Chemical Physics
MAST: Motif-Augmented Diffusion with Search Tree for Spectroscopic Molecular Structure Elucidation
Chenghao Jia, Mengdi Liu, Hong Chang et al. (+2 more)
Elucidating molecular structures from spectra is a foundational problem in chemical and materials characterization, yet remains challenging due to spectral ambiguity and the vast molecular space. Although recent
diffusion-based generators show strong promise for spectra-conditioned elucidation, existing methods struggle to learn robust spectra-structure relationships from limited paired data when relying solely on global spectral representation. Moreover, the repeated full sampling inference strategy incurs substantial computation overhead. To address these limitations, we propose MAST, a Motif-Augmented
diffusion framework with Search Tree, for joint 2D-3D spectroscopic molecular structure elucidation. MAST introduces explicit, interpretable motif priors as intermediate evidences throughout denoising, reducing conditional ambiguity and facilitating spectra-conditioned optimization. We further cast
diffusion sampling as reward-guided tree search to prioritize high-reward denoising trajectories, yielding a compact set of spectra-consistent candidates under limited budgets. On the QM9S multi-spectra
benchmark, MAST achieves 94.89% exact recovery and improves 3D fidelity, while preserving high chemical validity and stability. Code is available at https://github.com/Jia040223/MAST.
Read the full abstract ▾
All authors Chenghao Jia, Mengdi Liu, Hong Chang, Shiguang Shan, Xilin Chen
Information Theory
FTN Signaling: Spectral Efficiency from BPSK to 16-QAM
Emre Cerci, Adem Cicek, Melda Yuksel et al. (+4 more)
Faster-than-Nyquist (FTN) signaling can improve spectral efficiency by transmitting symbols closer together than the Nyquist limit, at the cost of additional inter-symbol interference (ISI). In this letter, we study the finite-alphabet symbol-wise achievable information rate (AIR) of FTN signaling for BPSK, QPSK, and 16-QAM under a transmit-power constraint and a common minimum mean square error (MMSE) channel-shortening (CS) detection framework, cross-validated for BPSK against an independent reduced-state Ungerboeck Bahl-Cocke-Jelinek-Raviv (BCJR)
benchmark. Our results show a clear trend. Lower-order constellations can benefit from more aggressive time acceleration, while higher-order constellations become less tolerant of acceleration and achieve their best spectral efficiency closer to the Nyquist limit. At a reference signal-to-noise ratio (SNR) of 6 dB the optimum acceleration factor shifts from $\tau^\star\approx0.65$ for BPSK to 0.75 for QPSK and 0.90 for 16-QAM. We further show that this trend is preserved in coded systems and remains robust to the detector
memory. The impact of FTN operation on the instantaneous-to-average power ratio (IAPR) is also examined. Overall, the results highlight how constellation order should be considered when selecting the operating point for FTN signaling.
Read the full abstract ▾
All authors Emre Cerci, Adem Cicek, Melda Yuksel, Gokhan M. Guvensen, Enver Cavus, Yaser Dalveren, Halim Yanikomeroglu
Optics
Minimizing round-trip loss for near-unity quantum efficiency in organic light-emitting diodes
Junho Kim, Seonjeong Lee, Seunghyup Yoo et al. (+1 more)
Modern organic light-emitting diodes (OLEDs) reach internal
quantum efficiencies close to 100%, yet in a planar device only 20-30% of
photons escape. External outcoupling structures on high-index substrates recover some of the trapped light, but external
quantum efficiencies (EQEs) have stalled at 55-60% across different optical designs, and the origin of this ceiling is not fully understood. Here we identify light extraction in such OLEDs as a multi-pass process: light that the outcoupling structure fails to extract returns to the OLED, is reflected back and may escape over multiple passes, with the efficiency governed by the absorption per round trip. Absorption loss is minor in conventional planar OLEDs, but it becomes dominant and accounts for the observed efficiency ceiling when an outcoupling structure is attached. Consistent with this picture, adding a low-loss auxiliary reflector raised the EQE of an OLED on a high-index microlens substrate from 51% to 77% without any change to the OLED stack or the outcoupling structure, the highest value reported for a single-junction OLED without a macroscopic extraction lens; for an OLED on a conventional glass substrate an estimated 92% of the light delivered to the substrate was extracted. We derive four design rules that keep round-trip losses to a few percent and predict that EQEs around 90% are feasible with established materials. Under these rules, microcavity tuning and emitter orientation largely cease to matter, and high efficiency is retained across a wide range of microlens designs and scattering layers.
