Papers

Seizure monitoring and detection are crucial for reducing the morbidity and mortality associated with seizures. Current epilepsy care, often involving expensive video-electroencephalography (VEEG) monitoring, requires specialized expertise and is limited to in-hospital settings, and intrusive in nature. Seizure diaries, on the other hand, suffer from unreliability due to under-reporting, leading to incorrect therapeutic decisions. Wearable non-invasive seizure detection may offer a more tolerable and feasible solution for long-term ambulatory monitoring. This study explores a wearable remote monitoring system utilizing a single wrist-worn accelerometer device and capable of detecting multiple types of seizures, including shorter duration events. We enrolled 79 patients under video-electroencephalography monitoring to wear accelerometer devices and collect data. Concurrent VEEG recordings were reviewed by board-certified epileptologists to produce annotations, including seizure onset, offset, and seizure type. Using this data, we constructed a deep neural network based on the time-series ResNet architecture, which could discriminate among seizure and non-seizure events. Our proposed approach achieved a seizure detection sensitivity of 95.65% and an overall false alarm rate of 0.15/24 hours during the evaluation, which spanned 5576 hours of total recording. Additionally, it resulted in an area under the receiver operating characteristic curve (AUC-ROC) of 0.98 and an area under the precision-re call curve (AUC-PRC) of 0.67 when averaged over 20 patients who experienced 46 convulsive seizures. These promising results suggest that the proposed seizure detection system can be effectively used for long-term ambulatory seizure monitoring. Future steps include validating our findings in larger datasets and assessing the utility of detection for additional seizure types.

Existing data-driven infrared small target detection (ISTD) methods typically require large-scale datasets with accurate pixel-level annotations for model training. However, such labor-intensive requirements are difficult to satisfy in real-world applications due to the heavy reliance on expert knowledge and the inherently weak distinctiveness of infrared small targets. Consequently, the presence of noisy labels during model training is inevitable, which can severely mislead the learning of target perception toward spurious patterns. To address this challenge, we propose Prompt and Refinement (PAR), a label-noise-robust asymmetric mutual learning paradigm for ISTD. Specifically, PAR comprises a pretrained Segment Anything Model (SAM) and an ISTD-specific detector trained from scratch, which learn collaboratively through a peer-teaching scheme. Coupled with local contrast regularity, the predictions of the two asymmetric peer models are mutually exploited as rectification cues for the supervisory masks of their counterparts. The interaction between complementary inductive biases effectively prevents the label correction process from degenerating into the self-confirmation loop of a single model, enabling progressive refinement of the annotations toward intrinsic target characteristics. In addition, the detector predictions are utilized as corrective mask prompts to facilitate task-specific adaptation of the vision foundation model. Moreover, an evidential uncertainty estimation strategy is introduced into the optimization process to further alleviate the adverse effects of noisy labels. Extensive experiments under diverse noisy label scenarios on three ISTD datasets demonstrate that PAR consistently achieves state-of-the-art performance.

In this work, we study bimerons (in-plane skyrmions) and their clusters composed of states with mixed topological indices. These clusters provide flexibility in controlling the skyrmion Hall angle, which depends on the cluster's total topological charge. First, using the simplest such cluster, a bimeron-antibimeron pair, we examine its stability with the geodesic nudged elastic band method and identify three mechanisms – bimeron separation, interchange, and annihilation – with comparable energy barriers that determine the pair's overall stability. This provides an upper bound for the stability of more sophisticated clusters of bimerons and antibimerons. Then, we numerically study the clusters' dynamics for currents applied in-plane and perpendicular to the plane, corresponding to the Zhang-Li and Slonczewski mechanisms. Finally, we analyze the dynamics of a wide variety of bimeron clusters within the Thiele approach and show that their velocities always lie on a specific ellipse whose parameters are fully governed by the model Hamiltonian.

Electric vehicles and heat pumps add large, time-flexible loads to low-voltage distribution feeders. When each household independently minimizes its own charging cost against the same electricity price, this flexible demand piles into the cheapest hours, and the resulting synchronization can stress the feeder more than uncoordinated operation does. This work quantifies that effect on a rural low-voltage feeder (97 buses, 250 kVA transformer) across 160 simulation runs spanning seven scenarios, four seasons, and electric-vehicle, heat-pump, and photovoltaic penetration from 0 to 90 %. At 30 % penetration of each device, this decentralized cost minimization drives the transformer to 253 % of rated capacity, more than twice the level of uncoordinated charging, and up to 478 % at 90 % electric-vehicle penetration, with the first undervoltage violations at the far end of the longest feeder arm. The ranking is device-mix dependent: each 30 pp of heat-pump penetration adds about twelve times more to the price-based peak than to any fixed schedule, so price-based scheduling becomes the worst strategy once heat pumps are present. Only a central schedule that explicitly caps the transformer removes these demand-driven violations, while at high photovoltaic penetration a summer floor from midday reverse power flow bounds what any demand-side strategy can achieve. A regulated time-variable grid tariff added to the spot price provides no relief for heat-pump-driven synchronization and only marginal relief for electric-vehicle-driven synchronization in autumn. The simulation and optimization code is available at https://github.com/lukas-wagner/SynchronizationEffectsAnalysis.

