Peer Revenue Equity in science

Papers

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

All of arXiv

Thursday 5 November

Instrumentation and Detectors

Study of neutron flux redistribution and shadowing effect in rod worth measurements with the rod drop experiment

This study investigates neutron flux redistribution and the shadowing effect between control rods during rod-worth measurements using the rod-drop experiment. A simulation-based instrumental calibration methodology is proposed to account for spatial effects on neutron detector responses and improve the accuracy of reactivity measurements. The IPEN/MB-01 research reactor was modeled using the MCNP Monte Carlo neutron transport code. Reactivity behavior during the initial transient of the rod-drop experiment (t < 10 s) was evaluated and compared with experimental measurements using the source-jerk method. The methodology quantifies control rod shadowing and detector-dependent neutron flux redistribution, allowing correction factors to be established for the measured reactivity. Control-bank worth values obtained through the stable-period method were adopted as reference measurements. The reported maximum deviations from the reference values were 0.5% and 11.3%, respectively. The results demonstrate the relevance of spatial neutron flux effects, control rod interactions, and detector calibration in rod-drop experiments, providing a computational and experimental basis for improving control rod worth measurements in research reactors.
Read the full abstract

All authors Giovanni L. Stefani, Guilherme E. S. Carvalho, Daniel A. P. Palma, Marcelo V. Silva, Cesar Raitz

Instrumentation and Detectors

MID: a muon detector for the ALICE 3 upgrade project

The success of the ALICE programme has allowed to study the hottest state of matter created in the laboratory: the strongly-interacting quark-gluon plasma (sQGP). However, some fundamental questions on the sQGP and other aspects of the strong interaction will remain open even after LHC Run 4. To address this, the ALICE collaboration has proposed a major upgrade of its detector, known as ALICE 3. Among the subdetectors considered for this upgrade is the muon identification detector (MID), whose aim is to contribute to the study of charmonium and exotic-hadron production in heavy-ion collisions. While the baseline design consists of plastic-scintillator-based chambers placed around a cylindrical magnetic iron absorber, different detector technologies are under evaluation for the MID, including multi-wire proportional chambers. This proceeding presents the results of two MID prototypes based on FNAL-NICADD scintillator bars, Kuraray fibers and Hamamatsu SiPMs, characterized in test beams at the CERN Proton Synchrotron with 3 GeV/$c$ muon- and pion-enriched beams. The muon efficiency as well as the hadron suppression obtained using machine learning techniques will be discussed.
Read the full abstract

Computational Engineering

Provenance-tracked, experimentally validated neural operator surrogates of three-dimensional laser melting with keyholes and free-form paths

A three-dimensional thermo-fluid-dynamic simulation of laser melting resolves the melt pool and vapor depression but costs from one to thousands of core-hours per track. We train FNO, RIGNO and GAOT neural operators that predict the temperature, melt-pool and free-surface fields. Our datasets comprise 3D temperature and melt geometry fields covering lack-of-fusion to keyhole conditions under varied laser power, scan speed, beam radius, beam shape and preheat, and include free-form scan paths with speed changes and up to four beams. We generate them with FLOW-3D WELD, using solver settings validated against published measurements for four materials. We implement an AiiDA framework to store the provenance of every simulation, model and evaluation, so each tabulated metric traces back to the solver and model inputs behind it. On held-out simulations, the best architecture scores below 2 % relative temperature error on quasi-steady straight tracks, below 4 % on single-beam free-form paths and about 5 % with two to four beams. Inference takes under a second per predicted arbitrary time step on our single-GPU setup. We extend a surrogate trained on Ti-6Al-4V to Inconel 718 by growing its training set autonomously, and the extended models reach a relative temperature error of about 3 % or less on both Inconel 718 and Ti-6Al-4V. At 20 Ti-6Al-4V and Inconel 718 conditions held out of training, the two-alloy RIGNO predicts melt-pool and vapor-depression depths within 30 % of published measurements, and its depths differ from FLOW-3D at the same conditions by 2 % on average. Furthermore, these results provide a scalable framework which can readily be applied to computationally intensive simulations in fields other than laser melting.
Read the full abstract

Genomics

AVA: Agentic Variant Adjudication for Automated, Auditable Read-Level Review in Cancer Diagnostic Sequencing

Identifying a tumor's DNA mutations is central to cancer diagnosis and treatment, but distinguishing genuine mutations from sequencing artifacts is error-prone. In practice, genuine mutations are identified using fixed variant-allele-frequency (VAF) and depth thresholds which cannot separate artifacts from genuine mutations in the borderline low-VAF zone where they overlap. A laboratory specialist must manually inspect the read-level sequencing data, a slow step that has high inter-rater variability and causes diagnosis delays. We present AVA: Agentic Variant Adjudication; an agentic framework that automates this read-level inspection with subagents that reason over evidence, generate investigative code, and return a real-or-artifact adjudications with an auditable reasoning trace all without training on existing data. On both the SEQC2 DNA-seq benchmark and patient RNA-seq data, restricted to a subsample of borderline variants where thresholds fail, both closed-source (Claude Sonnet 5) and open-weight (GLM 5.3 Flash and Gemma 4 31b) agents outperform the best achievable VAF/depth thresholds. Ultimately, AVA is suited for privacy-constrained clinical deployment, since the framework can run on locally deployed open-weight models with descriptive reasoning traces to keep specialists in the loop.
Read the full abstract

