Peer Revenue Now accepting positions

Papers

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

All of arXiv

Friday 6 November

Numerical Analysis

On the convergence of Lattice Boltzmann method for advection-diffusion equation with intermediate scaling

The purpose of this article is to study the convergence of a lattice Boltzmann method (LBM) for a simple diffusion equation. The LBM is proven to be stable and consistent in some sense, so that convergence results may be obtained. The approach is similar to asymptotic analysis in the literature, although the schemes are also studied for intermediate time scalings. Usually, LBM are used in the range of acoustic regime, where $\delta$t ___ $\delta$x, or in diffusive scaling, with $\delta$t ___ $\delta$x 2 . In this work, we study any intermediate scaling $\delta$t ___ $\delta$x 1+r with r $\in$]0; 1]. The goal is to prove that this scaling is capable of approximating an advection-diffusion equation, provided the relaxation rate is chosen in accordance with $\delta$x. Convergence results are established and match the previously known results in the diffusive regime. The second objective is to prove that several choices for the initial condition are suitable for convergence. In particular, the usual choice of equilibrium distribution at initial time guarantees that the LBM scheme yields an accurate approximation of the macroscopic quantities.
Read the full abstract

Game Theory

Attaining Optimal Values under Sure Constraints

A game may have, for each target below its value, a finite controller meeting that target, yet no finite controller attaining the value itself. We study finite stochastic games with a hard specification that must hold on every supported execution, including paths of probability zero. Max uses a pure finite controller; Min may use any pure public-history strategy. We distinguish three numerical questions: can a supplied target be met, can the unspecified optimum be attained, and what is one exact value? For explicit hard and Boolean soft Streett conditions, the strict and non-strict supplied-threshold problems are coNP-complete, attainment of the unsupplied supremum is $\mathrm P^{\mathrm{NP}}$-complete, and one exact reduced fraction is $\mathrm{FP}^{\mathrm{NP}}$-complete under polynomial metric reductions. The proofs preserve both the value and its finite attainability when fixing a single opponent, and encode adaptive queries without computing unknown subgame values. The same construction yields matching value-comparison and exact-output lower bounds for ordinary stochastic Streett games with unrestricted randomized history strategies. We also bound the memory needed for optimal finite implementations for arbitrary Muller hard conditions and rational rewards in $[0,1]$ determined by infinitely recurring states. One controller is sure from the entire original sure region and attains at every finitely attained start, with no attainment premise on the other states. For $n$ states in that region it uses $2^{O(n\log(n+1))}Q(1+\log R)$ modes, where $Q$ is the product of the least common denominators of full original chance rows sourced in that sure region and $R$ clears all reward values. Its joint dependence on $Q$ and $R$ is worst-case sharp at each fixed $n\ge6$ when $Q,R\ge2$. Optimal dependence on arena size and efficient general synthesis are not resolved here.
Read the full abstract

High Energy Physics (Phenomenology)

Roles of $f_0(980)$ and $a_0(980)$ in the process $\Lambda_c^+ \to pK_S^0K_S^0$: a clean probe of the $K\bar K$ final-state interaction

Based on the Belle measurement of $\Lambda_c^+ \to pK_S^0K_S^0$, we study the contributions of the light scalar mesons $f_0(980)$ and $a_0(980)$, dynamically generated from the $S$-wave pseudoscalar meson–pseudoscalar meson interaction within the chiral unitary approach, to the threshold enhancement in the $K_S^0K_S^0$ invariant mass squared distribution, and examine the possible contribution of the intermediate $\Sigma(1620)$ resonance in the $pK_S^0$ subsystem. We first argue that the $\Lambda_c^+ \to pK_S^0K_S^0$ channel is a cleaner probe of the $K\bar K$ final-state interaction than the charged channel $\Lambda_c^+ \to pK^+K^-$, since the $\phi$ and $\Lambda^*$ contributions are forbidden and the $\Sigma(1620)$ and $\Sigma(1660)$ couple weakly to the $\bar{K}N$ channel. The threshold enhancement in the $K_S^0K_S^0$ invariant mass squared distribution is well described by the $S$-wave $K^0\bar K^0$ final-state interaction alone, with the $f_0(980)$ contribution dominating while that of $a_0(980)$ is small, supporting the interpretation of $f_0(980)$ as a $K\bar K$ hadronic molecular state. Future high-precision measurements by Belle~II and the proposed Super Tau-Charm Facility will help test the roles of the $f_0(980)$ and $a_0(980)$ and of the possible $\Sigma(1620)$ contribution in this process.
Read the full abstract

All authors Li-Na Zhou, Le-Le Wei, Li-Juan Liu, En Wang

Atomic Physics

Independent nuclear charge radius determinations: bound-electron $g$ factor versus muonic atom spectroscopy

