Computer Vision
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
All of arXiv
Monday 9 November
Accurate and timely identification of road closures is critical for large-scale navigation and logistics systems, yet remains challenging due to the need to jointly reason over visual evidence, spatial context, and road network topology. Existing approaches either rely on manual map updates or focus on detecting closure-related objects without resolving which road segment is affected. In this work, we introduce a unified multimodal and multitask framework that jointly performs road closure classification (no, partial, complete) and road segment association within a single learning architecture. Our model integrates graph-structured map data, satellite imagery augmented with projected street-level detections, and street-view images using modality-specific encoders and cross-modal transformer attention, enabling explicit reasoning over visual cues and network topology. Evaluated on nearly 20K real-world scenarios collected from a large-scale commercial vehicle fleet, our approach achieves 72.7% closure classification accuracy and 60% road segment association accuracy, outperforming frontier vision-language models by over 28 percentage points on closure classification while being over 4,000x smaller and 100x faster at inference. Extensive ablations demonstrate the importance of cross-modal attention, edge-aware pooling, and graph message passing for accurate geospatial reasoning. This work establishes the first benchmark for joint road closure classification and segment-level attribution, and shows that task-specific architectural inductive biases substantially outperform general-purpose multimodal models for structured geospatial understanding.
Computer Vision
Comparative Evaluation of Deep Learning Models for Geographic Atrophy Segmentation in En Face OCT Images
Background: Geographic atrophy (GA) is an advanced form of age-related macular degeneration (AMD). It causes irreversible vision loss due to progressive degeneration of the retinal pigment epithelium (RPE) and photoreceptors. Accurate segmentation of GA in en-face image is crucial for early diagnosis, monitoring disease progression, and developing personalized treatment strategies. However, existing deep learning approaches are typically evaluated using single architectures and single-device datasets, limiting insight into relative performance and cross-device robustness. Methods: We trained and systematically compared five deep learning architectures for GA segmentation: Nested U-Net, ResU-Net, nnUnet, Swin-Unet and MedSegDiff model, together with a thresholding based classical baseline. The models were trained on a dataset of en-face OCT images derived from 269 OCT volumes from two different devices, following a three-step preprocessing technique involving normalization and inpainting to enhance the quality of the en-face images. Results: On the fixed held-out validation set comprising OCT scans from both Cirrus (R) (Carl Zeiss Meditec, Inc., USA) and Spectralis (R) (Heidelberg Engineering, Germany) devices, nnUnet achieved the highest mean Dice similarity coefficient (DSC) of 0.8882 +/- 0.0033. The other models yielded the following Dice scores: Nested U-Net (0.7932 +/- 0.0423), ResU-Net (0.8453 +/- 0.0074), Swin-Unet (0.6886 +/- 0.0294), and Diffusion model (0.7959 +/- 0.0242). Conclusions: nnUnet framework, demonstrated most reliable performance for GA segmentation in en-face OCT under limited data conditions. It offers the best balance of accuracy, stability, and efficiency for GA quantification.
Read the full abstract
Quantitative Methods
EvoSurv: Evolution-Aware Multimodal Survival Modeling Reveals Complementary Prognostic Signals
Accurate cancer survival prediction remains challenging due to tumor heterogeneity. Although multimodal survival models integrate clinical and molecular data, they often overlook evolutionary information encoded in genomic alterations. To address this gap, we developed eTime, a 38-dimensional evolutionary representation comprising eight feature groups that characterize clonal and subclonal mutation burden, mutational spectra, functional consequences, copy-number alterations, and anoikis-related genomic changes. We further developed EvoSurv, a multimodal survival framework integrating eTime with clinical variables and RNA expression. Across six TCGA cancer cohorts comprising 2,782 patients, EvoSurv achieved the highest mean concordance index (C-index) in all six cancers, averaging 0.663 and outperforming three state-of-the-art survival prediction methods. More importantly, we assessed whether evolutionary features provide additional prognostic information beyond clinical and transcriptomic predictors. eTime remained significantly associated with survival in head and neck squamous cell carcinoma (HNSC) and liver hepatocellular carcinoma (LIHC) after adjustment for clinical variables and RNA-derived risk scores. Feature-group analyses further revealed cancer-specific prognostic patterns. In HNSC, T>G substitutions were associated with increased mortality risk after clinical and molecular adjustment, including in HPV-negative patients, with supporting evidence from an independent CPTAC-HNSCC cohort. Overall, these findings demonstrate that eTime captures survival-associated evolutionary information and highlight its potential to provide prognostic value beyond clinical and transcriptomic predictors.
Read the full abstract
Theoretical Economics
On the Nonemptiness of the Core: Connecting Cooperative Games and Asset Pricing
We show that the Bondareva–Shapley theorem for infinite cooperative games can be derived from a version of the fundamental theorem of asset pricing. The link is the Choquet extension of a game, viewed as a possibly nonlinear pricing rule on the space of measurable simple payoffs. Under this identification, balancedness translates into no arbitrage, and a linear functional dominated by the pricing rule yields an allocation in the anti-core; the core statement follows by duality. Applying the arbitrage-free pricing theorem of Cornet (2025), we recover the characterization of the nonvacuity of the core initiated by Schmeidler (1967) and established in full generality by Bartl and Pinter (2023) for bounded-below games. Unlike the earlier derivation in Cornet (2025), our argument does not require the game to have bounded variation or the collection of coalitions to be a $\sigma$-algebra.
