Papers

This work concerns our submission for the Multilingual track of the 2026 BabyLM Challenge. We improve on our baseline using data that has been summarized and translated by an LLM, showing that higher-quality data can come from synthetic sources. We further improve with 2 simple changes: using a larger model, and using a character-level vocabulary. Both of these changes require LoRA-style fine-tuning for realizing their full potential. We dub our final Qwen3-based model Qwenfant, which, at the time of submission, ranks 1st in the Multilingual track.

Let $p$ be a prime $\ge 5$. We determine the semisimplifications of the mod $p$ reductions of irreducible two-dimensional crystalline representations of the Galois group of ${\mathbb Q}_p$ of slope $2$ and weights $k\ge3p+4$ with $k \not \equiv 6 \bmod (p-1)$. The proof relies on the compatibility between the $p$-adic and the mod $p$ Local Langlands Correspondences for $\mathrm{GL}_2(\mathbb{Q}_p)$.

Model routing can reduce the cost of LLM applications by assigning simpler requests to cheaper models without sacrificing the quality of the response. However, most routers treat requests as independent. Stateful agents do not operate under this assumption: a user turn may require multiple model and tool calls, its difficulty depends on the state produced by earlier turns, and changing models can sacrifice reuse of the conversational prompt cache. We study zero-shot, cost-aware routing at the user-turn boundary, systematically varying the routing rule, access to session history, candidate models that differ in size, reasoning effort, or model provider, and prompt-cache behavior. We evaluate on multi-turn versions of SheetCopilot and SpreadsheetBench 2, where persistent workbooks and cross-turn dependencies make isolated routing decisions especially brittle. At a matched offloading rate, giving the router session history improves task pass rate over the same router without history, with gains for every model pair; history-aware routing remains within one point of always using the expensive model. We also found that cost depends on execution dynamics rather than token prices alone: the cheaper model's per-task cost advantage narrows from 6.8x to 3.2x on longer sessions. A closed form cache analysis further shows that downward switches (expensive to cheap) recover their cache cost within two LLM calls, whereas upward switches do not break even under the evaluated conditions. Together, these results show that agent routing is a sequential decision problem: effective policies must account for what happened earlier in the session and for the asymmetric cost of switching models.

In so-called "Industry 4.0," integrated sensors, analytics tools, and other digital technologies supplement industrial workflows, potentiating enhanced visibility and automation. Prognostics and health management (PHM) and natural language processing (NLP) practitioners alike have explored the application of large language models (LLMs) to industrial asset management, but many prior works' focus on exam-like benchmarks has limited their assessment of LLMs' ability to automate practical operations. Herein, we introduce CoffeeBench, a fault diagnostics task for LLM agents, using a simulated appliance. The LLMs we evaluate struggle to consistently identify faults from the sensor information alone, but the best-performing ones can iteratively refine the fault-option space to eventually obtain a successful diagnosis. Our subsequent analyses, therefore, focus on: differentiating models by diagnostic efficiency (RQ1); examining the lagging performance of the three worst models (RQ2); and identifying viable heuristics for selective prediction (RQ3), including that initial diagnostic uncertainty and LLM-judged soundness of reasoning are high-signal predictors of diagnostic success. In addition to our benchmark results (Section 3.1), we provide practical insights for improving the performance of the open-weight models on CoffeeBench, such as by explicitly tracking episode state on their behalf to reduce timeouts (Section 3.2) or applying a deliberation loop to refine diagnostic hypotheses (Section 3.3).