We present an alternative method to determine absolute nuclear charge radii competitively via the bound electron $g$ factor of H-like ions. The method is demonstrated for $^{118}$Sn by extracting its radius with a precision of 0.1%. This is the second most precise radius ever determined from an electronic system. We benchmark our result against a state-of-the-art value derived from the spectrum of muonic $^{118}$Sn. This comparison reveals a 3.2$\sigma$ tension, prompting a careful reexamination of the calculations and calling for additional measurements. Crucially, after decades, we now have two independent methods to measure absolute charge radii with similar precision.
Read the full abstract

All authors F. Heiße, I. A. Valuev, J. Solanki, V. A. Yerokhin, B. Ohayon, N. S. Oreshkina

High Energy Physics (Theory)

Pfaffian sensitivity to temporal deformations of an N=5 background in the Lorentzian IIB matrix model

How spacetime dimensionality and anisotropy emerge from matrix degrees of freedom is a central question in the IIB matrix model. Our previous work constructed a family of constrained classical solutions at $N=5$, including a reference configuration with vanishing temporal extent and a pre-existing $3+6$ spatial hierarchy. Here we examine how the fermionic Pfaffian responds to temporal-matrix deformations of this configuration. The background geometry determines a distinguished real two-plane of temporal deformations. Numerical analysis of the full fermionic operator shows that the sensitivity of its lowest positive-eigenvalue doublet is strongly concentrated in this plane and dominates the first-order Pfaffian response. Under stated reflection-covariance and nonsingularity conditions, the other local one-loop contributions cancel at first order when combined, so the gradients of the logarithmic magnitudes of the total local one-loop weight and the Pfaffian agree. We prescribe a temporal deformation along the projection of this gradient onto the two-plane and determine the remaining matrix components by conditional classical stationarity. The existing spatial hierarchy is modestly strengthened at quadratic order, while the spectrum changes little at the small amplitude considered. This quadratic response is independent of orientation within the plane, although the first-order Pfaffian response distinguishes such orientations. These results connect background geometry, fermionic spectral sensitivity, and conditional classical spatial anisotropy.
Read the full abstract

Artificial Intelligence

MM-DyGraph: A Dataset, Benchmark, and Model for Multimodal Dynamic Graphs

Multimodal dynamic graphs are pervasive in real-world applications, where nodes and edges are accompanied by multimodal signals, including text and images. Despite their practical importance, benchmarks and methods that explicitly study multimodal learning in dynamic graphs remain scarce, hindering the full understanding and effective utilization of multimodal information for downstream tasks. To address this gap, we introduce MM-DyGraph, to the best of our knowledge, the first benchmark for multimodal dynamic graph learning, consisting of diverse real-world datasets with fine-grained temporal graph evolution and rich multimodal attributes. MM-DyGraph provides three evaluation tasks to comprehensively assess model performance in dynamic settings. We benchmark representative state-of-the-art dynamic graph methods and reveal that naive feature concatenation yields target-level inconsistencies: multimodal inputs benefit certain prediction targets while harming others. This observation indicates that the modality informativeness depends on the specific prediction target over time. Building on these observations, we propose a Target-aware Spatial-Temporal Multimodal Fusion (TST-MF) framework, which jointly conditions multimodal fusion on the prediction target and spatial-temporal relations, enabling context-adaptive cross-modal interplay for target-specific representations. Extensive experiments over diverse datasets and tasks show that TST-MF delivers robust and competitive performance, achieving clear gains in multimodal settings.
Read the full abstract

All authors Zongyuan Wu, Pengjie Wang, Lingshan Chen, Tianhang Wan, Yuan Meng, Xin Wang, Ling Feng, Wenwu Zhu

Computation and Language

DUET: Dual-Sparse Inference for Block Diffusion Language Models

Diffusion language models (DLMs) enable parallel token generation through iterative denoising, yet their inference remains costly due to repeated attention over a growing historical prefix and repeated processing of the entire current generation block. Historical attention becomes increasingly expensive as generation progresses, while all current-block positions are propagated through the Transformer at every denoising step even though only a small fraction are typically resolved at that step. In this paper, we introduce Duet, a training-free dual-sparse inference framework that addresses these challenges through prefix Key-Value (KV) sparsity and early query sparsity, respectively. Specifically, Duet exploits the concentration and cross-step stability of historical attention within each generation block to construct a compact historical KV subset that is reused across denoising steps. For query sparsity, Duet predicts update positions from shallow representations and propagates only the corresponding queries through the suffix Transformer layers, avoiding deep computation at positions unlikely to be updated in the current step. Experiments show that Duet maintains competitive accuracy while achieving 1.34$\times$/2.83$\times$ end-to-end/per-step speedups on LLaDA2.1-mini and 2.41$\times$/4.72$\times$ on SDAR-8B-Chat-b32 at a 32K context length.
Read the full abstract