Read the full abstract
Distributed Computing
One-Sided Expert-Parallel All-to-All for Mixture-of-Experts Inference over NVLink
Expert parallelism (EP) shards the experts of a Mixture-of-Experts (MoE) layer across GPUs, so every MoE layer requires two all-to-all communications, namely dispatch and combine. Compared with training, inference processes relatively few tokens per GPU at a time, and at small batch sizes fixed costs that do not grow with the number of tokens, rather than data movement, dominate the latency. We present NVLinkOneSided, a one-sided EP all-to-all for GPUs within a single NVLink domain. Each source rank pushes its tokens, without first exchanging metadata, into rank-major buffers on the target GPUs, and the grouped GEMMs read them in place instead of from a per-expert copy. We provide two protocols for the synchronization that the communication requires, namely Fence and Sentinel, for high-throughput and low-latency scenarios respectively. On Blackwell GPUs, NVLinkOneSided completes dispatch and combine 1.2-3.0$\times$ faster than DeepEP, with the largest speedups at small batch sizes. The resulting MoE layer is up to 2.4$\times$ faster than DeepEP with DeepGEMM. It even outperforms the fused DeepGEMM MegaMoE at up to 128 tokens per GPU, making it still competitive for decoding. We expose NVLinkOneSided through a modular interface, so that it is agnostic to the activation quantization format, decoupled from the MoE compute backend and easy to integrate into inference frameworks.
Read the full abstract
Statistical Methodology
Causal Learning with Invalid Instruments for High-dimensional Imaging Responses
Large-scale neuroimaging studies provide high-dimensional imaging phenotypes for understanding brain-related diseases and their causal relations with biological risk factors. While instrumental variable (IV) methods can address unmeasured confounding in such observational data, existing IV methods face significant limitations when applied to high-dimensional imaging data, particularly in handling large-scale data and leveraging the intrinsic spatial structure of images. To address these challenges, we propose High-dimensional Functional-imaging Instrumental Variable Estimation (Hi-FIVE), a novel framework for estimating spatially varying causal effects when a majority of the candidate instruments may be locally invalid. We first establish identifiability conditions under which the spatial continuity of the causal effect surface propagates identification from regions where valid instruments form a plurality into regions where they do not, provided that invalid-IV effects switch on non-smoothly. Motivated by this result, Hi-FIVE combines penalized multivariate splines over triangulations for the causal surface and region-wise penalties for spatially localized invalid-IV effects, under an anchor constraint that keeps at least one relevant instrument valid on every simplex. A distributed algorithm parallelizes the computation over simplices for two- and three-dimensional images. We study the asymptotic properties of Hi-FIVE, establish consistency and pointwise asymptotic normality, and construct bootstrap-calibrated simultaneous confidence regions whose finite-sample coverage is evaluated through simulations. Hi-FIVE is illustrated with an analysis of the effect of phosphorylated tau on FDG-PET imaging in the Alzheimer's Disease Neuroimaging Initiative, interpreted under the stated IV assumptions.
Read the full abstract
Machine Learning
Prediction-Powered LLM Ranking with Adaptively Collected Human Feedback
Ranking large language models (LLMs) through human evaluations is essential for downstream decision-making. Because human feedback is expensive, evaluation platforms often collect it adaptively, prioritizing informative model comparisons using previously observed feedback. To further improve precision with limited human feedback, evaluators can draw on abundant unrated responses and relatively inexpensive LLM judges; however, judge bias and adaptive dependence challenge standard i.i.d. inference frameworks. We propose Prediction-Powered Ranking (PP-Rank), a framework that combines LLM-judge predictions with human feedback to improve precision while providing valid inference under adaptive sampling. Data-driven prediction weights minimize the asymptotic covariance trace within the proposed class. In theory, we establish asymptotic normality for general evaluation targets and simultaneous ranking inference. A joint martingale analysis handles dependence between adaptive human feedback and auxiliary data. We further derive a rule for reducing exploration that prioritizes lower-cost comparisons while matching a conservative variance bound, and characterize allocation limits for regular $\sqrt{T}$-rate estimation of pairwise means. Semi-synthetic experiments on UltraFeedback-Extended show improvements over human-only baselines in score estimation, confidence interval width, and ranking certification. Synthetic experiments demonstrate robustness to judge quality and adaptive allocation, while showing that PP-Rank can substantially reduce evaluation costs without sacrificing precision.