We present Kandinsky S2V 1.0, an open family of speech-to-video models for controllable audio-driven human animation, comprising Kandinsky S2V 1.0 Lite (2B parameters) and Kandinsky S2V 1.0 Pro (21B parameters). From a single reference image, a speech waveform, and a text prompt, the models generate long-form 25 FPS video in which lip motion, expression, pose, and gesture follow the speech. The architecture draws text and audio conditioning from one Qwen2.5-Omni model and adds a reference-aware three-scale visual memory. Training starts from the Kandinsky 5.0 Video foundation models [1] and combines mixed reference-memory layouts with an area-adaptive face loss. The training corpus contains 1.17M single-speaker clips selected by active-speaker detection, with captions generated directly in English and Russian. A side-by-side human evaluation against the strongest open-source baseline and two proprietary commercial models shows that Kandinsky S2V 1.0 Pro performs comparably to or better than these models on average and is also preferred for lip synchronization. On common speech to video benchmarks, Kandinsky S2V 1.0 Lite and Kandinsky S2V 1.0 Pro rank among the top performing models according to mean rank over five reported metrics. Beyond the base models, we provide {\pi}-Flow distilled variants of both Kandinsky S2V 1.0 Lite and Kandinsky S2V 1.0 Pro. The distilled Kandinsky S2V 1.0 Pro model generates video in 16 sampling steps without classifier-free guidance, achieving a 5.4 times end-to-end speedup over the base model, while matching or exceeding it on nearly all reported benchmark metrics. To accelerate open research and deployment in speech-to-video generation, we release the code, model checkpoints, and diffusers integration under the MIT license.

The electric power grid is a critical infrastructure system subject to different types of failures which can have devastating impacts on society. One framework to study the resilience of the grid is the N-k interdiction model, which introduces an attacker and defender model to identify grid vulnerabilities. However, traditional N-k interdiction models characterize the power grid with steady-state assumptions. One key aspect missing from this analysis is the behavior of the grid as it evolves from the pre-failure grid state and post-failure grid state. During such a transition period, automated protection systems, transient instabilities, and other phenomena can trigger cascading failures that can occur before grid operators can respond. To bridge this gap, this paper first introduces the generalized induced failure N-k (N-k.k') problem, where an initial set of k failures induces an additional set of k' failures. In this context, we propose a disjoint optimization approach that decouples the computation of cascading failures from the interdiction model. This system response model, furthermore, is treated as a black box module without compromising the underlying performance guarantees of the interdiction model. Extensive experiments on several benchmark power grid case studies showcase the efficacy and scalability of the approach.

We study the linearization, with respect to zeroth-order perturbations, of the local Dirichlet-to-Neumann map of a formally self-adjoint elliptic operator $P=-\Delta_g+Q$ on a compact Riemannian manifold $(M,g)$ of dimension $n\geq 2$ with boundary, where $Q$ is of order at most one. Near a boundary point where the boundary, metric, and $P$ are analytic, we show that the linearized map is a classical analytic pseudodifferential operator of order $-1$ if and only if the zeroth-order perturbation $q$ extends analytically across the boundary. We also prove that local boundary data in an arbitrary open analytic boundary patch determines $q$ locally. In particular, if the measurement set meets every connected component of $M$, the local linearized DN map is injective on $L^\infty$ perturbations. Finally, we obtain a global completeness result for products of pairs of solutions whose boundary traces are supported in the measurement set. The global recovery proceeds by local steps using Runge approximation.

Compute cost in automated essay scoring (AES) is usually treated as a technical efficiency issue rather than a measurement design concern. Yet deployment decisions require more than reporting latency, model size, or inference expense. In high-stakes settings, total burden includes validation, monitoring, adjudication, compliance, and governance; in instructional settings, pricing, licensing, and integration may dominate cost. This paper reframes AES as a cost-constrained measurement system and introduces the Cost-Aware Measurement Design (CAMD) Framework, a four-layer structure consisting of metering, allocation, unit economics, and decision, for translating computational and workflow resource records into decision-relevant cost constructs. Direct cost is specified as monetary outlay plus attributable labor and time, and two criteria, design attribution and auditability, determine what the cost boundary excludes. Burdens that satisfy neither criterion, such as institutional capability and governance posture, are represented as constraints on which designs are admissible rather than priced. The paper defines cost per defensible score (CPDS) and argues that AES systems should be compared by quality-cost sufficiency: minimizing cost among defensible scoring designs that satisfy measurement requirements. Scoring volume conditions all four layers, so cost per defensible score is treated as a function of scale rather than as a single transferable number. CAMD provides a portable basis for representing, comparing, and evaluating cost across AES studies and deployment contexts.