Read the full abstract ▾
All authors Junho Kim, Seonjeong Lee, Seunghyup Yoo, Malte C. Gather
Computation and Language
ILM: An AI-Powered Storytelling Educational Tool
Suhaila Mohammed, Abdelaziz Serour and Allison Lahnala
Digital technologies have made Islamic narratives more accessible, but existing platforms provide limited support for structured learning and comprehension of these stories, particularly in Arabic and
multilingual settings. We present ILM, an interactive educational platform for Stories of the Prophets that combines Arabic natural
language processing, structured knowledge representation, and
retrieval-based question generation. Admin-approved Arabic narratives are processed by a Knowledge Graph (KG) Constructor Engine that identifies entities and narrative relationships and stores them as structured knowledge, enabling learners to explore stories through a visual story map and answer entity- and relation-based questions generated from the KG. Separately, a
multilingual retrieval pipeline retrieves relevant passages from the original narratives to generate multiple-choice and open-ended comprehension questions. For open-ended questions, an LLM-as-a-Judge evaluates learners' answers against the retrieved passages and reference answers to determine correctness. The platform also incorporates Quranic content as a separate enrichment layer, allowing selected narratives to be supplemented with source-supported information. By combining structured knowledge with passage-based
retrieval, ILM supports narrative exploration, comprehension, and assessment across Arabic and
multilingual content. The system demonstrates the feasibility of combining structured knowledge representation and
retrieval-based generation to support interactive learning of Islamic narratives. A demo is available at anonymous.4open.science/r/mml-5FCF.
Read the full abstract ▾
Functional Analysis
Hadamard powers of Gram matrices and Generalized Cauchy-Schwarz inequalities
Nathaniel Johnston and Mohammad Sababheh
The following inequality for vectors $\mathbf{v}, \mathbf{w} \in \mathbb{R}^n$ was considered in [Linear and Multilinear
Algebra 74:1779–1798 (2026)]: \[\|\mathbf{v}^p\|\|\mathbf{w}^p\| - \langle\mathbf{v}^p,\mathbf{w}^p\rangle \leq \|\mathbf{v}\|^p\|\mathbf{w}\|^p - \langle\mathbf{v},\mathbf{w}\rangle^p,\] where $\mathbf{v}^p$ is the vector that is obtained by raising each entry of $\mathbf{v}$ to the power $p$. It was shown that this inequality holds for all positive integers $p$, and it was conjectured that it holds for all real $p \geq 2$ when the entries of $\mathbf{v}$ and $\mathbf{w}$ are positive. A formula for the maximum amount by which the inequality's right-hand side can exceed its left-hand side for nonnegative unit vectors was also conjectured. We prove both of these
conjectures, and we determine exactly when the inequality holds with equality. We also present a new generalization of this inequality to three or more vectors.
Read the full abstract ▾
Optics
Observation of robust corner states in photonic crystals without global symmetries
Zitao Ji, Jianfeng Chen, Yidong Zheng et al. (+2 more)
Robust corner states have attracted considerable
attention in recent years, particularly in photonic systems. They are typically realized in second-order topological
insulators, including shrunken-expanded
lattices and breathing Kagome
lattices, where global
symmetries are essential in their formation. This reliance on
symmetry, however, can make these states highly sensitive to structural disorder. Here, we propose and experimentally demonstrate a class of robust corner states in photonic
crystals that require neither global
symmetry nor nontrivial
topology. By locally tailoring the corner geometry of an otherwise topologically trivial structure, we create strongly localized modes that remain stable under both position and radius disorder. These states are not accidental; rather they originate from a pair of Dirac points in the bulk band structure.
Geometric tuning shifts the flatband edge states connecting the Dirac points into an otherwise trivial bandgap, thereby inducing strongly localized corner states. Our findings uncover a previously overlooked class of
symmetry-independent, disorder-resilient corner states, extend robust localization beyond conventional topological protection, and provide a route toward light localization in all-dielectric photonic systems.
Read the full abstract ▾
All authors Zitao Ji, Jianfeng Chen, Yidong Zheng, Zhaogang Dong, Zhi-Yuan Li
Artificial Intelligence
When Should Agents Think? Adaptive Reasoning via Cross-Turn Estimation
Yiruo Cheng, Shen Huang, Xiaoshuai Song et al. (+5 more)
Large
language model (LLM)-based
agents have demonstrated strong capabilities on complex tasks. They typically perform
reasoning before each action throughout an interaction trajectory. However,
reasoning may not be necessary at every turn, as
reasoning produced earlier can continue to support subsequent actions. A key challenge is therefore to determine when existing
reasoning remains sufficient and when a new
reasoning step is needed, without relying on costly generation-based verification. We find that decreases in the likelihood of subsequent reference actions after removing additional
reasoning closely track whether those actions remain recoverable given earlier
reasoning, providing an effective and lightweight signal for estimating cross-turn action support. Based on this observation, we propose
Reasoning Adaptation through Cross-Turn Estimation (RACE), a training approach for adaptive
agent reasoning. RACE introduces a Likelihood-Guided Progressive
Reasoning Cover Detection (LoGiC) procedure that progressively identifies
reasoning turns whose removal has limited impact on the current and subsequent reference actions. The resulting removal signals are incorporated into both supervised fine-tuning and agentic reinforcement learning, enabling the policy to learn when to reason and when to act directly. Extensive experiments on four representative
agent benchmarks show that RACE substantially reduces
reasoning cost while maintaining or improving task performance.