All authors Caleb Chin, Sidharth Gupta, Arun K Anguraj Vadivel, Robert Siddaway, Siddharth Arya, Arianna He, Peter Zhukovsky, Cynthia Hawkins

Molecular Networks

Predicting transcriptional interactions with GrnPFN

Combinatorial therapies offer opportunities to improve treatment efficacy and overcome therapeutic resistance, but their rational design requires understanding how combined interventions interact beyond their individual effects. Combined genetic and chemical perturbations can elicit nonadditive transcriptional responses, yet the vast number of possible combinations limits experimental characterization. Predicting these interactions is particularly challenging when few or no combinations have been measured. Here we present GrnPFN, a prior-data fitted network that predicts nonadditive transcriptional responses to combinations from single-perturbation profiles. Pretrained on simulated screens generated from random gene-regulatory networks, GrnPFN incorporates measured combinations through in-context learning without retraining. We evaluated GrnPFN in single-cell screens spanning CRISPR activation, RNA knockdown and drug combinations, using separate cells to construct model inputs, training targets and test targets. Without measured combinations in the target screen, GrnPFN outperformed gradient boosting trained on other screens. With limited combination measurements, it outperformed kernel ridge regression fitted to the same measured pairs. These advantages extended to an additional screen combining CRISPR interference with a drug. Our results demonstrate that pretraining on simulated regulatory networks enables prediction of nonadditive transcriptional interactions across perturbation modalities when experimental measurements are scarce.
Read the full abstract

Representation Theory

Geometric Hammocks: Surface Laminations and the Ziegler Spectrum of String Algebras

We give the first complete topological, geometric, and ergodic classification of the Ziegler spectrum $\Zg(\Lambda)$ of a non-domestic string algebra $\Lambda$ over an algebraically closed field of arbitrary cardinality, via a dictionary with Thurston's theory of geodesic laminations on a cusped hyperbolic surface. This is a model-theoretic counterpart of Haiden–Katzarkov–Kontsevich's realization (Publ.\ Math.\ IHES 2017) of stability conditions on the bounded derived category of a gentle algebra via flat surfaces: pure-injective modules correspond to admissible geodesics, and the Ziegler topology to a corridor quotient of Chabauty space. Three consequences follow. The Krull–Gabriel dimension equals the surface's topological accumulation depth, and the string closure with all higher Cantor–Bendixson strata is globally homeomorphic to an explicit corridor space, with the pushout map to $\Zg(\Lambda)$ a homeomorphism if and only if $\Lambda$ is domestic. Prefix-free dyadic blocks in the geometric perfect core verify Puninski's cardinality-free two-chain criterion, resolving the open problem of superdecomposable pure-injectives over uncountable fields for all non-domestic string and special biserial algebras. Three purely algebraic invariants on the perfect core—localized pp-shape entropy, a $K_0$-weighted pressure function satisfying an algebraic Ruelle–Walters Variational Principle, and a pp-bifurcation index with a Gauss–Bonnet simplicity obstruction—strictly separate uncountable iso-entropic and iso-pressure families, and the inverse limit of the quotient pp-lattice recovers the transversal Cantor cross-section.
Read the full abstract

Graphics

Topology-Aware Reinforcement Learning for CAD Decomposition toward All-Hexahedral Meshing

Automatic hexahedral meshing for complex CAD models has long been regarded as the "holy grail" of mesh generation. The fundamental challenge arises from feature misalignment, where orthogonal or interacting features coexist, making direct hexahedral meshing infeasible and requiring decomposition into simpler, meshable blocks. While existing decomposition methods, including frame-field pipelines and learning-based techniques, have reduced manual intervention, they can remain sensitive to geometric and topological complexity and may fail to produce valid decompositions for challenging feature configurations. Specifically, their action spaces are typically restricted to surface-based splitting, which often yields invalid or unmeshable decompositions. To address these limitations, we propose DecompRL, a deep reinforcement learning framework for CAD decomposition toward hexahedral meshing. We design a set of topology-aware internal-splitting actions that leverage both global topological structure and local geometric property to effectively separate misaligned features. Furthermore, we introduce geometric rewards together with meshability feedback to guide the learning process toward valid decompositions and high-quality meshes. Extensive experiments on ABC, Fusion360, NIST, and DeepCAD show that DecompRL achieves the highest success rate among the evaluated learning-based methods on three of the four benchmarks, while maintaining competitive mesh quality. The measured LMFF-to-DecompRL runtime ratio reaches up to two orders of magnitude; this compares the full LMFF procedure with DecompRL decomposition alone, excluding subsequent mesh generation. The largest SR improvement is observed on Fusion360, where DecompRL reaches a success rate of 59.6%.
Read the full abstract

All authors Jiaming Peng, Zhihao Zheng, Rui Wang, Qinglin Wang, Jing Xiao, Jie Liu, Na Lei, Xinhai Chen