The mutual information of measurement of a local operator between two spatial regions characterizes the reduction in the uncertainty of measurement outcomes in one region given the outcome in the other. We study the Rényi-Shannon mutual information of measurement in conformal field theories (CFTs). Explicit calculations for scalar primaries and descendants in large $N$ CFTs are performed in the two extreme limits of the intervals being adjacent or largely separated. The results depend on only the cross ratio and the scaling dimension $\Delta$ of the operator being measured, with striking similarities to the mutual information of entanglement. We find a surprising quadratic dependence as $(\Delta-1/2)^2$ for adjacent intervals. This dependence superficially seems to violate the information processing inequality, which we discuss and resolve. We also offer a holographic interpretation of the mutual information of measurement.

We give matching two-sided estimates for heat kernels obtained from Schrödinger operators $\Delta+W$ on weighted graphs. The graphs we consider are transient and admit two-sided sub-Gaussian heat kernel estimates for the Laplacian $\Delta$, and the potentials $W$ decay to zero at infinity. We offer two proofs for one of our main results: one using the Feynman-Kac formula, and one without using it. Our results apply to fractal-type graphs such as graphical Sierpinski carpets, among other examples.

Chern insulators have optical properties that are strikingly different than ordinary chiral media. Unlike the natural optical activity of the latter, which is a finite-wavelength effect arising from spatial dispersion, the optical activity of a Chern insulator is already present in the long-wavelength limit, and exhibits a striking directional asymmetry in the propagation of circularly polarized light. We study this unnatural optical activity using a recently developed field theory for Chern insulators. For bulk Chern insulators, we derive the normal-incidence transmission and reflection coefficients of multilayer structures, identify the associated Faraday and Kerr angles, and show that these optical rotations provide a direct optical signature of topology at low frequencies. We also show that an interface between bulk Chern insulators can act as a kind of mirror, completely reflecting light at normal incidence when the Chern invariants have opposite signs. Moving to a more realistic model of a two-dimensional current sheet, we study the Faraday and Kerr effects in a thin film of $\mathrm{Mn}\mathrm{Bi}_2\mathrm{Te}_4$ in its Chern insulator phase. The predicted rotations are many orders of magnitude larger than those of optically active media of comparable thickness that are topologically trivial.

We examine the metamathematical significance of the Levy scheme $\Lambda$, which asserts, for each natural number $n$, the existence of an $n$-Mahlo cardinal $\kappa$ such that $V_\kappa\prec_{\Sigma_n}V$. Several initially unrelated constructions single out $\mathsf{ZFC}+\Lambda$: comparisons with the model theory of arithmetic and finite set theory; class theories in which the class of ordinals has the tree property; set theories with a proper class of indiscernibles; and interpretations of ordinary set theory in Cantorian parts of models of the extension $\mathsf{NFUA}$ of Quine–Jensen set theory $\mathsf{NFU}$. Taken together, these results provide a precise sense in which $\mathsf{ZFC}+\Lambda$ is an inevitable extension of $\mathsf{ZFC}$ if one attempts to generalize certain metamathematical results about the universe of all finite sets to the entire universe of sets.

Stress reshapes cortico-limbic dynamics across multiple timescales, with important consequences for healthy and pathological arousal responses. We develop and analyze a Wilson-Cowan model of prefrontal-amygdala circuitry, incorporating excitatory and inhibitory populations involved in emotional regulation. The effects of stress cortisol are represented as history-dependent modulations of two stress-sensitive interactions: recurrent prefrontal excitation and inhibition of basolateral amygdala principal cells. Using numerical simulations and bifurcation analysis, we examine how this self-interacting system organizes transitions between stable activity and infra-slow oscillations, and how this structure changes with cortisol-related temporal memory. To distinguish the effects of memory temporal distribution from those of its timescale, we compare weak and strong Gamma kernels with a discrete-delay limit, representing progressively different ways in which past activity can influence current coupling. Introducing temporal memory reshapes the parameter region supporting oscillations and alters their persistence. Weak and strong Gamma kernels produce remarkably similar global dynamics, whereas the discrete-delay limit generates a substantially richer bifurcation structure, including quasiperiodic activity. Together, these results suggest that the dynamical effects of cortisol depend not only on the characteristic timescale of its action, but also on the shape of the temporal memory through which past neural activity influences the present. Different forms of temporal integration may provide a mechanism by which neural systems implement distinct functional dynamics, while changes in this integration could promote atypical or pathological rhythmic regimes.