Read the full abstract ▾
All authors Yiruo Cheng, Shen Huang, Xiaoshuai Song, Jiejun Tan, Guanting Dong, Pengjun Xie, Ji-Rong Wen, Zhicheng Dou
Computer Vision
Look Back, Think Ahead: Visual Memory on Demand for Efficient Multimodal Reasoning
Yicheng Xue, Han Wu, Jufeng Yang et al. (+4 more)
Processing long visual
token sequences from high-resolution images makes multi-step
reasoning computationally expensive for
multimodal Large
Language Models (MLLMs). Existing one-shot
pruning and aggregation methods compress visual
tokens into a fixed context before decoding. However, visual evidence needs can shift as
reasoning unfolds, making it difficult for a fixed compressed context to retain all the details needed across stages. To address this challenge, we propose ViMoD, a lightweight framework that maintains a compact visual context while preserving access to original fine-grained evidence as
reasoning needs evolve. Deformable Aggregation of Region-wise
Tokens (DART) learns content-adaptive groups and aggregation capacities, constructing compact Coarse representations linked to recoverable original Fine
tokens. Temporal Routing for Adaptive Contextual Evidence (TRACE) integrates decoding history to anticipate upcoming evidence needs and select, retain, or replace active Fine-
token groups. Selected Fine
tokens augment the persistent Coarse context in the frozen backbone, enabling stage-specific evidence access without continuously attending to all visual
tokens. On Qwen3-VL-4B, ViMoD outperforms all evaluated baselines on all eight
reasoning benchmarks at a 20% target visual
token budget, improving the mean normalized score by 39.0% over the strongest evaluated one-shot baseline. These gains are achieved with only 0.0546% additional trainable parameters relative to the frozen backbone.
Read the full abstract ▾
All authors Yicheng Xue, Han Wu, Jufeng Yang, Minjing Dong, Xinghao Chen, Hanting Chen, Jianyuan Guo
Statistical Methodology
Design-Based Validation Method of EHR Algorithms to Address Partial Verification Bias
Rushi Tang, Maya Blasingame, Lillian Zheng et al. (+4 more)
Electronic health record (EHR) phenotyping
algorithms support clinical prediction, observational research, and quality improvement, but manual chart review for gold-standard validation is costly. For
algorithms targeting rare outcomes,
algorithm-negative patients often greatly outnumber
algorithm-positive patients. We propose a design-based validation framework that fully verifies
algorithm-positive patients and reviews a simple random sample (SRS) of
algorithm-negative patients. Horvitz-Thompson
estimators of false-negative (FN) and true-negative totals yield estimates of sensitivity, specificity, negative predictive value, positive predictive value, and accuracy. We derive finite-population-corrected variance
estimators and confidence intervals (CIs), with sample-size planning for a prespecified CI half-width. An optional Phase 2 extension uses pilot data to assess whether risk-guided Neyman allocation may improve precision. In simulations, the design-based approach substantially reduced bias relative to naive verified-sample analyses and achieved near-nominal CI coverage. Adaptive allocation reduced root mean squared error and CI width when FNs were concentrated across risk strata, but offered little benefit when the risk signal was weak. We illustrate the framework using a breast cancer recurrence
algorithm in 3,336 patients, with endpoint-specific
algorithm-negative review budgets of 125, 150, and 175 charts. The framework combines SRS-based validation and precision planning with optional pilot-guided adaptive allocation.
Read the full abstract ▾
All authors Rushi Tang, Maya Blasingame, Lillian Zheng, Joan Ac-Lumor, Saher Mubarek, Jennifer K. Plichta, Samuel I. Berchuck
High Energy Physics (Experiment)
Optimising Geometric Deep Learning for Varied Reconstruction Tasks in High Granularity Calorimeters
Matthieu Melennec and Frédéric Magniette
In the recent years, high energy physics discoveries have been driven by the increasing of luminosity and/or detector granularity. This
evolution gives access to bigger statistics and data samples, but can make it hard to process the detector outputs with current methods and
algorithms. Graph Neural Networks (GNNs), have been shown to be powerful tools to address these challenges. While GNNs can deal with the non-euclidean nature of the data extracted from these detectors, their deployment presents significant difficulties, mainly from the computational complexity of the
algorithms they are based on. We propose to use the known geometries of the detectors to optimise the
algorithms used for GNN pipelines, which usually have quadratic algorithmic complexities. We validate the validity of our optimised approach on various tasks, such as particle identification, energy regression and instance
segmentation. Finally, we present HIBOU, a generic modular architecture that provides a blueprint for how these GNN blocks are to be implemented within experiment-specific reconstruction frameworks.
Read the full abstract ▾
Data Structures and Algorithms
Faster Directed Hopsets, Distance Preservers, and Deterministic Shortcut Sets
Ben Bals, Daniel Dadush, Gary Hoppenworth et al. (+2 more)
While there have been many recent advances in getting better tradeoffs for directed distance structures such as hopsets, shortcut sets and distance preservers, most known
algorithms are inefficient. In this work we provide three results: - Faster
algorithms for directed $(1+\epsilon)$-hopsets, matching the state-of-the-art size/hopbound trade-offs of Bernstein & Wein [SODA23]. Our
algorithm is first designed for DAGs and then extended to general graphs (up to $n^{o(1)}$ factors) using the recent DAG projection result of Haeupler, Jiang, and Saranurak [STOC26]. - Faster deterministic
algorithms for constructing shortcut sets, matching the state-of-the-art size/hopbound tradeoffs of Kogan & Parter [SODA22A]. - Using our improved directed hopset construction, we get significantly faster
algorithms for computing source-wise distance preservers. This
algorithm is based on speeding-up reductions of Kogan & Parter [SODA22B] by using certain structural properties of our hopsets.