Read the full abstract
Astrophysics of Galaxies
Giant Radio Galaxies from the Third Data Release of the LOFAR Two-metre Sky Survey (LoTSS-DR3): Physical Characteristics and Environments
Giant radio galaxies (GRGs) are radio-loud active galactic nuclei (AGN) with projected linear sizes exceeding 0.7 Mpc, offering insights into the evolution of large-scale radio jets and their environments. We present 2333 GRGs from the LOFAR Two-metre Sky Survey Data Release 3 (LoTSS-DR3), with projected sizes of 0.7-2.9 Mpc, redshifts of $0.126 \leq z \leq 3.517$, and 150-MHz radio luminosities of $1.72 \times 10^{24}$-$3.04 \times 10^{28}$ W Hz$^{-1}$. Of these, 726 ($\sim 31\%$) lie at $z \geq 1$. We investigate their radio properties, lobe energetics, equipartition pressures, black hole masses, mid-infrared colours, and environments. Equipartition pressure decreases with increasing projected size, consistent with declining lobe energy density as radio structures expand, while its positive correlation with redshift persists after controlling for radio power and projected size. WISE colours highlight the diversity of the host-galaxy population, with approximately 75% of sources with available photometry falling within the empirical AGN/HERG region. For BCG-hosted GRGs, physical projected size shows no significant dependence on cluster mass. Although the normalized projected half-size, $D/(2r_{500})$, decreases with increasing cluster mass, this trend is consistent with the expected self-similar scaling of cluster radius rather than an independent suppression of radio-source growth in massive clusters. The substantial high-redshift population provides an opportunity to investigate GRGs at early cosmic epochs, although surface-brightness dimming and inverse-Compton losses against the cosmic microwave background can limit their detectability.
Read the full abstract
High Energy Physics (Phenomenology)
The KLN theorem meets collinear factorisation: Local Unitarity for hadronic cross-sections
We present an initial-state implementation of the Kinoshita-Lee-Nauenberg (KLN) theorem that is compatible with collinear factorisation. In this formulation, hadronic cross-sections are expressed as convolutions of parton distribution functions with partonic cross-sections obtained by summing over the number of incoming partons. By virtue of the KLN theorem, the partonic cross-sections are infrared finite by construction, without relying on dimensional regularisation as an infrared regulator. The Local Unitarity formalism provides a local realisation of this infrared finiteness, allowing partonic cross-sections to be evaluated directly in four dimensions by Monte Carlo integration. We demonstrate the framework at next-to-leading order and leading power. Using Local Unitarity, we compute the Drell-Yan and $t\bar{t}$ production cross-sections numerically; using Reverse Unitarity, we obtain the deep inelastic scattering and Drell-Yan cross-sections analytically. In all cases, a universal scheme conversion relates our results to those obtained in the $\overline{\text{MS}}$ parton model.
Read the full abstract
Software Engineering
AgentScript-Eval: Benchmarking LLMs and Agents for Code Generation in an Enterprise DSL
Domain-specific languages (DSLs) are widely used in enterprise software, giving a compact interface to complex tasks, from data access to infrastructure automation. Platforms increasingly extend them to agents, where a developer declares the topics, permitted actions, and guardrails of an agent rather than writing general-purpose code. Large language models (LLMs) write general-purpose code well, but that ability has been measured almost entirely on languages abundant in public pretraining data, and enterprise DSLs are neither abundant nor public. We introduce AgentScript-Eval, an execution-graded benchmark of 100 real tasks for AgentScript, the declarative DSL that Salesforce uses in Agentforce to build conversational agents. Tasks span authoring an agent from a specification, debugging a broken definition, and optimizing a working one. Evaluating five models zero-shot, we find that the regime is hard: the best model reaches only 25.0% pass@3, and every model scores 0% on authoring from scratch. We further find that model scale and price are weak predictors of success – a 31B open-weight model attains the highest compilation rate of any model tested (48.0%) and matches the strongest frontier model on partial-credit accuracy at the lowest token cost. Adding an agentic compile-and-edit loop is the single largest lever, roughly doubling partial-credit accuracy. Beyond that gain, the wording of domain guidance matters more than its packaging. The same skill pack scores 28% as plain documentation but only 14% as a live plugin. Rewriting one skill inside that plugin then raises it to 37%, our best result, with catalog and tooling untouched. Larger models did not reliably outperform smaller ones; what mattered was the harness surrounding the model and the guidance supplied to it.
Read the full abstract
Machine Learning
Making Strides in Symbolic Imitation Learning: Offline Demonstrations, Online Tuning
Deep reinforcement learning (DRL) methods have been the dominant paradigm in continuous control tasks due to their efficacy in function approximation, while interpretable evolutionary methods have lagged largely behind for the opposite reason. In this work, we introduce Symbolic Imitation Learning, a framework for distilling an oracle policy through an offline dataset into a decision list with linear model motion primitives. We show that this, in conjunction, with online Augmented-Random Search (ARS) tuning of the constants and motion primitives approximates the performance of the oracle across 20 benchmark RL tasks and exceeds the oracle in 11 of those 20 tasks while simultaneously producing parsimonious, interpretable policies that domain experts can better understand.