Understanding neural scaling laws is increasingly important for theory and practice, yet first-principles explanations remain limited even in simple settings. In this work, we study capacity-limited scaling in a bottlenecked bigram model with power-law (Zipf) token statistics, viewed as an analytically tractable toy model of a large language model. Fixed random embeddings serve as a proxy for representational capacity and induce an explicit bottleneck of width, allowing us to derive scaling laws for prediction regret as a function of model width, vocabulary size, and the Zipf exponent. Our results complement recent theoretical accounts based on optimization dynamics by isolating scaling due purely to representational constraints. In the appropriate regime, the theory recovers the functional form of classical neural scaling laws and suggests that heavy-tailed language statistics may underlie both optimization- and capacity-driven scaling. Experiments natural language datasets (WikiText2, OpenWebText) closely match the predicted curves.

With a 35-day orbital period and $\sim$0.5 eccentricity, TOI-2257 b occupies a sparsely populated corner of transiting mini-Neptune parameter space and may lie near the orbital distance at which tidal and atmospheric escape process become negligible. To test this hypothesis, we performed simulations of the combined stellar, tidal, and atmospheric evolution and confirmed that TOI-2257 b does lie near this boundary, regardless of whether it formed with or without a primordial hydrogen envelope. In the mini-Neptune scenario, we find that all cases retain an envelope of at least 0.02 M$_\oplus$ for 8 Gyr, but that the envelope mass can be halved if it was initially larger than 0.2 M$_\oplus$. We find that tidal evolution only occurs for cases with initial envelope masses larger than 0.02 M$_\oplus$ because their radii are large enough to generate significant torques. In the super-Earth scenario, the planet could lose $\sim$10 Earth ocean masses of surface water during the $\sim$75 Myr pre-main sequence and therefore could be habitable if it formed with $\gtrsim$2x Earth's surface water mass. The predicted tidal heating surface flux tends to be similar to Io, with no cases approaching the flux required to drive a runaway greenhouse. We find that if the planet is rocky, then eccentricities larger than 0.62 today are strongly disfavored, implying that larger eccentricities could confirm the gaseous nature of the planet. These results suggest that TOI-2257 b could provide valuable constraints on mini-Neptune properties and new insight into the radius valley shape as a function of eccentricity.

Modern machine learning models increasingly operate on variable-size structured data, such as sets and graphs. A central requirement is size generalization: the ability of a model trained on inputs of one size to generalize to inputs of a different, potentially much larger, size. Classical statistical learning theory, however, primarily considers fixed input domains, and its guarantees often deteriorate with input dimension. Although extensive work has established when models on sets and graphs can be transferred across input sizes, such results do not characterize the statistical cost of doing so: the sample complexity of size generalization, and its relationship to classical in-domain learning, remain largely unknown. We develop a statistical learning theory of size generalization and establish a sharp exponential separation in sample complexity. Letting $n$ denote the number of training samples, we show that polynomial source decay is insufficient for efficient size generalization: while it yields error $n^{-\Theta(1)}$ and hence polynomial sample complexity for in-domain learning, size-generalization error decays only as $(\log n)^{-\Theta(1)}$, yielding an exponential separation. Thus, polynomial in-domain sample complexity can become exponential across sizes. We then identify an exponential source condition under which size generalization achieves error $n^{-\Theta(1)}$, recovering polynomial sample complexity. These guarantees are independent of input dimension and depend only on the number of training samples, yielding dimension-free scaling laws. We also provide matching minimax lower bounds, establishing optimality. To our knowledge, this work provides the first dimension-free minimax theory of size generalization and a rigorous characterization of when, and at what statistical cost, learning can transfer across input sizes.