Read the full abstract ▾
All authors Ben Bals, Daniel Dadush, Gary Hoppenworth, Yasamin Nazari, Rajath Rao K.N
Artificial Intelligence
Structure Tax: How Structured Output affects LLMs Performance
Vineet Kumar, Kanishka and Bhuvanesh Mandora
Deploying large
language models in production often requires constraining outputs to structured formats such as JSON or XML, and prior work treats the resulting accuracy loss as an inherent `structure tax'. We re-examine this claim by evaluating a battery of models,
datasets and schemas, measuring task accuracy, confidence calibration, and hidden-state geometry. The tax turns out to depend on schema design rather than on structure per se:
reasoning-first field ordering matches or exceeds free-form accuracy, while answer-first ordering causes steep drops, particularly in smaller models. Format sensitivity scales inversely with a task's own structural constraints, and schemas that preserve
reasoning order also improve calibration with CKA showing greater separability between correct and incorrect representations in middle
transformer layers. Our findings indicate that properly designed structured formats can match or exceed free-form performance, reframing the critical question from `whether to structure' to `how to structure' for optimal
reasoning preservation.
Read the full abstract ▾
Astrophysics of Galaxies
Binding Energies of Astrophysically Relevant Molecules on CO$_2$ Clusters: Benchmarking and the Donor-Acceptor Origin of Surface-Dependent Binding
Aneesa Ahmad, Catherine Walsh and Stefan Vogt-Geisse
Carbon dioxide (CO$_2$) is among the most abundant interstellar ices and has been observed by JWST in water-rich mixtures and pure CO$_2$ environments. Astrochemical models adopt binding energies (BEs) measured on water ice because few BEs are available for CO$_2$ ice. We calculate BE distributions for twelve adsorbates (N$_2$, CH$_4$, HCl, C$_2$H$_2$, H$_2$S, HF, HCN, H$_2$CO, HCOOH, HNC, CH$_3$OH, H$_2$O) on CO$_2$ clusters containing two to five molecules using DFT calibrated against CCSD(T)/CBS, a coupled-cluster treatment extrapolated to the complete-basis-set limit. Binding strength depends on whether the adsorbate requires the surface to donate a hydrogen bond (H-bond). CO$_2$ can accept an H-bond but cannot donate one, and all six species compared with water ice from Ahmad et al. (2026) bind more weakly on CO$_2$. The ratio of the BE on CO$_2$ to water ice ranges from 0.66 for H$_2$CO and 0.75 for H$_2$O to 0.87 for CH$_3$OH, 0.89 for HCOOH and 0.92 for C$_2$H$_2$. Formaldehyde is predicted to desorb near 57 K from CO$_2$-rich ice, compared with 85 K from water-rich ice, and could contribute to cold gas-phase H$_2$CO. Although the mean binding energies are changing with cluster size for several species, the differences can be traced to local binding interactions likely to remain important in larger ice models. The BEs vary by more than 500 K between sites, while the desorption peaks range from 11 K for CH$_4$ to 132 K for HCOOH. The limited H-bonding capacity of CO$_2$ may favour separation of mixed H$_2$O:CO$_2$ ice into CO$_2$-rich domains.
Read the full abstract ▾
Group Theory
Self measurable embeddings of mapping class groups
Alexandra Gurieva
Given a finite-type orientable surface without boundary, outside of a few low-complexity surfaces, we show that any measurable
embedding of the associated mapping class group into itself is a measure equivalence. This is a measure-theoretic analog to the result of Ivanov and McCarthy that mapping class groups are co-Hopfian.
Read the full abstract ▾
Probability
Loop soup percolation and boundary dimension on metric trees
Christian Mönch
We determine the percolation threshold of the Brownian loop soup on a transient locally finite weighted tree with proper cable metric and no killing. It equals $1-D_G/2$, where $D_G\in[0,2]$ is the Hausdorff dimension of the boundary in the metric induced by the normalized Green function. This gives an analogue of Lyons' branching-number formula for Bernoulli percolation, with Green distance replacing graph distance. Spherically symmetric trees with unit conductances and branching number two realize every threshold in $[1/2,1]$, while
geometric conductances give explicit positive thresholds below one-half and a subdivided binary tree has threshold zero. Under the capacity assumption of Drewitz, Prévost and Rodriguez, we characterize the threshold one-half by a finite-energy condition, without a minimum-degree assumption. Non-percolation at one-half then follows from their
theorem and Lupu's coupling. For supercritical Galton-Watson trees with unit conductances, conditioned on survival, the boundary dimension is almost surely one and the threshold is one-half, with non-percolation at criticality. No offspring moment assumption is required, and leaves and vertices with one child are allowed. The classification also applies intrinsically to proper real trees spanned by their infinite rays.