This paper revisits the distributed learning problem for training a multinomial logistic regression model with the Federated Averaging ($\texttt{FedAvg}$) algorithm. We concentrate on a scenario with arbitrarily large stepsizes and heterogeneous update rules where the devices may perform a different number of local updates in each round. We show that, with linearly separable data, $\texttt{FedAvg}$ is stable with any stepsizes and the objective values converge to zero at the rate of ${\cal O}(1/R)$, where $R$ is the number of communication rounds. Our result also demonstrates that the effects of device heterogeneity vanish asymptotically. For sufficiently large $R$, the objective values decrease monotonically and is bounded by ${\cal O}( 1 / (R T_{\rm avg}))$, where $T_{\rm avg}$ is the average number of local update steps per communication round across devices. Numerical experiments support our findings.

An important role in the theory of conflict-avoiding codes is played by a quantity involving the multiplicative order of $4$ modulo $d$, with $d$ running over all positive divisors of an odd integer $n$. We establish results on its distribution as $n$ varies over the odd prime numbers, respectively odd integers, using techniques from the study of Artin's primitive root conjecture. We also point out some other interpretations of this quantity, with one of them making a connection with the factorization of $X^n-1$ into irreducibles over the finite field $\mathbb F_2$.

Task-fMRI connectomes reveal state-dependent neural reconfigurations, yet conventional methods marginalize these signals by aggregating distinct conditions into static pairwise graphs, thereby obscuring condition-specific multi-ROI organization. We introduce CoHyFuse, a condition-aware ROI-centered hypergraph framework that constructs a task-state-specific incidence matrix from condition-wise functional connectivity (FC)-profile embeddings, allowing the same ROI to form different multi-ROI hyperedges across task phases. Condition-specific neighborhood sizes $K_q$ further adapt the hyperedge scale to each task state, and the resulting condition embeddings are fused with a complementary whole-session FC branch for prediction. In the AABC cohort (N=1,074), CoHyFuse achieved the best mean out-of-fold predictive performance among evaluated baselines on FACENAME Fluid Cognition Composite (FCC) prediction (7.83$\pm$0.10 MAE, 0.439$\pm$0.026 \(R^2\)) and VISMOTOR age prediction (7.52$\pm$0.37 MAE, 0.592$\pm$0.022 \(R^2\)). In an auxiliary CMI-HBN attention-deficit/hyperactivity disorder (ADHD) classification benchmark (N=223), CoHyFuse obtained 72.0$\pm$2.1% macro-AUC and 74.2$\pm$2.9% accuracy. Ablation studies support the contributions of condition-wise incidence construction and dual-view fusion, suggesting that state-resolved ROI-set structure provides complementary predictive information beyond whole-session FC alone. Occlusion analysis identifies the Distraction condition as the primary driver of model prediction, pointing toward the Salience/Ventral Attention Network (SAN)–FrontoParietal Network (FPN) and within-SAN hyperedge-defined ROI-set motifs as candidate model-relevant patterns. This framework provides an interpretable, state-resolved view of the connectome for downstream cohort analysis.

The last mile toward enterprise AGI is a company that runs itself. Training and adapting such agents require longitudinal enterprise data, which remain scarce, costly to acquire, and often restricted by privacy constraints. Historical archives are also frequently incomplete and record only what actually happened. They cannot show the outcomes of alternative decisions. We introduce MiniCorp, an office simulator for studying how agents can collectively run a company while generating enterprise data at scale. Using an e-commerce company as a demonstration, MiniCorp connects two interacting worlds. The external world models customers, dynamic competitors, and market mechanisms. The internal world consists of agents that observe events, discuss their options, and make strategic decisions. These decisions have lasting effects on the market, and the resulting feedback informs the firm's later decisions. As the firm and market interact, MiniCorp continuously records the agents' communications and decisions. These records preserve the information available at the time and the business results that followed. Checkpointing allows the same situation to be replayed under different decisions, providing comparisons unavailable in static archives. We evaluate end-to-end fidelity against patterns reported in empirical studies of real markets. These evaluations provide agents with realistic market feedback and reduce the risk that they learn to exploit flaws in the simulator. Our experiments show agents coordinating across roles and adapting their decisions to market feedback. With explicit long-term strategic guidance, they also sustain advertising exploration despite weak early returns. MiniCorp thus provides an environment for studying AI-run companies and a scalable source of longitudinal and counterfactual enterprise data for agent training and evaluation.