High-performance camera-based perception pipelines must sustain per-frame real-time budgets on resource-constrained embedded GPU platforms. Availability attacks such as SlowTrack inflate the persistent state of multi-object trackers: data-association workload accumulates across frames, so per-frame latency increases and drives the pipeline past its deadline. Detector- and NMS-layer defenses do not directly control this persistent tracker-side resource. We present TrackShield, a lightweight runtime resource-management layer that protects the tracker at the point where attack-induced workload accumulates. TrackShield combines low-overhead anomaly monitoring with staged birth capping, pruning of young or weakly matched tracks, and per-stage association-input admission control. Its hard admission guard provides a deterministic bound on admitted association workload, independently of the heuristic monitor. On a Jetson AGX Orin, TrackShield reduces attacked tracking-stage latency from $139.90$ to $6.27$\,ms and whole-pipeline P99 latency from $190.9$ to $48.8$\,ms, at a benign MOTA cost of $1.06$ points. As an availability defense, TrackShield focuses on keeping the tracker responsive under attack instead of recovering the full tracking accuracy degraded by adversarial detections.

Lane-based RISC-V vector processors exploit data parallelism well but pay a recurring cost for two common patterns: reductions, which require a separate tree pass, and deep dependent expression chains, which cannot be fused and are issued as serial vector instructions. We present TALU, a 5-stage RISC-V core whose execute stage is an NxN grid of registered ALU tiles instead of a single ALU. The grid is a reconfigurable, feed-forward spatial fabric: tiles route operands only from non-backward neighbours, and added diagonal forwarding lets a reduction collapse spatially. A minimal custom extension configures the per-tile operation, routing, and per-tile constants, and streams operands through the array. On a single fabric, TALU runs ordinary scalar code, single deep-path expressions, and multi-lane vector kernels. On the benchmark suite of the RISC-V2 vector processor, TALU is 1.8-5.5x faster than RISC-V2 and 4.5-15.5x faster than Vicuna at equal lane count, with a reduction cost (dot{\div}add) of 1.05x versus 1.96x and 1.26x for the vector cores. A retimed implementation places and routes at 1.03 GHz in 45 nm in 0.30 mm2, half the area of RISC-V2.

Symmetry-protected electronic states in correlated oxides offer a route to combining topology, lattice tunability, and many-body interactions, but creating such topological electronic states in complex oxides remains challenging. Here, we show that oxygen octahedral rotations provide a symmetry-control parameter for stabilizing Dirac-related electronic states in epitaxial CaNbO3 films. By comparing compressively strained tetragonal films and relaxed orthorhombic films, we demonstrate that distinct octahedral rotation patterns preserve nonsymmorphic symmetry and lead to distinct transport signatures consistent with Dirac-like electronic states. The films exhibit metallic transport with high mobility of 138,000 cm2 V-1 s-1, multichannel conduction, and Shubnikov-de Haas oscillations from small high-mobility Fermi pockets. Landau-fan diagram analysis indicates finite Berry-phase contributions, while first-principles calculations reveal a strain-driven evolution from nodal-line band crossings in the orthorhombic phase to Dirac-electron band features in the tetragonal phase. These results establish oxygen octahedral symmetry as a design knob for engineering topological electronic states in correlated perovskite oxides.

Atomically well-defined interfaces are essential for realizing functional interlayer coupling in stacked and twisted oxide heterostructures, yet transferred freestanding oxides can develop structural inhomogeneities, including local curvatures, interfacial voids, and incomplete contacts with substrates. Here, we demonstrate that these transfer-induced structural disorders and imperfections evolve substantially and can be suppressed by post-transfer thermal treatments, yielding bulk-like crystalline freestanding oxide nanomembranes. Using freestanding SrTiO3 membranes as a model system, we report a systematic understanding of the role of "atomic ironing" and the associated evolution of lattice mosaicity, interfacial roughness, and optical bandgap. We further demonstrate atomically bonded conformal interfaces in twisted SrTiO3 heterostructures, highlighting the importance of post-transfer thermal processes. These results establish a synthetic strategy to restore the structural integrity and intrinsic bulk-like functionality of freestanding oxide nanomembranes, enabling stacked, twisted, and heterogeneously integrated oxide architectures with atomically well-defined interfaces for emerging quantum functionalities.