Read the full abstract ▾
Machine Learning (Statistics)
Diffusion Removes Langevin's Conditioning Dependence: A Sharp Gaussian Analysis
Adam Perbost, Francis Bach and Pierre Marion
Despite their empirical success, why
diffusion models overcome the bottlenecks of classical score-based samplers remains unclear. In this work, we leverage Gaussian distributions to isolate this phenomenon. We establish 2-Wasserstein convergence bounds for optimized hyperparameters, showing that
diffusion processes achieve a sampling error of $O(\sqrt{d\lambda_{\max}}\log N/N)$, where $d$ is the dimension, $N$ the number of sampling steps, and $\lambda_{\max}$ the largest
eigenvalue of the target covariance
matrix. Unadjusted and underdamped Langevin dynamics suffer from an additional $\sqrt\kappa$ factor, where $\kappa$ is the condition number. These rates follow from spectral bounds which are sharp: we confirm them via matching first-order asymptotics as $N\rightarrow\infty$. Our analysis provides a rigorous characterization, in the Gaussian setting, of how time-dependent score trajectories remove condition-number dependence during sampling. By contrast, in the learning phase, we show that estimating the unnoised score by gradient descent leads to essentially the same
estimator as estimating a noisy score, which suggests that the benefits of noising do not come from the learning phase.
Read the full abstract ▾
Instrumentation and Methods for Astrophysics
On-chip filled array hyperspectral kinetic inductance detectors (FHYKID) for line intensity mapping
Emile Prêle, Martino Calvo, Usasi Chowdhury et al. (+3 more)
The next generation Line Intensity Mapping (LIM) experiments will require very high throughput instrumentation, and as a consequence high pixel counts focal planes. For the study of the [CII] line in the
redshift range z=7-8, we develop a hyperspectral filled focal plane (FHYKID) operating in the range 210-235 GHz and with a spectral resolution reaching 120. The spectral filtering is achieved thanks to planar Slot Split Ring Resonators (SSRR) coupled to Lumped Element Kinetic Inductance Detectors (LEKID). This robust concept achieves, thanks to its simplicity, large
quantum efficiency and is scalable, even with current readout electronics systems, to several tens of thousands beams on Sky. Our FHYKID will be one of the four channels of the NIKONA instrument, currently seeking funding and to be installed at the IRAM 30-meter
telescope at Pico Veleta. Among the science goals, simultaneously mapping the local (
galactic and nearby-
galaxies) CO line and the cosmological [CII].
Read the full abstract ▾
All authors Emile Prêle, Martino Calvo, Usasi Chowdhury, Gregory Garde, Florence Levy-Bertrand, Alessandro Monfardini
Cryptography and Security
Protecting CPU AI On Edge TEEs: WebAssembly's Promise and Practical Challenges
Friedrich Vandenberghe, Lachlan J. Gunn, Bruno Volckaert et al. (+1 more)
AI models on edge hardware contain important intellectual property (IP), but an adversary can steal it when they achieve root access. Trusted Execution Environments (TEE) like Arm TrustZone protect against these Operating System (OS) level attacks. However, they are challenging to use. More precisely, it is difficult to run unaltered applications inside a TEE. This work presents a solution that allows the execution of unaltered AI models, compiled to WebAssembly, on the WebAssembly Micro Runtime (WAMR) in OP-TEE for Arm TrustZone. Additionally, this work provides an AI model distributor that encrypts the WebAssembly binary and places the encryption key in one of the device's fuses. This way, only the WAMR Trusted Application (TA) in OP-TEE can decrypt and execute the AI model. A thorough evaluation of our solution shows that it incurs an additional overhead of 22% in comparison to an application manually ported to OP-TEE, while also facilitating the execution of unaltered AI models with an additional inference latency of 6% without significant porting effort. Overall, there is a pressing need to safeguard the IP of AI models and this work shows that there is a real promise in WebAssembly, but there remain some practical challenges.
Read the full abstract ▾
All authors Friedrich Vandenberghe, Lachlan J. Gunn, Bruno Volckaert, Merlijn Sebrechts
General Economics
The Myth of Restrictive Working-Time Agreements: Firms' Working-Time Adjustment after Exit from Collective Bargaining
Vinzenz Pyka and André Rieder
Employers in Germany argue that collective agreements restrict their flexibility concerning working hours. Using data from the IAB Establishment Panel for the years 2008 to 2024, we investigate working-time adjustments after exit from collective bargaining. We find that exit is associated with only a modest increase in weekly working hours of around 0.3 to 0.6 per cent using fixed-effects and event-study models, corresponding to eight to thirteen additional minutes per week. This adjustment occurs shortly after leaving collective bargaining coverage. We further find no evidence that other margins of working-time adjustment, such as Saturday or Sunday work, are used more extensively after exit. In contrast to previous survey evidence, our findings call into question that working-hours constraints are one of the main drivers of exit from collective bargaining coverage.