Computation and Language

Language Defaults Before Content: A Causal Layer Map of Multilingual Generation

Multilingual generation requires large language models (LLMs) to determine both what to say and in what language to say it, yet these two commitments are difficult to disentangle from model outputs alone. We use translation as a controlled probe: for multi-parallel segments that a model translates well in one language direction but poorly in another, content is held fixed while language varies. Using activation patching, we write residual-stream states from successful runs into failing ones and compare content-matched states with content-independent controls, allowing us to trace content-specific and language-related signals across layers. Across three LLMs of different scales and architectures, we find a consistent three-stage pattern of multilingual generation. Intermediate layers first express a default English signal; content information becomes causally available one layer later; and the final target language is resolved only at later layers. The ordering of the English default before content also appears in a reading-comprehension task. Our sparse autoencoder analyses reveal a further dissociation: the intermediate language signal is only weakly represented by sparse features, whereas a later language distinction is clearly sparse, but resists intervention at the patched position. These results provide a mechanistic account of how language selection and content information emerge and interact during multilingual generation.
Read the full abstract

Fluid Dynamics

Dynamics of singular-jet formation during the collapse of drop-impact craters

The collapse of a drop-impact crater that bypasses the pinch-off singularity produces exceptionally thin, high-velocity jets, often referred to as singular-jets. We perform 2D axisymmetric DNS using Basilisk, employing adaptive mesh refinement to capture the intricate flow features. The numerical results for a drop diameter of $3.64$ mm at ambient pressure match the experiments well under different impact conditions. During the crater collapse, qualitatively different dimple shapes form, including an entrapped bubble beneath the dimple. Quantitatively, an entrapped bubble with a diameter of ~210 \mu m and a jet velocity of 61.7 m/s for the singular-jet case during cylindrical dimple collapse is obtained from simulations. Subtle morphological changes in crater evolution distinguish the jetting regime, with a critical angle of 109.2 between the crater side wall and its base in the singular-jet case. The simulation provides insight into the internal flow fields, enabling us to delineate the observed phenomenon. The pressure and velocity fields reveal that the high jetting velocity is associated with the formation and rapid collapse of a nearly cylindrical dimple. The resulting pressure distribution drives a strongly upward flow near the dimple base, while the nearly parallel dimple side walls redirect radial liquid motion towards the axis. We extend the simulations into the post-entrainment region and identify another singular-jet near the upper boundary of the entrainment region. The jet reaches a velocity of 63.93 m/s and exhibits morphological characteristics similar to the first singular-jet case. A regime map is constructed from simulations spanning Weber numbers ranging from 100 to 200 and Froude numbers from 45 to 92; the updated lower and upper boundaries for the entrainment region, based on singular-jet formation, are We = 45.5\, Fr^{0.237} and We = 53.5\, Fr^{0.251}, respectively.
Read the full abstract

Distributed Computing

HeteroTask: Multi-Variant Task Scheduling and Execution on Distributed Heterogeneous Clusters

Heterogeneous systems that combine CPUs and GPUs across multiple servers offer substantial compute capability but complicate scheduling. For DAG-based applications with multiple task implementations, effective execution requires coordinating variant selection, device placement, and data movement. Offline planning can exploit global dependency information, yet plans derived from isolated kernel profiles may diverge from execution because of runtime overheads. We introduce HeteroTask, a framework integrating kernel-wise profiling, PEFT-based multi-variant planning, host-buffer lifetime modeling, and schedule-driven MPI execution. A Bayesian-optimization loop calibrates empirical overhead corrections for repeated, fixed-DAG workflows on exclusively allocated nodes. We evaluate five benchmark and workflow-derived applications on up to six 8-GPU nodes. On two 8-GPU nodes, HeteroTask reduces application-level completion time by 19.7-72.1% relative to the compared StarPU implementations. For Canny Edge Detection, calibration reduces task-window makespan by 25.0% relative to planning without overhead corrections. Combined with CPU/GPU variant selection, calibration reduces it by 58.9% relative to an uncalibrated GPU-only execution.
Read the full abstract

All authors Hongzheng Tian, Alok Mishra, Haocheng Xu, Dejan Milojicic, Sitao Huang

Artificial Intelligence

AYO: A Safe and Privacy-Preserving Platform for Studying Children's Interaction with LLMs

Children and adolescents are increasingly interacting with generative AI and large language models (LLMs), yet little is known about how these interactions may influence their development, beliefs, and behavior. Existing research has relied heavily on surveys, while direct study of children's interactions with LLMs presents substantial challenges related to safety, privacy, consent, and experimental control. We introduce AYO (AI Youth Open Science), an open-source platform designed for safe and systematic research on children's interactions with conversational AI. AYO provides a child-friendly interface, configurable age-based safety guardrails, researcher monitoring, and privacy-preserving data practices. These capabilities enable cross-sectional, longitudinal, and experimental research on how children use and respond to LLMs and how AI design may shape trust, epistemic judgment, curiosity, and other aspects of development. AYO provides the infrastructure to generate evidence that guides safer AI systems, educational practices, and policies for children growing up with AI.
Read the full abstract