Read the full abstract
Astrophysics of Galaxies
Dust Reverberation Mapping of Low-Mass Active Galactic Nuclei
We perform a systematic dust reverberation mapping study of low-mass ($M_{\mathrm{BH}}\sim10^6$-$10^7 M_\odot$) active galactic nuclei (AGNs) selected from a literature-compiled catalog, using data from the Zwicky Transient Facility and the Wide-field Infrared Survey Explorer. Through quality assessments, we obtain a core sample of 65 sources with W1- and W2-band lag measurements relative to optical light curves. The W2-band lags are generally longer than the W1-band lags, with a median ratio of $\tau_{\mathrm{W2}}/\tau_{\mathrm{W1}}=1.24^{+0.33}_{-0.21}$. We further investigate the relations between dust radius and AGN luminosity ($R_{\mathrm{dust}}-L_{\mathrm{AGN}}$), using different lag measurements and luminosity proxies. Combining our sample with a higher-luminosity AGN sample, we establish unified $R_{\mathrm{dust}}-L_{\mathrm{AGN}}$ relations spanning nearly seven orders of magnitude in bolometric luminosity. The slopes are $0.47\pm0.01$ and $0.44\pm0.01$ for the W1 and W2 bands, slightly below the expectation of $0.5$ from dust radiative equilibrium. In the combined sample, the lag deviations show a negative correlation with Eddington ratio, potentially related to the self-shadowing effect of slim disks. The luminosity-dependent SED variations revealed by the SPHEREx composite spectra suggest that individual luminosity proxies may not fully trace the intrinsic dust-heating luminosity. Both effects may contribute to the shallow $R_{\mathrm{dust}}-L_{\mathrm{AGN}}$ slopes. Future joint optical and near-infrared monitoring with the Vera C. Rubin Observatory's Legacy Survey of Space and Time and the Nancy Grace Roman Space Telescope will further extend dust RM studies to lower-luminosity AGNs hosting lower-mass black holes.
Read the full abstract
Computer Vision
X Marks the Spot: Diffusion for Image Geolocalization with Location Representations
Image geolocalization predicts the location at which a photo was taken. Accurate geolocalization requires discriminating between both physically and semantically adjacent areas on Earth. However, prior diffusion-based approaches predict locations by generating coordinates, which are semantically unstable over space. We avoid this limitation by learning geolocalization behavior in the pretrained representation spaces of location encoders, which provide both positional and semantic understanding. To make this space usable for coordinate prediction, we introduce a Location Autoencoder (LAE), which pairs a frozen location encoder with a learned decoder mapping embeddings back to coordinates with kilometer-level accuracy worldwide. In the LAE space, we train DIG, a latent diffusion model that transports noise to location embeddings conditioned on image features. DIG consistently outperforms coordinate-based generative models and sets a new state of the art for parametric geolocalization on three benchmarks: OSV-5M, Im2GPS3k, and YFCC4k. Notably, DIG is the first generative model to outperform retrieval-based methods, and further improves retrieve-and-rerank systems when used as a prior. Code and models will be released.
Read the full abstract
Plasma Physics
High-power X-Point Target divertor experiments on the TCV tokamak
Stationary power exhaust in the X-Point Target (XPT) divertor configuration is investigated in high-power ($2.5\,\text{MW}$) TCV L-mode discharges, in which the Lengyel detachment scaling metric $P_{\text{SOL}}B_{0}/(R_{0}n_{e,\text{u}}^{2})$ lies in the range expected for reactors (SPARC, ITER, ARC). Compared with a conventional single-null (SN) divertor, the XPT improves power exhaust performance, reducing the peak outer target heat flux by 40–60% and enabling divertor detachment with nitrogen seeding under conditions in which the SN cannot be detached. Detachment evolves asymmetrically across the multiple XPT outer strike points. The sensitivities of the peak target heat fluxes and impurity emission front positions to the X-point separation ($\text{d}R_{\text{u},X2}$) are characterized and found to decrease at higher nitrogen seeding levels. Field-reversal experiments reveal the critical role of $E\times B$ drifts inherent to the magnetic topology of the secondary X-point, which drive an unconventional redistribution of density and impurity radiation in the XPT divertor and thereby cause the asymmetries between the XPT outer strike points, largely confirming the expectations from recent SOLPS-ITER modelling. A framework for understanding the power exhaust physics of the XPT is presented, detailing the roles of SOL splitting, connection length peaking, $E\times B$ drifts, and vessel geometry. Together, these results establish an experimental physics basis for the XPT divertor as a potential power exhaust solution for tokamak reactors.