Read the full abstract ▾
Computer Vision
Do Not Train Away Uncertainty: Early Uncertainty Anchored Calibration
Yutong Xie, Jiawei Tang, Zhenglin Hua et al. (+6 more)
Deep neural networks, including large
language models, have achieved remarkable performance across various tasks. However, they are prone to overconfidence during training or fine-tuning. In this work, we observe a consistent phenomenon across different models that the early model is better calibrated, while later training or fine-tuning yields marginal accuracy gains but substantially increases calibration errors. Our analysis suggests that the early model retains uncertainty awareness in both its predictions and features, which is gradually lost with continued training. To avoid training away this uncertainty awareness, we propose EUA-Cal, a novel method that exploits the Early model as an Uncertainty Anchor for Calibration. EUA-Cal introduces early prediction regularization to preserve early predictive uncertainty and prototype structure regularization to exploit uncertainty reflected in the early feature space, jointly mitigating overconfidence. Extensive experiments on image classification and multiple-choice question answering across eight diverse models demonstrate that EUA-Cal outperforms state-of-the-art calibration methods.
Read the full abstract ▾
All authors Yutong Xie, Jiawei Tang, Zhenglin Hua, Yuxiang Ma, Si Qin, Yaxin Hou, Hui Liu, Junhui Hou, Yuheng Jia
Probability
Characterization and scaling limit of three-dimensional loop-erased random walk and random-walk loop-catchers
Gefei Cai, Xinyi Li and Daisuke Shiraishi
Random-walk loop-catchers (RWLC) are a one-parameter family of random subsets of the trace of a random walk, recently introduced in [Cai 2026, arXiv:2607.18070], which interpolates between loop-erased random walk (LERW) and the full walk trace, satisfying a recovery property. We prove that the continuous version of this recovery property uniquely characterizes the scaling limit of LERW and RWLC in three dimensions, using a Green-function test via entangled multipath LERW, as outlined in [Cai 2026, arXiv:2607.18070]. For LERW, this resolves the 3D case of the
conjecture of Sapozhnikov–Shiraishi [Sapozhnikov–Shiraishi 2018, Probab. Theory Related Fields 172, 615–662] and gives a new and axiomatic proof of the existence of the scaling limit, which was first proved in [Kozma 2007, Acta Math. 199, 29–152]. For RWLC, this gives the construction of 3D Brownian loop-catchers. The characterization also yields inversion invariance of the 3D LERW scaling limit.
Read the full abstract ▾
Robotics
SkillWeave: Weaving Heterogeneous Demonstrations into Long-Horizon Manipulation Skills
Ryosei Tamura, Xiaoxiang Dong, Uksang Yoo et al. (+4 more)
Dexterous manipulation requires both large-scale task progression and precise contact-rich interaction, making it challenging to collect demonstrations that effectively support both regimes. We present SkillWeave, a heterogeneous demonstration framework for long-horizon dexterous manipulation that combines teleoperation for coarse reaching and transport with kinesthetic teaching for precise, contact-rich skills. To address the visual mismatch introduced by the demonstrator's presence during kinesthetic data collection, we propose an object-mask-conditioned
diffusion policy that uses offline object
segmentation for training supervision and a lightweight learned mask predictor at deployment, avoiding online
segmentation and image inpainting. To mitigate distribution shift between independently trained sub-task policies, we introduce successor-aware terminal steering, which selects among actions sampled from the predecessor policy to guide the system toward states supported by the successor's demonstrated initial-state distribution. Across three real-world long-horizon tasks, SkillWeave achieves 27% average end-to-end success. Mask-conditioned kinesthetic policies improve dexterous sub-task success to an average of 65%, while successor-aware handoffs achieve an average composition efficiency of 87%. These results show that matching demonstration modality to interaction regime, explicitly addressing kinesthetic visual mismatch, and steering policy handoffs toward successor-supported states substantially improves long-horizon dexterous manipulation.
Videos and code are available at skillweave-authors.github.io .
Read the full abstract ▾
All authors Ryosei Tamura, Xiaoxiang Dong, Uksang Yoo, Yuemin Mao, Romina Mir, Jonathan Francis, Jeffrey Ichnowski
High Energy Physics (Theory)
Mass quench in interacting finite-size quantum field system
Grachik A Simonian, Alexey G. Mikhaylenko and Andrew G. Semenov
In this work, we theoretically investigate the non-
equilibrium evolution of the correlation functions of a
quantum scalar field following a mass
quantum quench in a finite-size system. The self-consistent equation for the Keldysh (statistical) correlation function is solved numerically in the one-loop approximation. A specific spatio-temporal effect is observed: at a point in time approximately equal to half the system size, the effective mass experiences a strong disturbance. It is shown that this effect is driven by the interference of counter-propagating wave fronts of fluctuations that propagate within the light cone and is induced by periodic boundary conditions.