All authors Huiwen Alex Yang, Rachel Skwersky, Bill D. Thompson, Celeste Kidd

Functional Analysis

On the dynamics of weighted shifts on Fréchet sequence spaces over trees

In this paper, we give characterizations of Li–Yorke chaos, mean Li–Yorke chaos and distributional chaos for weighted backward shifts acting on general Fréchet sequence spaces over directed trees. As corollaries, we obtain explicit characterizations in the setting of Köthe spaces $\lambda_p(A,V)$, where $p\in\{0\}\cup[1,\infty)$ and $V$ is a directed tree. In particular, we obtain criteria for the case of the classical Banach spaces $\ell^p(V)$, $p\in[1,\infty)$, and $c_0(V)$.
Read the full abstract

Information Theory

Cross-Domain Channel Estimation for MU-MIMO-OFDM

Conventional channel estimation for multiple-input multiple-output orthogonal frequency-division multiplexing (MIMO-OFDM) relies on the pilot interpolation without explicitly identifying the sparse delay-angle (DA) components. To exploit DA sparsity while maintaining backward compatibility with existing MIMO-OFDM systems, we propose a cross-domain channel estimation scheme for multiuser MIMO-OFDM (MU-MIMO-OFDM) systems. The scheme exploits the circular convolution structure in the DA domain to identify candidate DA pairs by correlating delay-domain observations with DA pilot responses. These pairs and the known pilots then define a time-frequency (TF) dictionary for refining the DA coefficients from TF observations through least-squares or sparsity-regularized fitting. The refined coefficients and candidate pairs reconstruct the aggregate channel across antennas and subcarriers aligning with the geometric channel model. Numerical results show approximately 10 dB and 7 dB lower normalized mean-square error than the conventional OFDM estimation schemes for integer and fractional delay and angle indices, respectively.
Read the full abstract

All authors Mingcheng Nie, Shuangyang Li, Ruoxi Chong, Giuseppe Caire, Yonghui Li

Machine Learning

All roads lead to Markov kernel networks: A representation theorem for discretization-stable learning

Operator learning seeks to approximate function-to-function maps in a way that remains stable when the input discretization changes, a property sometimes called "discretization invariance". We consider a quantitative refinement of this idea, called $\textit{uniform discretization stability}$, which builds on a common representation of input functions and their discretizations as probability measures and provides a Hölder bound relating output changes to the Wasserstein distance between these measures. Our main result is a representation theorem showing all uniformly discretization stable operators can be written exactly as a $\textit{stable Markov kernel network}$—an abstract architecture that is a joint generalization of transformers, Fourier neural operators (FNO), Flowers, and other successful learned operators. Our representation theorem characterizes a design space for discretization-stable operator learning; designing new architectures becomes a question of which parts of a Markov kernel network to parameterize, and how. We further show that the read kernels of these networks can be realized as solutions of regularized variational problems, with particular choices of costs and regularizers recovering the kernels of transformers, FNO, Flowers, their hybrids, and beyond. This variational formulation provides a systematic approach to architecture design, together with sufficient conditions for uniform discretization stability.
Read the full abstract

All authors S David Mis, Trevor Teolis, Ivan Dokmanić, Matti Lassas, Maarten V. de Hoop

Machine Learning

Selected-Risk Calibration for Feasibility at Returned Decisions

When an optimization problem uses a surrogate constraint, a decision that is feasible for the surrogate need not be feasible for the true constraint. We propose Selected-Risk Calibration (SRC), a framework that calibrates a learned constraint against the optimizer's returned decisions for objectives drawn from a given distribution. SRC guarantees that, with probability at least $1-\delta$, every margin it certifies on held-out objectives has selected risk, the probability that the returned decision is infeasible, of at most $\gamma$. We also show that the random-input coverage guarantee of split conformal prediction does not bound the selected risk, and that likelihood-ratio reweighting for feedback covariate shift does not apply when the learned constraint is active at the returned decision. Our experiments show that a conformal margin covering about $90\%$ of random inputs covered as few as $17\%$ of the returned decisions, and every certified margin had a violation rate below its target on unseen objectives. Moreover, at equal cost, SRC's correction had fewer or equal infeasible solutions to the baselines in 57 of 62 comparisons. The SRC framework applies to any setting in which a surrogate model chooses a decision whose feasibility can be verified for a sample of instances.
Read the full abstract

All authors Raghav Thind, Youran Sun, Ling Liang, Haizhao Yang

Artificial Intelligence

LoRazor: Trimming Adapter Execution for Efficient Multi-LoRA Serving

LoRA offers a parameter-efficient adaptation strategy, supporting multi-user workloads via customized adapters on a shared backbone. However, serving concurrent requests with different adapters introduces substantial computation and memory-management overhead. We ask the research question: Do all requests need their adapters throughout the whole generation process? Our motivational study reveals that employing LoRA adapters at every decoding step is unnecessary, releasing them early has limited impact on accuracy while substantially accelerating generation. Although skipping adapters early offers theoretical workload reduction, translating these computational savings into concrete throughput gains requires tailored GPU kernels. We present LoRazor, a multi-LoRA serving framework that combines early adapter release with active-row aware GPU kernels. LoRazor determines when to permanently release a request's adapter and continue decoding with the backbone alone. By dynamically resizing LoRA kernel launches as adapters are released, LoRazor trims idle GPU blocks. Across two backbones and four tasks, LoRazor achieves up to 3.02x request throughput and 2.53x output-token throughput over vLLM multi-LoRA, with a mean relative task-accuracy decrease of 2.89%.
Read the full abstract

All authors Xiang Li, Pengcheng Wang, Azam Ikram, Saurabh Bagchi

Artificial Intelligence

Veterans' Mental Health Crisis Severity Annotation: Do LLMs Agree More Than Human Experts?