Read the full abstract
Computer Vision
Multimodal Judges Flag, Humans Decide: A Review-and-Edit Framework for AI-Generated Product Video Ads
Generative pipelines can now turn every product page of an e-commerce catalog into a short video ad, but deciding whether such an ad is usable is not one good-or-bad call. We describe how we operationalized the review and editing of these ads so that both can keep pace with generation: reviewers go through each ad scene by scene and answer six questions, on the product, scene quality, pronunciation, script, sound and captions; judges built from vision-language, speech and language models file each finding under the matching question, and reviewers accept or reject each finding and add what the judges missed. Each confirmed issue then opens an edit tool matched to it. The questions and the routing from a finding to its fix do not depend on the models that generate the ad. Across 120 ads from our pipeline, reviewed with judge findings on screen, reviewers marked the product wrong on 60% of scenes, and script, caption or sound problems on under 10% each. On 60 of these ads, they marked the product wrong on 58% of scenes, far more than the 26% marked by two reviewers who later saw the ads with the judges hidden; one reviewer who saw the ads both ways marked 57% with judge findings and 19% without. On 120 blind-labeled scenes, a product judge that watches the clip with the catalog photos reached AUC 0.84, while video-quality and photo-similarity scores were close to chance. We tune the judges for recall: the product judge catches nearly all defects but also flags many clean scenes, so reviewers decide; together, the judges widen what reviewers mark and reviewers add what the judges miss. Without judge findings, one reviewer's marking faded over a session. We release a review-portal template and our review data (120 ads and 562 scenes with judge findings and human marks, plus 120 blind-labeled scenes with baseline scores) at https://github.com/anoubhav/ad-review-and-edit.
Read the full abstract
Computer Vision
GroundCascade: Selective Second Opinions for GUI Grounding
Graphical user interface (GUI) grounding maps a screenshot and a natural-language instruction to the screen location where an agent should act. Zoomed views can reveal small targets but may lose the context needed to identify them. We introduce GroundCascade, a training-free controller that uses agreement between a grounder's full-screen and zoomed predictions to decide when to consult a complementary model. It selects between zoomed predictions while preserving corroborated full-screen locations. An optional Finalizer predicts from an unmarked joint crop when disagreement remains (Full); omitting it yields the two-stage Selective variant. GroundCascade reaches 85.33% on ScreenSpot-Pro with Indeed-UI-32B/8B/32B and 88.43% on the visible-target OSWorld-G-Refined subset with Qwen3.6-27B/3.8-27B/3.8-27B, gains of 2.66 and 2.35 percentage points over the stronger standalone two-pass grounder. On UI-Vision, GroundCascade reaches 72.20% category-balanced accuracy, with Selective slightly higher at 72.28%. Across these three evaluations, GroundCascade in Full mode uses 2.42-2.73 calls per request, 31-39% fewer than unconditionally acquiring both models' two-pass predictions. At matched second-model call budgets, selective acquisition gains 1.52-2.56 points over standalone Two-pass, compared with 0.22-0.69 points from stratified random acquisition. GroundCascade leads the official ScreenSpot-Pro leaderboard and exceeds reported GUI-specialist comparisons on OSWorld-G-Refined and UI-Vision, subject to differences in external evaluation protocols.
Read the full abstract
Distributed Computing
SHADOWPREFILL: Keeping Your Next Model Warm for Per-Step Model Routing in Agentic LLM Serving
Per-step model routers send each LLM call of an agent to the cost-optimal model dynamically within each session. However, every model switch discards prefix-cache reuse: KV caches are per model, so the new model must prefill the relevant context before it emits a token. Across representative agentic workloads, including a production coding-agent trace with a mean context of 63K tokens, the prefix hit rate falls from 80-89% on same model calls to 44-65% after a switch, and a switch call takes 1.5-3.2x as long. We present SHADOWPREFILL, which prefills the session's context on the model the router has not chosen, in capacity that instances serving normal traffic leave idle, so the next switch finds its prefix cached. SHADOWPREFILL harvests the per-iteration SLO slack of requests in flight, protects routed calls, spills shadow KV to host memory, and falls back to a cold switch whenever the shadow is incomplete. With LLMRouter routing 24 model pairs over five agentic benchmark suites, SHADOWPREFILL cuts switch-call latency by 8.5-21.4%, LLM-call latency by 3.7-17.7%, and routed token cost by 22.5-46.1%, while normal traffic keeps its SLO and energy per request rises by 0.4%.
Read the full abstract
Quantum Physics
Quantized memory of measurement records near the light cone of quantum circuits
Quantum processors now measure qubits in the middle of a circuit, and the stream of outcomes is the raw output of error correction, measurement-based state preparation and monitored dynamics. For a scrambling circuit, predicting the statistics of this stream is as hard as simulating the processor, so there is no exact reference to check it against once the device outgrows classical simulation. We show that near the light cone the problem becomes exactly solvable, for any two-qubit gate. In a brickwork circuit with one polarized qubit in a maximally mixed background, measurements along the edge of the light cone form a two-state Markov chain, and measurements $d$ sites inside the edge are generated by $\lfloor d/2\rfloor+1$ qubits that move with it, while the rest of the chain, however large, acts on them as a memoryless bath. The memory of the record is therefore quantized, one bit on the edge and at most one more for every site of displacement, a bound that generic gates attain. The same causal structure determines which hardware noise the record can see, removes that sensitivity at no gate cost, and yields a tomography-free witness of dual-unitarity. On a 156-qubit superconducting processor, a demonstration with its analysis specified in advance finds the predicted memory one and two sites inside the light cone, and none where the theory forbids it. This gives dynamic circuits an exact reference at any system size, and shows that the collision models of open-system theory arise exactly inside chaotic dynamics.