Read the full abstract ▾
Strongly Correlated Electrons
Strange-metal behavior in correlated molecular conductors
Savita Priya, Maxim Wenzel, Christian Prange et al. (+10 more)
The $\kappa$-BEDT-TTF salts are prototypical Mott systems, realizing a correlation-driven electronic transition from a
Fermi-
liquid state to a Mott
insulator. The metallic compounds with spatially modulated Hg-containing anions, $\kappa$-(BEDT-TTF)$_4$Hg$_{3-\delta}X_8$ ($X$ = Br, and Cl), however, fall outside this classification. Our temperature-dependent transport and broad-range optical investigations reveal a $\rho(T)\propto T$ dependence of the resistivity up to elevated temperatures and a power-law behavior of the frequency-dependent conductivity $\vert\hat{\sigma} (\omega)\vert \propto \omega^{-0.7}$; well-known hallmarks of “strange metals”. The linear-in-frequency scattering rate of these strange metals scales to the Planckian dissipation. We demonstrate that this response only occurs in systems with incommensurate anion layers and establish a direct link to large-scale spatial modulations. Our conclusions can be extended to other strongly correlated strange-metal compounds
Read the full abstract ▾
All authors Savita Priya, Maxim Wenzel, Christian Prange, Bertold Rasche, Christian Sindlinger, Takuya Kobayashi, Hiromi Taniguchi, Dita Puspita Sari, Yasuyuki Ishii, Kazuya Miyagawa, Kazushi Kanoda, Simone Fratini, Martin Dressel
Solar and Stellar Astrophysics
Volume Dependence of Helicity-based Eruption Diagnostics in Recurrent Coronal Jets
Chang Zhou, Jinhan Guo, Yang Guo et al. (+3 more)
Recurrent
solar coronal jets originate from the same source region, providing a useful setting for studying the buildup, redistribution, and release of
magnetic energy and helicity. However, how these quantities can characterize
magnetic evolution and the onset conditions of the eruptions remains unclear. We perform a full three-dimensional (3D) thermodynamic magnetohydrodynamic simulation of recurrent jets driven by converging motions imposed at the bottom boundary. The simulation naturally produces two successive eruptions. The current-carrying helicity $H_J$ decreases during the first jet and gradually recovers between the two jets. In contrast, the free
magnetic energy $E_{\mathrm{free}}$ shows no clear accumulation after its first release. This difference indicates that $H_J$ provides information on
magnetic evolution beyond that provided by $E_{\mathrm{free}}$. Furthermore, the commonly used helicity ratio $H_J/H_V$ reaches different peaks near the onset of the two jets and is sensitive to the integration volume. By comparison, the locally normalized quantity $H_J/\Phi_{\mathrm{closed}}^2$ reaches similar peaks near the onset of both jets and is relatively insensitive to the choice among the tested integration volumes, where $\Phi_{\mathrm{closed}}$ is the single-polarity
magnetic flux of the closed-field domain. These results reveal the local nature of the eruptions and suggest that $H_J$ normalized by the square of the closed
magnetic flux may be a useful diagnostic of eruptive conditions.
Read the full abstract ▾
All authors Chang Zhou, Jinhan Guo, Yang Guo, Jiajia Liu, Zekun Lu, Yuhang Gao
High Energy Physics (Experiment)
Using Geometric Deep Learning for the Detection of the Diffuse Supernova Neutrino Background in Water-Cherenkov Detectors
Matthieu Melennec, Andrew D. Santos, Thomas Mueller et al. (+1 more)
The study of the Diffuse
Supernova Neutrino Background (DSNB) is an essential approach to improve our understanding of the formation of Core-Collapse
Supernovae, as well as investigating fundamental
neutrino properties. The detection of the DSNB flux however is a difficult task, as it requires the disentanglement of its signal with many competitive backgrounds in the low energy range. The main DSNB detection channel is the Inverse Beta Decay (IBD), which is characterised by a distinct, time-separated two-signal signature. This work introduces the use of Graph Neural Networks (GNNs) to handle critical detection and reconstruction tasks for the DSNB through the IBD channel. In particular, we present a model to recover the energy of the first signal from a low energy positron. Additionally, we develop a signal-to-background classifier optimised for the secondary signal, which corresponds to the characteristic
photon emission following a
neutron capture event. We test our models on two simulations in a Hyper-Kamiokande-like water Cherenkov detector geometry, and show that we obtain performances that are comparable to the state of the art.
Read the full abstract ▾
All authors Matthieu Melennec, Andrew D. Santos, Thomas Mueller, Frédéric Magniette
Materials Science
Emergent gold-induced interfacial magnetism in monolayer FePS$_3$
Giuseppe Buccoliero, Anita Polli, Marli dos Reis Cantarino et al. (+10 more)
Owing to their atomically clean interfaces and extreme thinness, two-dimensional (2D) materials are ideal hosts for strong proximity effects, whereby contact with an adjacent material reshapes their ground state. Here, we demonstrate such an effect within a hybrid
heterostructure, between the 2D antiferromagnet FePS$_3$ and an ultrathin gold film. Rather than a passive support, the latter profoundly changes the nature of the magnetism in the 2D magnet, inducing a sizeable interfacial moment. Large-area flakes from Au-assisted exfoliation, combined with element-selective X-ray absorption mapping, enable a layer-by-layer study from monolayer to bulk, benchmarked against an inert SiO$_2$ substrate reference. At 9 T and 5 K, the monolayer on gold shows a three-fold enhanced X-ray
magnetic circular dichroism relative to SiO$_2$, with a $\sim$40% reduction of the orbital-to-
spin moment ratio. Resonant inelastic X-ray scattering reveals a collapse of the
crystal-field (d-d) excitations and a massive spectral-weight transfer to the charge-transfer region, unveiling the microscopic mechanism behind this change. Supported by atomic multiplet calculations, we identify interfacial charge transfer and Fe 3d-Au hybridization as the origin of these effects, which decay over the first three layers and vanish at the pentalayer, establishing layer number combined with substrate engineering as an atomically precise handle.