Large language models (LLMs) offer a scalable approach for annotating mental health and crisis-related social media content, but their reliability in safety-critical tasks remains unclear. This study evaluates LLM-based crisis-severity annotation of veteran-related social media posts against annotations from eight Veteran Subject Matter Experts (SMEs). Using 149 tweets and five severity categories, we evaluate 15 open-source and 10 proprietary LLMs across eight prompt variants, examining annotation performance, inter-model agreement, prompt sensitivity, and high-severity crisis detection. Human reviewers showed higher ordinal than nominal agreement ({\alpha} = 0.494), indicating that disagreements often occurred between neighboring severity levels. In contrast, selected LLMs demonstrated substantially higher inter-model consistency, with ordinal agreement ({\alpha}) reaching 0.8336 under the baseline prompt. Proprietary models generally showed better agreement with expert annotations and greater stability than open-source models. GPT-5.6-Terra achieved the best overall performance, while the open-source Qwen2.5-7B showed the strongest performance in identifying Acute Crisis with Suicidal Ideation cases. However, higher model consistency did not necessarily correspond to agreement with human judgments or reliable detection of rare high-severity cases. Label-flipped demonstration experiments also revealed model-specific sensitivity to misleading in-context examples, while explanatory demonstrations did not consistently improve performance. Overall, LLMs show promise as scalable complements to expert annotation, but require careful evaluation before use in high-stakes mental health applications.
Read the full abstract

All authors K M Sajjadul Islam, Pranav Yaramalla, Nadiyah Johnson, Mark Flower, Praveen Madiraju, Zeno Franco

Computers and Society

Public Perceptions of Public Participation in Local Government Mediated through a GenAI Tool

City council meetings offer a site for direct democracy through public comment periods. Yet, this process can lack transparency, and research on local political participation has documented that people who speak at public meetings are often unrepresentative of their communities. However, a lack of representation does not establish whether other people view these participants as advocates for their concerns. In fact, little is known about how they are perceived. To address this gap and improve transparency into this democratic process, we built a generative AI (GenAI) tool, VoicesView, that allows individuals to retrieve and read public comments. We deploy this tool, along with a novel survey, in a study with 72 participants in Ann Arbor, Michigan, and Seattle, Washington, in order to understand people's perceptions of public commenters. We find that participants generally viewed public commenters favorably, and that these views became more positive after using VoicesView. For example, the share who agreed that public commenters advocate for issues they care about rose from 49% to 62%. Knowledge of how public comment works in their city was moderate and also improved: the proportion of participants who correctly answered a procedural knowledge question about council meetings increased from 53% to 74%. Participants viewed public commenters more favorably than local government, yet their sense of being heard by local government also grew: agreement that their own voice is heard rose from 28% to 46%. These findings suggest that giving residents opportunities to engage with and learn from records of local political participation can foster positive civic outcomes.
Read the full abstract

All authors Jason Yan, Yanna Krupnikov, Nicholas Weber, Sabina Tomkins

Machine Learning

Topological Deep Learning Challenge 2026: Bridging the Gap

This paper describes the 2026 edition of the Topological Deep Learning Challenge: Bridging the Gap, hosted at the second Topology, Algebra, and Geometry in Data Science (TAG-DS) Conference. Where previous editions grew the modelling and data ecosystem of Topological Deep Learning (TDL), this edition set out to place Graph Neural Networks (GNNs) and Topological Neural Networks (TNNs) side by side under a single evaluation pipeline. Participants implemented recent state-of-the-art architectures in TopoBench, with Track 1 covering GNNs and Track 2 covering TNNs. They evaluated these implementations on shared synthetic benchmarks generated with GraphUniverse, which independently controls homophily, average degree and degree-distribution heaviness. The challenge yielded 68 qualifying submissions from 34 teams, spanning the graph, simplicial, cellular, hypergraph and combinatorial domains. For the first time, submissions were also tracked on a public leaderboard, which we summarize and analyze here. Our analysis finds that the differences between the submitted tracks are strongly task-dependent: the mean accuracy gap is modest on community detection, while on triangle counting the Track 1 mean error is 1.85 times the Track 2 mean. Several structurally informed GNNs also outperform the Track 2 mean on triangle counting.
Read the full abstract