Read the full abstract
Quantum Physics
How wide must an eavesdropper be? Leakage of a toric-code logical bit to a strip and the Nishimori point
Fault-tolerant quantum computers hand pieces of each logical qubit to separate decoder modules and shared readout lines. How much does a piece reveal? We prove that for any region and any Pauli noise, including correlated noise, an observer's optimal advantage in learning a logical bit is exactly 1 - 2 P_fail, where P_fail is the failure probability of maximum-likelihood decoding confined to that region. Two lessons follow. Geometry, not noise, provides the protection: a region that misses a single column of the code learns exactly nothing at any noise level, while a region that spans the code is protected by noise only at impractical distances (d > 350 for one row at a 10^-3 flip rate). And width is a poor guide to risk: under i.i.d. noise a two-row strip learns exactly as little as one row, yet below threshold an eavesdropper needs a width growing only as ln d. Through the identity, the strip's operational threshold is the Nishimori point of the random-bond Ising model, and a Walsh-Hadamard transfer matrix up to width 16 locates it to within a few thousandths. On two IBM Heron processors (d = 3, 5), an emulated strip adversary shows the predicted jump at width three, and non-uniform qubit errors let a trained attacker profit from the second row that uniform noise makes useless. We turn these results into design rules for partitioned decoding and multiplexed readout. Code and data are included.
Read the full abstract
Systems and Control
DOE-Constrained Distributed Battery Coordination in Unbalanced Distribution Networks
Dynamic operating envelopes (DOEs) provide injection and withdrawal limits for distributed energy resource scheduling while respecting network operational constraints. This paper develops a risk-aware DOE framework for distributed battery operation in unbalanced three-phase networks using a semidefinite program (SDP). We derive sufficient conditions for exactness of the SDP relaxation, ensuring that the recovered operating points corresponding to the DOE vertices satisfy all voltage and nodal power injection constraints of the original AC network model. The method computes hourly active-power DOEs under uncertain load conditions and incorporates them directly into a 24-hour battery operation problem. On a modified IEEE 13-bus feeder, empirical comparisons with unconstrained arbitrage and local Volt–VAR/Volt–Watt control show that DOE-constrained battery scheduling eliminates the observed out-of-sample voltage violations while retaining 94.6% of the unconstrained arbitrage profit.
Read the full abstract
Computer Vision
Continuous Angular Control in Text-to-Image Generation via Geometric Token Encoding
Text-to-image diffusion models can render a photorealistic analog clock but cannot control the time it displays. Across four pretrained models, fewer than 1% of generated clocks place both hands within 6 degrees of the requested time; the models instead collapse onto visually dominant configurations such as the canonical 10:10 pose. The same failure appears when a specific fraction of a circular object is requested, where models return complete or conventionally sliced objects. We observe that these two failures share a geometric structure: both target quantities can be parameterized by angular variables. We introduce a geometric token encoding framework that conditions generation directly on these angles. Each angular component is encoded as a periodic sine-cosine feature and projected into a role-specific token embedding that is injected into the prompt before a frozen text encoder. To spatially ground these tokens without requiring localization input at inference, we introduce Stochastic Localization Relaxation (SLR), which stochastically relaxes attention supervision from component-level to object-level to unconstrained localization during training. For clocks, we additionally introduce a hand-swap preference objective that discourages assigning the correct angles to the wrong hand roles. On analog clocks our method reaches 63-78% Acc@6 degrees across evaluation methods, compared with below 1% for the baselines. On circular-fraction generation, it achieves 89.5% Acc@0.10, while maintaining 88.8% on unseen object categories and 84.9% on unseen target values, compared with roughly 20% for pretrained baselines.
Read the full abstract
Group Theory
Smooth actions of Baumslag-Solitar groups on the interval
We study representations of the Baumslag-Solitar group $BS(m,n)=\langle a,b\mid ab^ma^{-1}=b^n\rangle$ inside the group of orientation-preserving diffeomorphisms of the interval $[0,1]$. We prove that for integers $n,m\in \mathbb Z\setminus\{-1,0,1\}$, the group $BS(m,n)$ embeds into $\mathrm{Diff}_+^1([0,1])$ if and only if $|n|=|m|$. Furthermore, we show that for any $\alpha\in (0,1)$, the group $BS(m,n)$ embeds inside $\mathrm{Diff}_+^{1+\alpha}([0,1])$ if and only if $1=|m|=|n|$ or when $1=m\leq n$.
