$ is finite? We show that, in general, the answer to this is negative. The main result obtained in this paper is the following theorem: Let $G$ be a profinite group in which $|x^G|$ is finite whenever $x$ is a commutator of prime power order. Assume that $\pi(G)$ is finite. Then: (1) The second commutator subgroup $G''$ is finite; (2) $G$ has an open normal subgroup of nilpotency class at most $3$.
Read the full abstract