All authors Guillermo Bernárdez, Lev Telyatnikov, Mathilde Papillon, Marco Montagna, Louisa Cornelis, Louis Van Langendonck, Antonio Purificato, Bruno Neri, David Leko, Donatella Genovese, Eleonora Grassucci, Jan-Willem Van Looy, Johan Mathe, Lucia Testa, Luka Benić, Melanie Weber, Olga Zaghen, Pavlo Vasylenko, Raffael Theiler, Scott Mahan, Gaurav Khanal, Valentina Sánchez, Dario Loi, Gabriele Onorato, Prasanna Devadiga, Taejin Paik, Anier Velasco Sotomayor, Jonas Müller, Aarav Gupta, Riyan Patel, Grold Otieno Mboya, Pushparaj Devadiga, Adam Gohain, Dominik Filipiak, Michael Banf, Alireza Mohammadian, Amirreza Akbari, Gabriel Chaviano Díaz, Marcos Chávez Mesa, Ankit Grover, Rémi Bourgerie, Alessio Comparini, Ankit Kumar, Mohammad Yusuf, Devesh Kaushal, Francesco Orsi, Ghazal Mostofi, Kolya Lettl, Salwa Khattami, Gabriel Franco, Yelizaveta Falkouskaya, Jorge Luiz Franco, Kajal Singla, Fernando David Espinosa Iñiguez, Raj Patil, Eric Dolores-Cuenca, Luigi Petti, Chenguang Xu, Stefano Blando, Christine Zeh, Lennart Bastian, Kaden Zheng, Vihan Tiwari, Sud Athreya, Hisataka Kajita, Hugo Micheron, Rémi Devaux, Antoine Jardin, Tegan Emerson, Olga Fink, Nina Miolane

Artificial Intelligence

Don't Blame The Preferences: Examining The Emergence of Sycophancy in LLM Training

AI sycophancy, the tendency of large language models (LLMs) to flatter users and affirm their beliefs, can undermine accuracy and contribute to misinformation, dependence, and overconfidence. Although prior work has benchmarked sycophancy in released models and proposed mitigations, its origins during training remain poorly understood. Preference alignment methods such as RLHF and DPO are often blamed, but their effects are rarely compared with earlier training stages. We investigate sycophancy across the training pipelines of three model families and three benchmarks. We find that sycophancy is largely established by the end of supervised fine-tuning (SFT), rather than introduced by preference alignment. Across all nine model benchmark combinations, SFT checkpoints attain at least 92% of preference aligned models' sycophancy scores and at least 85% of final models' scores; on average, they are about as sycophantic as final models. Preference alignment increases sycophancy in only four of nine cases, by an average of 4.2 percentage points when it does. To probe the underlying causes, we isolate model representations associated with sycophancy using mechanistic interpretability and measure their activation on training datasets from each stage. SFT datasets show roughly the same average activation as other training data but contribute substantially more examples, and the highest-scoring individual datasets all come from SFT. Our findings suggest that addressing sycophancy requires interventions throughout the training pipeline, particularly in data selection and augmentation for pre-training and SFT.
Read the full abstract

All authors Chao Peter Yang, Catherine Zehr, Pranav Bijith, Brandon Fain

Computation and Language

CulturalCrosswords: A Multilingual Benchmark of Answer Constraints and Grid Completion

When several answers fit a riddle, exact-match scoring can penalise a valid response. CulturalCrosswords tests how answer constraints affect large language models (LLMs) solving cultural riddles, whose solutions draw on knowledge specific to a community. We constructed 53 crosswords from crowdsourced riddles in 42 languages and evaluated 11 LLM configurations on five versions of each clue set: plain batches, batches with lengths, batches with lengths and partial-answer hints, and crosswords requesting an answer list or a completed grid. On 962 clues with alternative answers, lengths improve agreement with the reference by 22.8 percentage points, compared with 8.4 points when accepted alternatives also receive credit. Part of the gain therefore appears to come from selecting a required answer form. Crossword input increases the share of all crossings with compatible answers by 9.7 points; requiring a completed grid reveals large differences between models. The benchmark distinguishes agreement with a reference from the ability to produce answers that fit together and occupy the correct cells.
Read the full abstract

All authors Isaac Saxton-Knight, Gabriel Saxton-Knight, Diego Oliveira, Tidor-Vlad Pricope, Aston Garratt, Jocelyn D'Arcy

Image and Video Processing

Covariate-Conditioned One-Step Diffusion for Multi-Horizon 3D Brain MRI Prediction in Aging and Neurodegeneration

Predicting longitudinal brain changes from magnetic resonance imaging (MRI) can help characterize structural progression associated with aging and neurodegeneration. However, diffusion-based forecasting typically requires iterative denoising, limiting inference efficiency. We propose a covariate-conditioned one-step diffusion framework for multi-horizon prediction of 3D gray-matter density (GMD) maps. Rather than directly generating future anatomy, the model predicts longitudinal GMD changes through residual prediction in a single denoising step. The predicted changes are added to the baseline GMD map to reconstruct future anatomy. The model is conditioned on baseline anatomy, age, diagnosis, APOE4 allele count, and continuous follow-up time. Training combines mean squared error with a Pearson correlation loss to improve the spatial agreement between predicted and observed GMD changes. A unified model supports prediction horizons from 2 to 10 years. On the held-out ADNI test set, our method achieves a $\Delta$-Pearson of 0.478, mean squared error (MSE) of $5.176 \times 10^{-3}$, and structural similarity index measure (SSIM) of 0.979, with $\Delta$-Pearson reaching 0.582 at 10 years. Without retraining, external evaluation on AIBL yields a $\Delta$-Pearson of 0.450, MSE of $4.016 \times 10^{-3}$, and SSIM of 0.985. Ablation studies further demonstrate the benefits of residual prediction, covariate conditioning, and correlation-based supervision. Code: https://github.com/natenaile/OneStepBrainMRI.
Read the full abstract