All authors Siyuan Yan, Shen Gu, Xiaoxing Ma, Chun Cao, Jingwei Xu

Graphics

Audiovisual Disparities in VR: Impact on Spatial Perception

Virtual reality (VR) experiences often leverage rich and spatialized multimodal environments to increase immersion and engagement. This demands a consistent spatial perception of audiovisual stimuli, since perceived discrepancies can disrupt the sense of presence. In this work, we investigate the consequences of two types of spatial audiovisual disparities: true disparity, where there is a measurable spatial offset between auditory and visual cues, and perceptual disparity, where users report misalignment despite cues being colocated. Unlike most previous studies that employed controlled but simplified experimental setups, our research focuses on complex, realistic VR environments, allowing us to assess the actual implications for VR content design. Our experiments indicate that users are highly sensitive to true audiovisual disparities in controlled environments, detecting even minor misalignments. However, when engaged in additional tasks within realistic settings, their ability to notice such discrepancies diminishes significantly. We also observed that previously found perceptual disparities persist in complex audiovisual environments. However, we identify self-initiated head rotations as a key factor; its absence prevents the effect entirely. We hope our findings offer practical insights for designing more immersive and perceptually coherent VR experiences.
Read the full abstract

All authors Edurne Bernal-Berdun, Mateo Vallejo, Qi Sun, Ana Serrano, Diego Gutierrez

Algebraic Geometry

The Woo-Yong conjecture

We prove the Woo-Yong conjecture: the non-Gorenstein locus of a type A Schubert variety is the union of its generically non-Gorenstein maximal singular components. More strongly, at every scheme point, local Weil divisor classes are detected at the generic points of the maximal singular components containing that point. For arbitrary finite crystallographic type, the analogous localization map is injective on the torsion subgroup of the source local class group. This determines Cartier indices of locally $\mathbb{Q}$-Cartier divisors and gives the Woo-Yong criterion on the $\mathbb{Q}$-Gorenstein locus. A type D family shows the necessity of this hypothesis: the ambient variety is generically Gorenstein along its singular locus but not $\mathbb{Q}$-Gorenstein along a deeper nonempty locus. The type A proof and the counterexamples are geometric; the general torsion proof also uses bounded finite inputs with explicit certificates and reproducible reconstruction programs.
Read the full abstract

General Relativity and Quantum Cosmology

Axion cloud survival in nuclear star clusters

Stellar perturbations can ionize a superradiant boson cloud while transferring energy and angular momentum from the perturbing orbits. We follow the cloud after superradiant saturation by coupling its occupation to a two-dimensional, orbit-averaged Fokker–Planck equation for the stars. The evolution includes stellar relaxation, capture, orbital backreaction, and gravitational-wave (GW) emission from both the cloud and the stellar orbits. We evolve 24 models with black hole masses $M=10^5$–$10^7M_{\odot}$ and couplings $0.04\leq\alpha\leq0.40$. Stars with semimajor axes of a few Bohr radii, $r_c=M/\alpha^2$, dominate ionization. Stellar replenishment from the outer cluster sustains ionization over hundreds of Myr, particularly for extended clouds. Of the 24 models, 21 lose at least $99\%$ of their initial cloud mass. For compact clouds, intrinsic GW emission dominates the depletion.
Read the full abstract

Machine Learning

Controllable Molecule Generation via Sparse Representation Editing: An Interpretability-Driven Perspective

Controllable molecule generation is crucial for diverse scientific applications, such as drug discovery and materials design. While large language models (LLMs) show great promise, their dense and entangled representations impede precise control over the generation of molecules with bespoke substructures or properties. To address this, we propose Sparse Representation Editing (SpaRE), an interpretability-driven framework for fine-grained and precise control in LLM-based molecule generation. The crux of SpaRE is to learn an overcomplete sparse feature space that disentangles LLM representations into a compact set of latent features corresponding to chemically meaningful concepts. Within this space, we can directly manipulate these concept-aligned latent features to achieve (1) local control, by generating target atoms and functional groups at specified positions; and (2) global control, by customizing the overall structural and physicochemical properties within defined ranges. In this way, our framework advances interpretability from post-hoc analysis to actionable generative control. Experiments show that SpaRE can generate chemically desirable molecules under complex constraints in real-world scenarios, while offering mechanistic insights for quantitative structure-property analysis. The code and demo are available at https://github.com/WanyuGroup/ICML2026_SpaRE.
Read the full abstract

All authors Zhuoran Li, Xu Sun, Chang Wen Chen, Wanyu Lin

Software Engineering

Integrating MBSE Principles with AI to Improve the Quality of Requirements in SMEs Context