Atmospheric turbulence creates complex spatial and temporal distortions that significantly reduce image quality. This work compares two deep learning methods for turbulence reduction: Generative Adversarial Networks (GANs) and Transformer-based models. Both methods are trained and tested on the SOTIS dataset, which includes multi-frame sequences under different turbulence conditions with ground truth references. Our results show an important finding: GAN-based models perform better in severely degraded, low-quality situations, while Transformer-based models achieve better results in moderate to high-quality settings, especially in correcting geometric distortions. We also enhance the Turbulence Mitigation Transformer (TMT) by introducing a refined training strategy, an augmented input design that includes a temporally averaged frame as prior structural information, and a modified loss function that incorporates the Structural Similarity Index (SSIM). These improvements lead to consistent performance gains across various turbulence conditions. Overall, this study provides an in-depth evaluation of deep learning methods for turbulence reduction and offers practical guidance for selecting models in different imaging environments.

We study the $\alpha$-Ricci flow $(g(t),\alpha(t))$, which is a Ricci flow coupled with a harmonic 1-form heat flow, on the 2-torus $M$. We prove global existence for this flow as well as a Type III curvature estimate. The long-time behavior of the $\alpha$-Ricci flow depends on the cohomology class $[\alpha] \in H^1(M,\mathbb{R})$, which is invariant under the flow. If $[\alpha]$ is rational, then as $t\to\infty$, $(M,t^{-1}g(t))$ Gromov–Hausdorff converges to $S^1$ with length $\ell([\alpha])/\sqrt 2$ where $\ell([\alpha])$ is the primitive length of $[\alpha]$. On the universal cover, the rescaled flow $(s^{-1}\tilde g(st),\tilde\alpha(st))$ smoothly converges to an expander $(g^*(t),\alpha^*(t))$ on $\mathbb{R}^2$ as $s\to \infty$ where $\alpha^*(t)$ is nonzero and constant in spacetime and $g^*(t)$ is flat and expanding in the $\alpha^*$ direction. If $[\alpha]$ is irrational, then there is a sequence $s_i\to\infty$ such that $(M,s_i^{-1}g(s_it))$ Gromov–Hausdorff converges to a point. Moreover, on the universal cover, $(s_i^{-1}\tilde g(s_it),\tilde\alpha(s_it))$ smoothly converges to $(g^*(t),\alpha^*(t))$.

Choosing which information to use as the context changes is a basic cognitive operation. We introduce a new task of this kind for quantum reservoir computing (QRC): a binary cue indicates which of two simultaneous input streams the reservoir must reproduce. In simulated spin reservoirs with a trained linear readout, infinitely many measurements (exact expectation values) let the reservoir reproduce the cued stream accurately. At a finite number of measurements \(N_M\), the score exceeds that of a linear readout of the raw inputs only at large \(N_M\) because of sampling noise. Performance increases with both entanglement and coherence in the computational basis, and then saturates. These results show that QRC can select and reproduce a cued stream, and they are a starting point for harder cognitive tasks.