All authors Natenaile Asmamaw Shiferaw, Ali Farki, Deepika Koundal, Jussi Tohka

Mathematical Physics

Positivity Improving Quantum Markov Operators

A Markov operator on a matrix algebra, that is, a unital completely positive map, is called positivity improving if it maps every nonzero positive element to an invertible one. This is the noncommutative counterpart of a stochastic matrix with strictly positive entries. We show that positivity improvement is self-dual and equivalent to a faithful quantum Markov–Dobrushin–Doeblin minorization. We also compare the scalar trace minorization with the optimal common-lower-bound construction of the quantum Markov–Dobrushin coefficient. We then characterize positivity improvement through the Kraus space of the map: a Markov operator is positivity improving if and only if the Hilbert–Schmidt orthogonal complement of its Kraus space contains no operator of rank one. As a consequence the Kraus rank of a positivity improving Markov operator on the algebra of $d\times d$ matrices is at least $2d-1$, and an explicit Hankel-type channel shows that this bound is attained in every dimension. Although deciding positivity improvement is NP-hard in general, we obtain complete and effectively checkable classifications for qubits and qutrits, together with families of examples (Pauli, Weyl and Werner–Holevo channels) showing that all our estimates are optimal. Finally we show that, in infinite dimensions, positivity improvement no longer implies any contraction of the Markov operator.
Read the full abstract

Accelerator Physics

Physics in Safety and Safety in Physics: Lessons from the LHC and its Underground Experiments

Large underground particle-physics facilities are usually discussed by separating scientific performance from the systems that make their operation safe. We develop a complementary view in which safety is treated as an intrinsic part of the physical and operational structure of an experiment. Using the LHC, ATLAS, and CMS as principal case studies, we introduce the concepts of physics in safety and safety in physics: physical laws, measurements, and models determine hazards and protective barriers, while those barriers in turn define the machine, detector, access, and intervention states in which physics can be performed. We formulate this relation through a safe operational phase space and through admissible trajectories connecting different configurations, emphasizing that safety constrains not only states but also transitions between them. The framework is applied to beam and magnet protection, cryogenic and oxygen-deficiency hazards, detector-specific risk configurations, personnel access, maintenance, and radiation protection, with particular attention to the different architectures of ATLAS and CMS. We argue that validated protection does not merely restrict scientific activity but can enlarge the domain of sustainable operation. This viewpoint provides a unified physical and operational framework for understanding safety at the LHC and for incorporating protection, maintainability, and recoverability into the design of future accelerator and underground facilities such as HL-LHC, FCC and muon collider.
Read the full abstract

Software Engineering

SPRING-5: Measuring Tool Documentation Repair for LLM Agents by Efficiency

Methods that repair tool documentation automatically are mostly evaluated on task accuracy. We report a benchmark that scores them on cost. SPRING-5 is 20 tool families and 344 tasks, each with a deliberately degraded specification, a hand-written honest one, a deterministic backend and an answer key, scored on calls charged against the ideal call path. A wrong answer is charged the full call budget, so this cost index largely tracks the completion rate, and we report calls per correct task beside it, which carries no such charge. Over 2,644 baseline sessions, 54% of wrong answers arrive with no rejected call, no refusal and no retry. The benchmark carries three validity instruments. A build-time ceiling assertion requires the honest specification to reach the ideal path and fails systematically on only one family. A dud rule names the families a stronger model solves cold and catches four. A directly measured noise floor shows that two evaluations of a byte-identical specification differ by a 95% factor of 0.64 to 1.56, so about one comparison in eight crosses a conventional win threshold on noise alone, and the cheap way of estimating that floor is biased by which families happen to go unedited. Against these instruments, a schema-only method closes 1% of the baseline-to-ideal cost gap, an online self-evolution loop 13%, one batch rewrite over every training transcript 33%, and a variant given correctness labels 58%. The batch advantage over the loop is the evidence it reads: held to the loop's exposure it is no longer detectable. With five and ten tools in context the effect survives, the baseline cost does not change, and only machine-written descriptions sent calls to the wrong tool. On BFCL, 62% of held-out queries are already answered from the shipped documentation, which explains a pooled null there. Code, data and runs are released.
Read the full abstract