Multi-view panoptic segmentation assigns a semantic class and a scene-level instance ID to every pixel of an unordered set of images, and recent feed-forward 3D models predict these labels for the input views in a single forward pass. Their predictions, however, have been evaluated with the scene-level PQ (PQ^scene) borrowed from per-scene optimization methods, typically on rendered held-out views. PQ^scene tiles all views of a scene into a single image, so that a missed appearance or a change of ID lowers the score of the matched pair only in proportion to its area. We propose View-Consistent Panoptic Quality (VC-PQ), which extends PQ from a single image to a set of input views, counts equally every view in which an instance is visible, and penalizes a prediction that is not visible in the same views as its ground truth. A decomposition of VC-PQ attributes the score a method loses to mask accuracy, view consistency, and the matching threshold. A single additional parameter recovers the area weighting of tiling for comparison. Under a fixed evaluation protocol on ScanNet++ and ScanNetv2, recent feed-forward methods are evaluated with VC-PQ and PQ^scene, and the decomposition shows where each of them loses its score. Controlled perturbations of the ground truth show that VC-PQ responds to the number of views in which an instance is missed or changes ID, whereas PQ^scene responds to their area. The aim of this work is to make view consistency part of the evaluation of multi-view panoptic segmentation, with VC-PQ reported alongside PQ^scene. The evaluation code and the input sets of the protocol are available at https://github.com/Renem24/vcpq.

Real-world thermal super-resolution (SR) is constrained by limited sensor resolution and the difficulty of obtaining corresponding high-resolution (HR) observations for direct model supervision. Conventional SR methods typically construct training pairs by treating captured thermal images with real-world degradations as HR references and applying predefined degradation to generate synthetic low-resolution (LR) inputs. Such a construction not only introduces a domain gap between synthetic and captured LR observations but also retains acquisition degradations in the supervision. To address this issue, we propose AstraSR, a real-world thermal SR method guided by GPT-6 Astra, a frontier multimodal generative model endowed with emergent and transformative visual capabilities. Specifically, we construct a dataset of image pairs by using captured LR thermal images to condition GPT-based HR reference. We develop a direct generative supervision strategy that learns from captured thermal inputs paired with GPT-generated HR references. Pixel, gradient, and perceptual losses jointly supervise the transfer of intensity patterns, structural boundaries, and visual details from the generated references. Qualitative comparisons with seven existing state-of-the-art real-world SR methods show continuous object contours, distinct structural boundaries, and smooth intensity transitions in the thermal scenes. These results demonstrate that AstraSR outperforms existing real-world SR methods in both thermal clarity and structural coherence.

We study an area-preserving curvature flow for planar open curves whose end points lie on skew lines and meet the lines at general contact angles. In previous studies of moving boundary problems for higher-order geometric gradient flows, the right-angle condition has often been imposed to ensure that the evolving curves naturally satisfy the boundary conditions. However, under general contact-angle conditions, the evolution must be described by a free boundary problem. Under suitable assumptions on the contact angles $\theta_\pm$, the angle $\theta$ between the skew lines, and the initial data $\gamma_0$, we prove global-in-time existence and uniqueness up to reparametrization. We further prove that, after choosing a constant-speed parametrization at each time, the solution converges exponentially to a circular arc with the same signed area as the initial curve. Our analysis is based on a formulation of the evolution as a free boundary problem together with the derivation of uniform a priori estimates.

Recent Text-to-Image (T2I) models achieve remarkable visual image generation performance, but they can still generate NSFW (Not-Safe-For-Work) contents, including violent or explicit images. Existing safety checker mechanisms are largely confined to pre-generation filtering (e.g. prompt-level text classifiers) or post-hoc moderation applied after an image is completely synthesized. However, adversarial attack methods operate over a much broader space. This imbalance highlights the need for a safety mechanism that intervenes during the generation process. We propose an in-generation safety framework that monitors the denoising trajectory and detects emerging NSFW signals from intermediate representations. Rather than merely detecting NSFW generations, our method applies reinforcement learning to generate safe images from NSFW prompts. By coupling in-generation detection with controllable steering, our approach mitigates unsafe trajectories even when NSFW signals emerge after generation has already begun. Experiments results show that our method consistently outperforms existing safe image generation methods across both standard and adversarial evaluation sets, while preserving perceptual quality and prompt fidelity. Code will be released upon acceptance.