Model-Based Systems Engineering (MBSE) is a promising approach to improve systems engineering practices, particularly by enhancing requirement management, traceability, and early validation and verification (V&V). However, its adoption remains challenging, especially for small and medium-sized enterprises (SMEs), which often lack structured processes and solutions adapted to their scale and constraints. Our goal is to propose a lightweight MBSE methodology tailored to the needs and constraints of SMEs. This paper focuses specifically on requirements refinement, traceability, and early V&V. The approach integrates MBSE principles with artificial intelligence-assisted analysis to improve requirements quality while remaining compatible with existing tools and practices. A proof of concept is developed and applied in an industrial case study to evaluate the feasibility of the approach. The results demonstrate how early V&V activities can be supported within an SME context, while improving traceability across engineering artifacts. These findings highlight the potential of a lightweight, integrated MBSE approach to support the progressive adoption of MBSE in SMEs, thereby improving overall engineering efficiency.
Read the full abstract

Machine Learning

Cross-Layer TopK: Dynamically Allocating Sparsity Across Layers in Transcoders

Cross-Layer Transcoders (CLTs) decompose the activations of large language models into human-interpretable features and trace how these features interact across layers to form circuits, making them a key tool for mechanistic interpretability. However, training CLTs effectively and efficiently is hard. Beyond the usual trade-off between sparsity and reconstruction quality, a CLT has to model all layers at once, which adds a further challenge. The common JumpReLU recipe controls sparsity only indirectly through a penalty, while the direct extension of TopK selects a fixed number of latents in every layer. Yet reconstruction difficulty differs widely across layers, and intermediate layers are much harder than others, so giving every layer the same budget is suboptimal. We propose Cross-Layer TopK, which sets a single L0 budget for the whole model and dynamically allocates it across layers during training. In a comparison with JumpReLU and other TopK variants on Qwen3-8B-Base, this gives lower FVU while using 20% fewer active latents than JumpReLU. The better reconstruction carries over to circuit tracing, yielding more complete circuits with higher replacement and completeness scores. The fixed per-token budget of Cross-Layer TopK also makes it easy to pair with sparse decoding, which cuts both computation and communication and gives a 2.9x speedup per training step. We further show that Cross-Layer TopK follows scaling laws in both latent size and sparsity. We have open-sourced the weights of our CLT at https://huggingface.co/ByteDance-Seed/Cross-Layer-TopK-CLT.
Read the full abstract

All authors Anyuan Zhuo, Shaohua Zhang, Tianle Li, Yuan Lin, Diandian Gu, Jun Wang, Pinyan Lu

Artificial Intelligence

RAG-Enhanced LLMs for Interactive Explainability in Clinical Decision Support Systems

Artificial Intelligence (AI) is increasingly embedded in high-risks decision-making systems across domains such as healthcare, finance, and the judicial system. As reliance on predictive models grows, so does the demand for Explainable AI (XAI) to ensure transparency, trust, and interpretability. However, existing XAI approaches often fail to present explanations in a format accessible to domain experts, limiting their practical utility. Large Language Models (LLMs) have recently emerged as a bridge between complex models and end-users, offering natural language explanations. Yet, their use introduces new risks, including hallucinated outputs and lack of source traceability. To address these challenges, we propose a Clinical Decision Support System (CDSS) that not only combines standard predictive models with XAI techniques but also incorporates a Retrieval-Augmented Generation module as an interactive enhancer. This module grounds explanations in verified medical knowledge, reformulates technical outputs into a human-like format, and enables users-such as clinicians-to actively query the system. By fostering an interactive, human-in-the-loop environment, our approach empowers domain experts to explore model decisions, contextualize explanations, and build trust in AI-assisted diagnostics. The study concludes with preliminary experiments validating the proposed methodology.
Read the full abstract

All authors Lorenzo Mannocci, Francesca Naretto, Lucia Passaro, Anna Monreale

Statistical Methodology

Power improvements and exact shortcuts for closed permutation maxT testing with discrete data

In permutation tests with discrete data, especially in small samples, different permutations often yield the same test statistic. The resulting ties can make the test conservative and reduce power. This conservatism can be particularly pronounced for maximum statistics: permutations may have the same maximum but differ at other endpoints, so the remaining endpoint statistics are discarded. We develop two power improvements for closed permutation maxT tests by refining the treatment of ties. The first method uses a single uniform random variable shared across all local tests in the closed testing procedure. This shared-uniform construction maximizes the rejection probability of every elementary hypothesis among all ways of coordinating the same randomized local tests. It also preserves an exact linear step-down shortcut. The second method is the deterministic two-endpoint augmentation (tea) test, which uses the second-largest endpoint statistic to resolve ties in the maximum. This additional information can favour rejection under alternatives while preserving null validity, without auxiliary randomization. Its full closure admits an exact quadratic shortcut. More generally, augmentation using any fixed number of endpoint statistics admits an exact polynomial shortcut. Both procedures retain every rejection made by classical closed maxT procedures and provide strong finite-sample control of the family-wise error rate under intersection-null permutation invariance. In simulations with sparse binary outcomes, both methods improve power over classical closed maxT testing. When many endpoints are tested, the tea test also often outperforms shared-uniform randomization.
Read the full abstract