On-policy distillation provides dense supervision on the student's own samples, but its targets still come from the teacher. Verifier reinforcement learning (RL) instead uses an outcome signal, but it provides no direct constraint on capabilities outside the verifier. Combining the two is difficult for diffusion language models (dLLMs), because bidirectional denoising makes the policy likelihoods required by ratio-based RL unreliable. We develop a likelihood-free pipeline for a mixture-of-experts dLLM. Each rollout group is routed by verifier outcome. Mixed groups use ratio-free, group-centred REINFORCE. If every student rollout fails but a teacher succeeds, the model receives dense KL supervision on the student's denoising canvases. All other groups are skipped. A KL term anchored to the distilled checkpoint limits changes outside the verifier reward. The tool-calling teacher succeeds on only 18.6% of the training prompts. At 24 denoising steps, distillation exceeds the teacher by 6.77 percentage points, and routing improves the distilled checkpoint by a further 18.31 points. Routing lowers reasoning-trace accuracy by 0.71 points but raises direct-answer accuracy on the same items by 13.21 points. The anchored arm restores trace accuracy. Its tool-calling score differs by -0.16 points, with a confidence interval that includes zero, although this comparison also removes a small fallback distillation route. A domain-conditional recipe retains a 16.37-point tool-calling gain; a second seed gives 18.23 points and reproduces both preservation results, but the math result is seed-dependent. We pre-registered every arm and evaluated each comparison on paired items in a single sweep on one machine.

The ecological insurance theory posits that diverse forests are more resilient and better able to adapt to changing environmental conditions. Increasing drought and heat stress in Central Europe promotes mixed forests for maintaining ecosystem services. However, a diverse species distribution complicates remote sensing-based species mapping, as distinguishing pure from mixed species compositions at the pixel level remains challenging. We evaluated the potential of recurrent neural networks (RNNs) for tree-species classification from Sentinel-2 time series and investigated the spectral-temporal features underlying class discrimination, with particular emphasis on separating pure from mixed Norway spruce pixels. We compared RNN classification models with Random Forest (RF) and Extreme Gradient Boosting (XGBoost) and used Integrated Gradients (IG) to quantify feature attributions across spectral bands and time. The best-performing RNN achieved a Macro-F1 score of 0.851 on the test data, outperforming XGBoost by 0.012 and Random Forest by 0.103. Beyond classification performance, the attribution analysis revealed distinct spectral-temporal patterns associated with the separation of pure and mixed Norway spruce pixels. Attribution values increased particularly after approximately day of year (DOY) 125, while B11 showed the strongest contribution during periods of increasing spectral divergence between the two classes. These findings show that recurrent models can provide competitive classification performance while also revealing the phenological periods and spectral regions that are most relevant for distinguishing closely related forest compositions. Combining time-series classification with explainability therefore provides additional insight into the spectral-temporal mechanisms underlying tree-species discrimination and may support more detailed mapping of mixed forest structures.

Periodic solids with strongly interacting electrons display phenomena – from Mott localization to high-temperature superconductivity – that escape conventional band theory. Efforts to address this unconventional physics have been typically cast in the language of Green's function theories and the associated frequency- and momentum-dependent self-energies capturing strong correlations beyond the reach of mean-field theories. Here, we underscore that frequency- and momentum-dependent self-energies emerge also from mean-field solutions that break translational symmetry, and that these self-energies can display the characteristic fingerprints of strong correlations for certain, physically meaningful, spatial patterns. We showcase this perspective by studying the paradigmatic case of the 2D Hubbard model. We find that already at the mean-field level the doped antiferromagnet displays, in addition to stripes, many quasi-degenerate self-consistent solutions with antiferromagnetic textures organized into antiphase domains separated by hole-rich interfaces, reminiscent of the phase-separation scenario. Remarkably, these solutions map into spectral functions that display very distinctively the phenomenology of the pseudogap, including the nodal-antinodal dichotomy of the self-energy, Fermi arcs, and a complementary Luttinger surface. The result is a conceptually simple description of such an intricate many-body state: the pseudogap emerges as the momentum-space spectral portrait of the real-space interplay between antiferromagnetic domains and the metallic interfaces, which echoes the role of spin-fluctuations identified by nonperturbative methods.