Read the full abstract
General Relativity and Quantum Cosmology
Depletion of Dark Matter Spikes by Massive Black Hole Mergers
Dark matter over-densities around massive black holes, known as dark matter spikes, can dephase gravitational waves emitted by intermediate- and extreme-mass-ratio inspirals (IMRIs/EMRIs), offering a promising way to indirectly detect dark matter. This detection channel, however, requires the spike to survive over cosmological time-scales. We investigate the impact that a host black hole's merger history has on its dark-matter spike. We focus on how the inspiral of a massive-black-hole (MBH) binary modifies and disrupts these spikes within the central $\sim 100$ Schwarzschild radii of the black hole. We find that a single merger with mass ratio $q = 0.03$ is capable of reducing a spike's density by around 3 (5) orders of magnitude at 100 (5) Schwarzschild radii across all polar angles relative to the binary's plane, independent of the primary MBH's mass, with larger mass ratios inducing even stronger depletion. In fact, a single past merger with mass ratio as small as $q = 0.005$ reduces central densities by 2–3 orders of magnitude, making gravitational-wave dephasing difficult to detect. We relate these results to the MBH merger history by analysing cosmological simulation data from Horizon-AGN. We find that mergers become increasingly common as the MBH mass increases, with mergers of mass ratio q > 0.03 having occurred in 50% of MBHs of mass $3\times10^7M_\odot$ and 95% of mass $10^8M_\odot$ at z=0. These fractions are reached at even lighter masses ($2\times10^7M_\odot$ and $5\times10^7M_\odot$, respectively) at z=3. We find that accretion-induced adiabatic MBH growth is unable to cause appreciable spike re-growth between such mergers in Horizon-AGN. Thus, in combination with previous studies of EMRI-induced depletion, this work constrains the mass range over which massive black holes can host detectable DM spikes to $10^6M_\odot$–$5 \times 10^7M_\odot$ for $z \leq 3$.
Read the full abstract
Econometrics
Identification, estimation, and robust inference of structural autoregressions for multidimensional data
Many modern datasets such as cross-country panels have a multidimensional structure. Conventional approaches on structural vector autoregressions for identifying structural shocks are not directly suited to these settings where data are naturally organized as matrices or tensors, since the number of parameters grows rapidly and identifying restrictions become hard to impose. We introduce a class of structural autoregressions for multidimensional data that achieves tractability and global identification by assuming a Kronecker structure on the model coefficients. We develop a minimum-distance estimator and establish its asymptotic distribution. To accommodate approximate Kronecker separability, we develop confidence intervals for the structural impulse response functions that are robust to local misspecification, and we show their asymptotic validity. Monte Carlo simulations demonstrate the validity of the asymptotic approximations and good finite-sample performance of the estimator and their (robust) confidence intervals. We illustrate the methods on a country panel dataset to analyze the dynamic and spillover effects of US monetary policy shocks.
Read the full abstract
Machine Learning
RadRef: A Reference-Free Metric for Radiology Report Generation
AI-generated radiology reports are commonly evaluated against radiologist-authored references, limiting the applicability of standard reference-based metrics when no reference is available. We introduce RadRef, a reference-free metric that scores report claims against image-derived beliefs over a clinical finding vocabulary, weighted by relevance to the clinical indication. By evaluating findings beyond the report's stated claims, RadRef assesses omissions as well as unsupported assertions. On ReXVal, RadRef achieves $73.1\%$ agreement with radiologist rankings among non-tied same-radiograph report pairs, with $61.8\%$ pair coverage and $64.3\%$ tie-inclusive agreement. Its non-tied agreement is close to GREEN ($73.9\%$) and RadFact ($72.6\%$), with its strongest relative results on omissions. We further evaluate $100$ model-generated and $100$ radiologist-authored reports from $100$ MIMIC-CXR studies, independently annotated by two board-certified radiologists. On same-image omission rankings, RadRef attains Kendall's $\tau_b = 0.295$, compared with $0.386$ between the readers. In separate paired comparisons across generated reports, no evaluated comparator significantly outperforms RadRef on omissions. The same image-grounded representation provides a false-finding readout, RadRef-FF, with $\tau_b = 0.32$ on ReXVal and $0.29$ in the reader study. Controlled perturbations confirm that RadRef depends on the radiograph, report, and clinical indication. RadRef enables image-grounded evaluation before a radiologist-authored reference exists, supporting quality monitoring of report-generation systems in clinical settings, and complements reference-based metrics when one is available.
Read the full abstract
Artificial Intelligence
AI Alignment Is Neither Solvable nor Sufficient
Humans and AI systems increasingly act within the same social and economic environments. Alignment, the goal of ensuring that AI systems act in accordance with human values, is widely regarded as a central challenge for ensuring that AI benefits society while avoiding harm. We argue that this endeavor is misguided. First, alignment is not well defined, as it presumes a sufficiently stable and identifiable set of human intentions or values. Second, where a concrete notion of alignment can be specified, it is not solvable. Third, where alignment appears solvable, solving it would not necessarily mitigate the negative consequences of AI. A key reason is that alignment is typically framed as a property of individual AI systems, while the resulting hybrid societies of humans and machines are complex systems shaped by interactions and feedback loops. Emergent risks, and not only those posed by individual misaligned systems, are therefore likely to become a major source of harm. We conclude that AI safety is an ongoing challenge and not a problem that can be solved once and for all by verifying a particular model property.