All authors Giovanni Corradini, Marvin N. Wright, Ronja Foraita, Livio Finos

Quantum Physics

Reachability Constraints and Monotones for Quantum-State Texture

How a quantum resource can be quantified depends on which operations are free. Yet two proposed quantifiers of quantum state texture can increase under free operations: the Wigner-Yanase-Dyson skew information relative to the unique textureless state and the squared Hilbert-Schmidt distance from that state. We show reachability from every state orthogonal to the textureless state, whether pure or mixed: each can be freely transformed into any other state. Therefore, every texture monotone must assign the same maximal value to all such states, a requirement that neither candidate satisfies. We then construct valid texture monotones by adding the population outside the textureless state to the skew information and by maximizing the squared Hilbert-Schmidt distance over free operations. We prove faithfulness, convexity, and invariance under textureless ancillas, characterize additivity properties and strong monotonicity, satisfied only in the Wigner-Yanase case, and derive tight bounds for both monotones at fixed overlap with the textureless state and characterize all saturating states. Together, these results show how allowed state transformations impose operational constraints on the quantification of quantum state texture.
Read the full abstract

Quantum Physics

Exploring Variational Regularization of Krylov Subspaces

Krylov subspace methods efficiently approximate extremal eigenstates and eigenvalues of many-body Hamiltonians, but non-orthogonal bases can suffer from ill-conditioning. This challenge is amplified in quantum computing, where direct state access and thus straightforward orthogonalization are unavailable. We investigate a variational Krylov-like framework that iteratively optimizes a parametrized sequence of states to mimic Lanczos-like couplings while penalizing linear dependencies. Analyzed against exact Krylov and Lanczos on a finite Anderson impurity model, ansatz expressivity strongly influences the performance: Hamiltonian-inspired ansätze preserve Krylov-like convergence while improving the conditioning, whereas shallow ansätze remain stable but miss low-energy directions.
Read the full abstract

All authors Alexandre Perrin, Jérémie Messud, Sébastien Burdin, Matthieu Saubanère

Human-Computer Interaction

Rhythmic Fitness Beat 'Em All using Real-Time Monocular Human Pose Estimation

Short and frequent physical activity exercices mitigate the health risks of sedentary behavior. Yet, the lack of motivation remains a critical barrier to adoption. Exergames gamify physical activity, but most rely on dedicated hardware, and outside dance and rhythm games they rarely use the musical beat. We present a work-in-progress beat 'em up exergame played with the whole body in front of a single smartphone, which runs the entire pipeline: an off-the-shelf real-time 3D human mesh recovery model on the phone's neural processing unit, gesture recognition, and the game, shown on any TV through a USB-C to HDMI cable. Our contributions are (1) a 3D recognition of five shadow-boxing moves (guard, punch, slip, squat, jump), each defined by one metric measure in a body frame aligned with gravity and one threshold with hysteresis, personalised during a tutorial; (2) a rhythm stage whose discrete cues are placed on the song's kick drum and judged on the movement's capture time, with a delay learned for each player; and (3) a phone-only architecture, running at 30 frames per second, that turns a phone and an HDMI cable into an exergame console.
Read the full abstract

Operator Algebras

Property Gamma and McDuff Factors from Shift-Similar and Houghton-like Groups

We initiate the study of von Neumann algebras of shift-similar and Houghton-like groups, introduced by Mallery and Zaremsky as analogs of self-similar and Röver-Nekrashevych groups, respectively. Given a shift-similar group $G$ of permutations of $\mathbb{N}$, one can form the Houghton-like group $H_n(G)$, a “mashup" of the classical Houghton group $H_n$ with $G$, which acts on $X_n$, the disjoint union of $n$ copies of $\mathbb{N}$. One main result is that when $G$ and the canonical shift map $\sigma$ admit a joint Følner sequence for their action on $\mathbb{N}$, then both group von Neumann algebras $L(G)$ and $L(H_n(G))$ are II$_1$ factors with property Gamma and hence $G$ and $H_n(G)$ are inner amenable. More generally, let $\Gamma$ be an ICC group of permutations preserving the ends of $X_n$. We prove that if the natural near action on $\mathbb{N}$ of the germs of almost permutations arising from $\Gamma$ along some ray is amenable, then $L(\Gamma)$ has property Gamma; if this group of germs is itself amenable, then $L(\Gamma)$ is McDuff. Then we construct a shift-similar group $G$, using property (T) groups with certain actions, such that neither $G$ nor $H_n(G)$ is inner amenable. Finally, we produce extremely broad and diverse sources of shift-similar groups satisfying the joint amenability condition. These examples come from residually finite groups and from wobbling groups of a variety of metric spaces. In particular, we completely characterize all metrics $d$ on $\mathbb{N}$ for which the wobbling group $W(\mathbb{N},d)$ is shift-similar in terms of a coarse geometric property—namely, that $\sigma^{-1}$ is bornologous with respect to $d$. As a consequence, this gives us necessary and sufficient conditions for when the wobbling group of any countable metric space admits a shift-similar embedding.
Read the full abstract