Benchmarks shape how model capabilities are measured and interpreted, yet their scores can obscure whose conventions, contexts, and judgments define successful performance. We introduce a user-centric audit framework that distinguishes users who contribute benchmark content from those whose needs evaluation results are intended to represent. The framework operationalizes seven dimensions across benchmark content, metadata, documentation, and scoring procedures. Across 16 resources, we find that American-centered conventions account for 62.4-94.4% of matched signals in 11 text benchmarks, while geographic visibility in five image-text resources is concentrated and strongly associated with national GDP share (mean Pearson r = 0.780; Spearman p = 0.778). Documentation reveals a systematic source-target gap: all 16 resources operationalize source-user context, but none operationalize target-user context. Paired scoring experiments further show that meaning-preserving answer variants reduce DROP exact-match accuracy from 100.0% to 33.3%, while TruthfulQA exhibits metric-dependent reductions in reference similarity. Our framework and documentation checklist make user assumptions auditable, clarifying what benchmark scores represent and where evaluation claims are supported.

Online memes are often carriers of unsafe content such as disinformation and propaganda. Their rapid evolution in response to news events and strong dependence on contextual information make it difficult for platforms and moderators to detect and moderate them effectively. Moreover, existing moderation studies typically classify memes as safe or unsafe based on individual images, lacking the ability to perform broader provenance analysis, i.e., where a meme originates, how it spreads, what cultural symbolism it represents, whether it poses potential risks to society, etc. To bridge this gap, we introduce MemeTrace, a multimodal LLM-based agent that automatically traces the provenance and potential risks of emerging memes. We evaluate MemeTrace on Meme100, a dataset of 100 real-world memes, using both human annotation and automatic metrics. The results demonstrate that MemeTrace substantially outperforms the VLM baseline. More importantly, MemeTrace achieves higher accuracy in detecting unsafe memes, especially those involving disinformation and sensitive content, and remains robust under visual perturbations.

Existing illumination estimation methods primarily recover lighting distributions for rendering, but typically leave the contributions of individual physical light sources at specific scene locations implicit. We address this gap by leveraging the scene understanding capabilities of vision-language models (VLMs) to connect local illumination with its physical sources. We introduce LAMP, a framework for point-conditioned, source-aware illumination reasoning from a single RGB image with marked query points. LAMP constructs a shared inventory of active light sources, ranks contributing sources for each query, and predicts their 3D query-to-source directions. We supervise these relationships using procedurally controlled indoor scenes, reusing independently rendered source responses to derive contribution targets across lighting configurations and obtaining exact directions from source and query positions. On the Infinigen test set, fine-tuning Qwen3-VL-8B increases dominant-source 3D PCK@0.5m from 0% to 70.6%, with 94.5% of directions within 12.5 degrees of ground truth among correctly localized sources. Combined with a lightweight network predicting dense RGB irradiance, LAMP achieves probe-rendering quality comparable to prior lighting-estimation methods while additionally localizing individual physical sources in 3D and explicitly attributing local illumination to them.

The scarcity of real-world ground penetrating radar (RealGPR) data is a major barrier to the development of deep learning (DL) methods for GPR data processing. Although simulated GPR (SimGPR) datasets can be used to train DL models, differences in hardware configuration and system parameters between SimGPR and RealGPR create a sim-to-real domain gap that leads to poor generalization of DL models to RealGPR measurements. To address this gap, we present a physics-based far-field GPR domain transfer (FFDT) method to transfer far-field SimGPR data to the corresponding RealGPR-domain data in the same scenario. In FFDT, SimGPR and RealGPR are modeled using angle-dependent transfer functions that capture three-dimensional antenna radiation characteristics. These transfer functions are then combined with ray tracing to enable the sim-to-real data transfer. Extensive studies across diverse objects and soil conditions demonstrate that FFDT-transferred data closely match RealGPR ground truth. To further demonstrate its value for RealGPR-domain dataset generation, FFDT is implemented to build a RealGPR-domain dataset for training a DL model for clutter removal. The model achieves robust performance on RealGPR measurements. With one-time SimGPR and RealGPR system calibration, FFDT enables RealGPR-domain dataset generation without requiring task-specific RealGPR field measurements for DL model training or fine-tuning, thereby addressing the RealGPR data scarcity bottleneck and supporting reliable development and deployment of DL methods for real-world GPR applications.