Data Structures and Algorithms

Efficient classical simulation of Fermi-Hubbard models with Kramers symmetry

Quantum Monte Carlo methods are widely used to solve quantum many-body problems. However, rigorous guarantees for their performance are poorly understood, especially for simulating interacting fermions. In this work, we develop a quantum Monte Carlo algorithm for a broad family of fermionic systems exhibiting Kramers time-reversal symmetry, and prove that their partition functions can be estimated in polynomial time. This family, which contains many important classes of Hubbard-type models, is known to both avoid the sign problem and obey a Lee-Yang zero-freeness theorem. Thus our results suggest that these two properties together may enable efficient classical algorithms for quantum many-body systems more broadly.
Read the full abstract

All authors Chaithanya Rayudu, Jun Takahashi, Ojas Parekh, Andrew Zhao

Cryptography and Security

Towards a Privacy-Preserving Reliability-Aware Entity Resolution for Attributed Graphs [Vision] (Technical Report)

Entity resolution (ER) on attributed graphs aims to identify nodes across multiple graphs that correspond to the same real-world entity and merge them into an integrated graph. When input graphs vary in quality and contain sensitive information, ER faces a dual challenge: while reliability information is crucial for accurate graph data integration, its presence in the published integrated graph can heighten privacy risks. This paper presents a research vision that unifies reliability assessment and privacy preservation within a single graph ER process. We highlight the inherent tension between these two objectives, propose a three-level coupled framework, and outline the remaining research challenges to enable the publication of integrated graphs that are both accurate and privacy-preserving.
Read the full abstract

Cryptography and Security

Toward Safety-Aware AI Agents in Realistic Social Engineering Attack Scenarios

Personal AI agents are increasingly deployed on user devices to execute actions involving real money and sensitive data, but their security under social engineering attacks has not been systematically evaluated. We propose SEA (Social Engineering Attack Benchmark), the first interactive environment for evaluating the intrinsic security capability of AI agents in social engineering attack scenarios, covering 8 daily-life domains, 30+ simulated applications, a 4-level difficulty gradient, and a 5-dimensional evaluation framework, with 320 evaluation tasks designed in consultation with anti-fraud specialists and spanning two culturally distinct settings in Chinese and English. Evaluation on 11 frontier open-source and closed-source models, with GPT-6 evaluated only in Chinese, reveals that all models are breached to varying degrees (Chinese attack success rates ranging from 16.9% to 84.3%) proactive verification rates universally below 32%. We further propose META-SEA, a co-evolutionary framework for attack tasks and defenders. With META-SEA training, Qwen3-8B achieves defense performance comparable to closed-source models while effectively reducing ungrounded action claims.
Read the full abstract

All authors Zhitao He, Chengru Han, Chi Him Zhang, Yi R Fung

Plasma Physics

Magnetohydrodynamic spectra and line invariants in the presence of axions

A new form of the mean-field equations for magnetic fields, which accounts for the interaction with axions, is studied. It is shown that the asymptotic ergodic invariants of magnetic field lines play a key role in the evolution of magnetic fields, which determine the type of the magnetohydrodynamic spectrum of the solutions.
Read the full abstract

Software Engineering

Ninada: Sub-Clause Split-Graph VITS Streaming and Capability-Routed 24 kHz Speech Synthesis Across Twelve Scripts

End-to-end VITS text-to-speech models synthesize natural speech in a single forward pass, yet standard ONNX exports execute monolithically over an entire clause, delaying Time-To-First-Audio (TTFA) to several seconds and spiking activation memory. Client-side deployment across Indic scripts faces four further bottlenecks: 138.8 MB float32 checkpoints per language strain browser caches, mobile WebGPU stalls for 6-14 s compiling convolution shaders while still running slower than real time ($1.44\times$ RTF), small WebAssembly pthread pools stall at barriers on big-LITTLE mobile CPUs, and mobile Web Audio pipelines drop frames or mute output under AudioWorklet contention and iOS WKWebView rules. We present Ninada, an open-source 24 kHz TTS system covering English, Classical Sanskrit, and ten Indian languages (sa, en, hi, kn, mr, bn, ta, te, ml, gu, pa, or) across a capability-routed 48-checkpoint fleet: 24 studio HiFi-GAN v1 specialists (27.9 MB W8A32+FP16 ONNX) for desktop/WebGPU and 24 distilled c8c8_384 iSTFTNet specialists (21.7 MB ONNX, $2.64\times$ fewer vocoder MACs via an exact NOLA dual-window ConvTranspose1d inverse STFT with zero complex ops) for mobile WASM, plus a 0.0 MB zero-dependency Klatt fallback. At runtime, Ninada splits a single ONNX file in memory in $< 15$ ms into prior, flow, and windowed vocoder sub-sessions streaming 12-frame halos with raised-cosine crossfades. On $N = 480$ held-out 24 kHz studio utterances, Ninada achieves 3.22 dB 0-8 kHz MCD (3.09 dB mobile iSTFTNet; 1.33 dB 24-voice distillation MCD, vs. 5.17 dB for the 1,526.8 MB 16 kHz Meta MMS fleet) and 6.07 dB 8-12 kHz error, delivering 89-155 ms audible startup on desktop WebGPU and 538-649 ms zero-underrun streaming at 0.75-0.76x RTF on a physical Google Pixel 10 Pro XL.
Read the full abstract