Functional Analysis

A Fermat–Torricelli convex-hull characterization of inner-product spaces via Birkhoff–James orthogonality

The classical location problem posed by Weber [Über den Standort der Industrien,Tübingen: J. C. B. Mohr, 1909] has motivated various geometric characterizations of inner-product spaces. Benítez, Fernández and Soriano [Trans. Amer. Math. Soc. 354 (2002), 5027–5038] characterized real inner-product spaces through the convex-hull property of Fermat–Torricelli medians.Pakhrou [J. Funct. Anal. 285 (2023), 110078] further established a characterization using $\gamma$-centers and equilateral triangles. Building on these results, we show that the Fermat–Torricelli convex-hull criterion remains valid when restricted to triples satisfying a one-sided Birkhoff–James orthogonality condition.The proof combines norming functionals, radial rescaling,Alexandrov differentiability, and the area formula.No completeness, smoothness, or strict convexity assumptions are required.
Read the full abstract

Artificial Intelligence

Fused-Softmax for efficient quantized attention

Low-precision matrix throughput on modern GPUs is growing far faster than vector-arithmetic throughput. Attention therefore cannot realize the available GEMM performance because softmax has become the pipeline bottleneck. We introduce Fused-Softmax Attention, a training-free algorithm that reformulates and approximates softmax to fold its dominant operations into the attention GEMMs, moving work from vector units to matrix cores. Implemented for the AMD CDNA 4 architecture, our MXFP8, MXFP6 and MXFP4 kernels reach approximately 3.0, 4.0 and 4.6 PFLOP/s on an AMD Instinct MI355X GPU, yielding speedups of $2.1\times$, $2.8\times$ and $3.2\times$ over the BF16 attention. Importantly, the additional error introduced by Fused-Softmax remains below the error floor set by quantizing the GEMM operands. On Wan2.2-T2V-A14B, a hybrid MXFP6/4 schedule delivers up to $2.05\times$ end-to-end speedup while preserving the visual quality of the outputs. The kernels ship in AMD's open-source AITER library, and they powered AMD's MLPerf Inference v6.1 Wan2.2 submission, which ran $1.7\times$ faster than AMD's previous-round result on the same hardware. Our results show that redesigning softmax to utilize the matrix pipeline enables attention to capitalize on continued growth in low-precision matrix throughput.
Read the full abstract

Computer Vision

HMMF-Net: A Lightweight Hierarchical Multi-Frame Multimodal Fusion Network for Accurate Fish Feeding Intensity Assessment

Automated fish feeding intensity assessment (FFIA) is important for precision aquaculture and feed cost optimization. However, unimodal methods are vulnerable to acoustic noise and visual disturbances, while existing multimodal frameworks often rely on computationally intensive 3D convolutions or Transformers, limiting their use on edge devices. To address these limitations, we propose HMMF-Net, a lightweight multimodal architecture that combines early multi-frame visual encoding and multi-axis acoustic modeling through a hierarchical bidirectional messenger cross-modal fusion (BMCF) mechanism. The visual branch captures temporal variations in feeding behavior by stacking temporally distant frames as a multi-channel input to an adapted 2D EfficientNetB0 backbone. The acoustic branch employs adaptive multi-axis residual units (AMARU) to extract temporal, frequency, and local time-frequency representations from log-mel spectrograms. BMCF progressively exchanges information between the two modalities at intermediate representation levels through compact messenger tokens, reducing the computational overhead of dense cross-modal attention. Evaluated on the AV-FFIA dataset using five-fold cross-validation, HMMF-Net achieves 96.77% accuracy and 96.88% F1-score on the held-out test folds, with 3.987M parameters and 1.002 GFLOPs. Compared with the best-performing visual configuration (95.49% accuracy), HMMF-Net reduces the classification error by 28.38%. Deployment on an NVIDIA Jetson Orin Nano using TensorRT with FP32 precision maintains comparable accuracy, with an inference time of 11.54 ms per 2-second sample and a model size of 18.81 MB. These results demonstrate that HMMF-Net combines effective multimodal representation learning with computational efficiency, enabling accurate and practical FFIA on resource-constrained edge platforms.
Read the full abstract

All authors Phi-Hung Hoang, Van-Manh Tran, Nam-Thuan Trinh, Thi-Thu-Hong Phan

Computation and Language

KLEAR: Knowledge-Matched LLM Evaluation in Korean across Application Requirements via Controlled Benchmark Transformation