Read the full abstract ▾
All authors Giuseppe Buccoliero, Anita Polli, Marli dos Reis Cantarino, Rachel Nickel, Aloïs Arrighi, Roberto Sant, Tristan Riccardi, Valérie Reita, Nedjma Bendiab, Laëtitia Marty, Kurt Kummer, Johann Coraux, Nicholas B. Brookes
High Energy Physics (Phenomenology)
Constructing $SU(N)$ strongly coupled dark sectors by extending the Standard Model: mixed representations, flavor structure, global symmetries and various portals
Yi Chung
Motivated by the potential existence of a dark QCD sector, we construct a $SU(N)$ strongly coupled dark sector by extending the Standard Model (SM) gauge group to an enlarged $SU(N+5)\times SU(N+1)\times U(1)$, under which the dark
quark candidates naturally arises from the same multiplets as the SM
fermions. The dark QCD sector from the extension inherits the rich structure of the SM, featuring three distinct properties: (1) Mixed representations: There are dark
quarks of three different representations, including fundamental ($\chi_f$), antisymmetric ($\chi_a$), and symmetric ($\chi_s$), which can be traced back to the
chiral structure of the SM. (2)
Flavor: Each representation features three generations, with their
flavor structures and Yukawa interactions directly related to those of the SM. (3) Global
symmetries: Alongside the conventional dark
baryon number $D$, we identify an exotic dark
baryon number $\widetilde{D}$ from a conserved $(B-L)-\widetilde{D}$
symmetry. These features provide novel perspectives on the structure of dark QCD sectors. Furthermore, the extension of the scalar and gauge sectors introduces various portals between the dark QCD sector and the SM sector. We systematically study these portals, including
quark portal,
lepton portal,
Higgs and $Z$ portal, heavy $Z'$ portal, kinetic mixing portal, dipole portal, and dark pion/
axion portals. With a concrete UV model, their origin become robust, providing additional guidance for experimental searches of the dark sector.
Read the full abstract ▾
Machine Learning
MPGE: A Multi-Perspective Graph Explainer for Molecular Classification Explanation
Mahtab Sarvmaili
Graph neural networks (GNNs) predict molecular properties from chemical graph data, but predictive accuracy does not explain how graph information supports an individual decision. A compact prediction-preserving rationale does not necessarily reveal which changes reverse the decision or which modifications the model tolerates. We propose the Multi-Perspective Graph Explainer (MPGE), unifying factual support, counterfactual sensitivity, and exemplar tolerance for a frozen classifier. The factual view, originally termed prototype (PT), seeks a compact retained edge set with the same label and required confidence. Counterfactual (CF) explanations seek bounded prediction-changing deletions; exemplar (EXE) explanations seek non-trivial bounded deletions that preserve the label and confidence. A shared constrained formulation connects prediction behavior, compactness, and edit cost, while separate objectives generate the three views. Our graph-classification extension of CF-GNNExplainer learns symmetric edge rankings and verifies discrete candidates, recording unsuccessful searches. A separate BBBP fragment backend returns RDKit-sanitized molecules. We evaluate the primary GCN implementation on MUTAG, Mutagenicity, AIDS, COX2_MD, and BBBP using semantic coverage, conditional quality, stability, and runtime. Successful factual masks retained 8.6%–15.5% of input edges on average across
datasets; bounded counterfactual coverage was 4.8%–67.6%, and exemplar preservation coverage was 98.9%–100.0%. Exploratory controls reveal the influence of hard projection and retained node information. Quantitative comparisons and molecular visualizations characterize model support, sensitivity, and tolerance without treating them as validated chemical mechanisms.
Read the full abstract ▾
Signal Processing
Reputation-Graded Wavelet-Packet Subtree Allocation for IoT Sensor Networks
Surya Jayakumar, Arun Narayanan and Indrakshi Dey
Access control in Internet of Things (IoT) sensor networks usually acts above the physical layer: the fusion centre decides which reports to accept, while the
radio of a compromised node keeps occupying the shared channel. This paper places the entitlement in the waveform: wavelet packet division multiplexing (WPDM) builds the channel as a binary tree of orthogonal pulses, and each sensor's reputation sets the depth of the subtree it may use: trusted sensors hold wide subbands, and low-trust sensors hold narrow subbands that the fusion centre withdraws one at a time. We prove that the allocation isolates sensors exactly under perfect timing, and that under a timing offset the energy leaking out of the lowpass leaf falls fourfold per tree level. The locally optimum detector of out-of-subtree transmission in impulsive Class-A noise weights each leaf in proportion to this predicted leakage. We bound the time needed to confine a compromised node and derive the optimal tree depth. In simulation, with demoted nodes radiating within their allocation, the scheme lowers the probability of a false global decision by 68% relative to fixed-allocation WPDM and to OFDM with the same reputation tiers, and by 44% relative to cryptographic re-attestation, which leads once more than 24% of the sensors are compromised.
Read the full abstract ▾