Read the full abstract
Strongly Correlated Electrons
Study of the frustrated SU(3) Heisenberg chain with SU(3)-symmetric matrix product states
We study the SU(3) Heisenberg chain with nearest- and next-nearest-neighbor antiferromagnetic exchange $J_1$ and $J_2$. Field theory predicts that the trimerized valence-bond solid of this chain has a finite extent in $\alpha=J_2/J_1$, followed by a gapless SU(3)$_1$ phase, but earlier numerical estimates of the upper boundary range from $\alpha\approx2.1$ (exact diagonalization) to $\alpha\approx3.8$ (iDMRG). We use the variational uniform matrix product state (VUMPS) algorithm with the full non-Abelian SU(3) symmetry imposed on the virtual bonds, reaching bond dimensions up to $D\approx8000$, and locate the phase boundaries by finite-entanglement scaling. Resolving the transfer-matrix spectrum and the entanglement spectrum by SU(3) irreducible representation, we find that the trimerized phase extends to $\alpha_{c2}\gtrsim3.5$ and end below $\alpha\approx5$, excluding a boundary near $\alpha\approx2$. At larger $\alpha$ our numerical analysis of the effective central charge is consistent with the predicted massless flow from SU(3)$_2$ to SU(3)$_1$.
Read the full abstract
Biological Physics
Chemotactic Horizons Regulate Threshold Dynamics in the Biological Carbon Pump
Microbial encounters with sinking marine snow regulate organic carbon remineralization, acting as a critical knob in the ocean's biological carbon pump. Yet how the vast diversity of microbes in the open ocean respond to natural chemotactic cues coupling sedimentation-induced flow to microbial encounters in the ocean remains unresolved. Here we report shipboard measurements of microbial motility around freshly collected marine snow under imposed flows spanning ecologically relevant sinking speeds. Via high resolution single cell tracking at sea, we map a highly diverse repertoire of microbial motility, including rapid switching between motile and non-motile states. We identify a chemotactic horizon from measured reorientation towards chemical gradients, defining a spatial region around an aggregate within which chemotactic alignment overcomes flow-induced rotation due to sedimentation. This interplay produces a non-monotonic dependence of encounter rates on aggregate sinking speed, with a peak at 14.4 m/day followed by a sharp decline beyond 78.1 m/day, with a tracer-like regime at high effective sinking speeds. We incorporate the measured encounter rates into a kinetic transport equation in position-velocity phase space, with an effective nonlinear damping arising from velocity-dependent remineralization. The resulting particle dynamics reveal a threshold-induced separatrix dividing overdamped and ballistic microbial populations, leading to velocity segregation. Our study connects direct measurements of microbial motility and sedimentation-induced flow to the emergent phase-space for fate of marine snow in the biological pump, establishing a quantitative multiscale framework for investigating ocean carbon export.
Read the full abstract
Numerical Analysis
Sobolev inequalities for the Discrete de Rham complex and an application to the Navier–Stokes equations
We prove Sobolev inequalities for the discrete de Rham (DDR) method, an arbitrary-order discretisation of the de Rham complex on polyhedral meshes. We show that, for any polyhedral domain $\Omega$, the DDR counterparts of the spaces $H({\bf grad},\Omega)$, $\boldsymbol{H}({\bf curl},\Omega)$ and $\boldsymbol{H}(\mathrm{div},\Omega)$ satisfy the same Sobolev embeddings as their continuous counterparts. For the curl and divergence spaces, this establishes on quasi-uniform meshes an $L^r$-bound (with $r>2$ in general, and $r=6$ if $\Omega$ is convex) on discrete vector fields that are orthogonal to discrete gradients (resp.~curls), in terms of the $L^2$-norms of the vector fields themselves and their discrete curl (resp.~divergence). These Sobolev embeddings are the first such results for a polytopal complex. Their proofs hinge on recently introduced quasi-interpolators and conforming liftings. As an application, we use these estimates to prove the convergence of a DDR scheme for the Navier–Stokes equations in the curl–curl formulation. The convergence proof uses a Maxwell compactness theorem for DDR and requires no smallness assumptions on the data.
Read the full abstract
Machine Learning
High Probability Risk Control for Online Policy Learning
We study risk control for online policy learning, where an agent adaptively collects data and aims to learn a new policy whose risk stays below a tolerance threshold at every round with high probability. Such policy updates and data collection induce feedback covariate shift (FCS): observed data depend on history through evolving policies and the data-collection process. FCS brings two substantive problems for risk control: 1) efficient risk evaluation for each candidate policy and 2) valid policy selection with high-probability risk control. To address these, we propose a permutation-based weighted risk evaluator with weights efficiently approximated by the Sinkhorn algorithm and select policies whose estimated risk upper confidence bounds (UCB) fall below the target threshold, with the UCB obtained through a proposed Gaussian bootstrap method. Theoretically, we prove that our risk evaluator has a no-larger variance upper bound than that of the standard importance-weighted evaluator. Our policy control method also yields an asymptotically valid simultaneous risk bound and provides asymptotic high-probability risk control for the selected policy. Empirically, we evaluate our method on the online contextual bandit problem, safe question answering benchmark, and Medical Claim Filtering tasks. Our method maintains the target risk level, achieves lower risk than baseline methods, has lower variance in risk estimation, and achieves reliable policy selection with high-probability risk control. Further ablation studies support our theoretical claims.