Existing Korean benchmarks for Large Language Models (LLMs) evaluate performance across diverse knowledge domains, but their accuracy scores cannot distinguish whether incorrect responses arise from a lack of required knowledge or a failure to apply that knowledge correctly. Our analysis of existing Korean benchmarks reveals that most questions can be answered through factual recall alone, while those demanding greater application also differ in required knowledge, making it difficult to isolate application ability. To address this, we introduce KLEAR, a framework that constructs knowledge-matched contrasts by varying application demand while preserving the required knowledge. During construction, KLEAR identifies the knowledge underlying each source question as an explicit anchor bundle to guide generation of questions requiring more complex application. Generated candidates then undergo automated and human validation, with iterative revision based on evaluation feedback. Experiments demonstrate that the resulting knowledge-matched sets largely satisfy the intended constraints, enabling controlled evaluation across diverse LLMs. These evaluations reveal that increasing application demand exposes failures that persist even when the required knowledge is provided.
Read the full abstract

All authors Seoyeon Kim, Jeonghyun Kang, Hamin Koo, Barom Kang, Heondeuk Lee, Jaehyung Kim

Computation and Language

Khmer PFGen-Bench: Building a Judge-Free Benchmark for Open-Ended Generation in a Low-Resource Language

Benchmarks for open-ended generation often rely on large language model (LLM) judges, whose reliability in low-resource languages is uncertain. We introduce Khmer PFGen-Bench, to our knowledge the first judge-free benchmark for multi-sentence answers to questions composed directly in Khmer, a low-resource language. Khmer PFGen-Bench adapts the scorer of PFGen-Bench, a benchmark for Japanese that scores responses with keyword rules and n-gram statistics from LLM-generated references. This scoring requires references that are both valid and varied, which LLMs may not produce in Khmer, so we examine how to build such references. We find that post-trained models can give more valid but less varied references than their base variants, and that mixing model families increases variety for the same number of references. Using references from post-trained models in nine families, the benchmark yields model rankings that are stable across reference sets and consistent with a native speaker's ratings. These findings offer guidance for building judge-free benchmarks in other low-resource languages.
Read the full abstract

All authors Mengsay Loem, Kentaro Imajo, Ryota Ozaki, Masanori Hirano

Analysis of PDEs

Boundary singularities for the Lane-Emden equation

For every $n\ge3$ and $(n+1)/(n-1)<p<(n+2)/(n-2)$, we construct positive solutions of $\Delta u+u^p=0$ in the half-space with zero Dirichlet data away from the origin, the homogeneous singular profile there, and Poisson-kernel decay at infinity, resolving a question of del Pino–Musso–Pacard. Under a weighted non-degeneracy condition, these half-space solutions serve as building blocks for positive very weak solutions in smooth bounded domains, singular along prescribed compact smooth boundary submanifolds of normal dimension $n$. We verify non-degeneracy in the initial stable interval and at and just beyond its threshold when this lies below $(n+2)/(n-2)$. For isolated boundary singularities, the construction applies throughout the strict subcritical interval outside a locally finite exceptional set.
Read the full abstract

Algebraic Geometry

Centres and semisimple deformations of contraction algebras

Let $A_{\rm{con}}$ be the contraction algebra of a $3$-fold flop of an irreducible curve, with genus-zero Gopakumar-Vafa invariants $n_1, \dots, n_l$. We show that $\prod_j M_j(\mathbb{C})^{n_j}$ is the only semisimple algebra to which $A_{\rm{con}}$ deforms. Hence $A_{\rm{con}}$ encodes its Gopakumar-Vafa invariants intrinsically, answering a question of Wemyss. Hua and Toda showed that this deformation exists, and Jelisiejew and Smoktunowicz proved its uniqueness in types A and D. The key step in our proof is that the centre of $A_{\rm{con}}$ has dimension $n_1 + \dots + n_l$. To prove this, we lift the contraction algebra of the universal flop to the Cartan subalgebra, where it becomes a corner of a deformed preprojective algebra. We show that the universal contraction algebra and its zeroth Hochschild homology are free over a polynomial ring, with presentations whose determinants are products of roots. Pulling these presentations back along the arc that classifies the flop gives the count, and recovers Toda's formula for $\dim A_{\rm{con}}$. We also prove Hua and Toda's flatness conjecture, and determine which Gopakumar-Vafa invariants occur in lengths 3 to 6.
Read the full abstract

Optics

When is an Optical Design Problem a Mass Transfer Problem?

This paper investigates whether optical design problems can be represented as Monge–Kantorovich mass transfer problems with some cost function. An optical problem is specified by a function defining a particular family of supporting surfaces for the optical element. We introduce the notion of geometric equivalence between optical and transport problems, which is directly related to the possibility of formulating an optical design problem as a mass transfer problem. A necessary condition for such equivalence is derived in the semi-discrete setting. It allows a transport representation to be ruled out without choosing a cost function in advance. The condition is applied to the near-field refractor problem and to the design of two refracting surfaces transforming a spherical beam into a spherical or collimated beam. For these problems, the necessary condition is shown to fail in physically admissible configurations. We also study an optical element with a single working surface and an arbitrary regular incident wavefront. In the far field, geometric equivalence to a mass transfer problem requires a planar or spherical wavefront, whereas in the near field with a planar target, the necessary condition fails for an arbitrary smooth wavefront.
Read the full abstract

Software Engineering

BOXBOX: Recursive Harness Self-Improvement for Long-Running Performance Optimization

Automated harness optimization improves agents by learning from trajectories and task feedback. Existing offline methods use reference answers to optimize harnesses for reuse on held-out tasks. However, many real-world tasks are open-ended, while long runs expose new difficulties that offline-optimized harnesses may not address. To study harness optimization in these settings, we introduce PerfRace, a benchmark for long-running performance optimization of software projects, and investigate whether agents can use task-specific experience to recursively improve their harnesses during an ongoing task. To this end, we introduce BOXBOX, which enables agent-triggered harness revision and kernel-managed hot reloading while preserving execution context. We evaluate BOXBOX on PerfRace under matched test-time budgets, comparing it with Frozen Harness, offline Harness Evolution, and Harness Scaling. BOXBOX achieves higher speedup with more efficient use of task time and model cost, reaching a Speedup@4h of 1.70x, compared with 1.54x for the strongest baseline.
Read the full abstract

All authors Yuhao Tan, Hanyun Jiang, Fangkai Yang, Yuan Yao, Taolue Chen, Bo Qiao, Yu Kang, Lu Wang, Pu Zhao, Xiaoxing Ma, Qingwei Lin, Saravan Rajmohan, Qi Zhang

Machine Learning

Provable State Estimation with Recurrent Models

Estimating the state of a dynamical system from observations is a crucial task in robotics and control. Despite their empirical success, recurrent models often lack formal guarantees on convergence and robustness. We bridge this gap by analyzing these architectures from the perspective of nonlinear observer theory. By framing RNNs as a time-discretization of continuous latent dynamical systems (Neural ODEs), we demonstrate that enforcing a contraction condition on the latent flow guarantees the existence of a smooth, injective map of the true physical state into the latent space. We prove that this architecture acts as a tunable observer: by increasing a single scalar gain that scales the latent vector field, the estimation error can be driven to an arbitrarily small neighborhood of zero at an exponentially fast rate. We further establish quantitative bounds on the MLP complexity required to decode this latent state, and extend our formal guarantees to the discrete-time residual RNN implementation. We validate the theory on a real mobile robotic task, where a provably contracting model outperforms a vanilla GRU with substantially more stable training. We also report that unconstrained GRUs converge to contracting dynamics on this task, suggesting that contraction is an emergent property of well-trained sequence models.
Read the full abstract

All authors Vincent Andrieu, Pauline Bernard, Lucas Brivadis, Laurent Praly, Christian Wolf

High Energy Astrophysical Phenomena

Quasi-spherical fallback in binary neutron-star mergers: fuel and obstacle for extended emission

Extended emission (EE) lasting hundreds of seconds has been observed in a fraction of short gamma-ray bursts (sGRB) containing both nonthermal and thermal components, and representing a significant puzzle for models involving binary neutron stars (BNSs). Recent GRMHD simulations of BNS mergers found that the late-time fallback is approximately quasi-spherical and motivated an r-process powered thermal model for EE. However, the available radioactive-heating power may be insufficient to sustain the luminosity inferred by such a model. We therefore propose a fallback-fed secondary-jet scenario in which a fraction of the fallback material circularizes on the black-hole equatorial plane and powers the jet, while polar fallback impedes its propagation. Using a semi-analytic toy model, we follow the propagation of the jet-head through the fallback envelope and identify jet-dominated, cocoon-dominated, and choked scenarios. The boundaries of these regimes depend primarily on the circularization fraction $f_{\rm circ}$, on the jet half-opening angle $\theta_j$, and on the polar refilling factor $\xi_{\rm pol}$, with the approximate scaling $\widetilde{L}\propto f_{\rm circ}/(\xi_{\rm pol}\theta_j^2)$. By post-processing data from BNS-merger simulations we find that the fraction of bound fallback material that avoids prompt capture is $f_{\rm npc}\simeq0.52$, thus providing an upper bound on the circularizable fallback supply. Under suitable prescriptions for nonthermal jet emission and thermal emission from the jet head, the estimated isotropic-equivalent $0.3$–$10$ keV energies reproduce well both the observed EE distribution for the jet-dominated branch and its low-energy end for the cocoon-dominated branch. Overall, these results suggest that fallback-fed secondary jets represent a dynamically feasible and energetically consistent explanation of the EE in sGRBs.
Read the full abstract

All authors Zhiwei Chen, Marie Cassing, Paramvir Singh, Luciano Rezzolla