We developed a highly sensitive and long-term stable diamond quantum magnetometer by integrating a dual-feedback system that compensates for common- and differential-mode frequency shifts in double-quantum (DQ) 4-Ramsey interferometry. Detecting extremely weak magnetic-field signals requires extensive signal accumulation to enhance the signal-to-noise ratio. However, deviations from the optimal operating frequencies caused by thermal drift or environmental magnetic noise severely hinder the long-term stability of such a sensing system. To overcome the limitation, we first derived analytical solutions for the signal response to common- and differential-mode detunings based on a detuning-inclusive Hamiltonian to theoretically clarify the robustness limit in DQ 4-Ramsey sensing. Guided by this theoretical model, we introduced synchronized frequency modulation and a dual-feedback scheme to independently track and compensate for these frequency shifts while maintaining an optimal magnetic-field response. The proposed method enabled stable continuous operation against temperature variations of 2.8 K. By maintaining a magnetic field sensitivity of 7.0 pT/Hz^1/2 in the frequency range of 5-100 Hz for a prolonged duration of 4 hours, we achieved an extremely low noise floor of 64 fT/Hz^1/2. During prolonged continuous sensing, the system exhibited the expected 1/T^1/2 scaling with integration time T, and prevented both the underestimation of amplitude and waveform distortion. These results establish a robust approach for long-term DQ 4-Ramsey diamond magnetometry and demonstrate its potential for detecting weak biomagnetic signals.

The flat voltage plateau of lithium iron phosphate (LFP)/graphite cells makes electrochemical-state errors and voltage-measurement abnormalities produce similar innovations, complicating state-of-charge (SOC) estimation. This work proposes an electrochemically constrained residual-bias compensation dual extended Kalman filter (RBC-DEKF) with uncertain-initialization commissioning. A thermal control-oriented parameter-grouped single-particle model (CPG-SPMT) provides paired-electrode dynamics and terminal-voltage prediction, while a separately configurable voltage-discrepancy state accommodates systematic residuals. When the initial SOC is uncertain, residual adaptation is suspended over a verified-healthy startup window. Buffered data support electrochemically constrained trajectory matching, followed by calibrated state replay and residual-channel reactivation. Frozen cross-cycle healthy-residual calibration supports commissioning and remains in the voltage prediction after handover. Evaluation covers 216 additive-bias and 168 multiplicative-gain profiles over 24 A123 temperature-drive-cycle trajectories. Relative to Single-EKF, the original RBC-DEKF reduces mean SOC RMSE from 7.646 to 0.170 percentage points for additive bias and from 22.646 to 0.356 percentage points for gain faults. Separately, 18 nonzero-initial-error profiles on three representative 50%-SOC-anchored segments give a full-record mean SOC RMSE of 1.747 percentage points, including startup, for the commissioned realization, versus 11.192 and 11.400 for the always-on RBC-DEKF and conventional joint EKF. Representative delayed bias, 10-s ramp, and gain tests show that the corrected SOC trajectory remains stable after fault introduction. The central contribution is electrochemically constrained temporal coordination of uncertain-state recovery and subsequent voltage-discrepancy accommodation.

The effect of spin-orbit coupling (SOC) in monolayer graphene has been thoroughly studied in the past. Its impact is commonly disregarded since the electronic structure of graphene is almost unaffected. Consequently, SOC has also been ignored in other graphitic systems. In this work, we calculate the intrinsic SOC in stacks of graphene layers from a model that includes all the valence orbitals of carbon as well as the intra-atomic SOC.The leading spin-orbit term in a graphene multilayer combines a Rashba-like spin-flip process with interlayer hopping. We find that this intrinsic term modifies the phase diagram and Fermi surfaces of finite rhombohedral stacks, and it leads to a distinct density-dependence of the magnetization in weak magnetic fields. This Letter shows that, unlike in the single-layer case, SOC has important consequences in multilayer graphitic systems and it should not be neglected when studying their physical properties.

Inspired by a recent gauge field description of shallow water equations [SciPost Phys. 14 (2023) 102], in this work we build a simple Lagrangian formulation for the case of constant potential vorticity. An attractive feature of the construction is that it does not appeal to a Clebsch-like parametrization or the use of auxiliary fields. It is also emphasized that on the equator, where the Coriolis parameter vanishes, the shallow water equations exhibit extra SO(2,1) conformal invariance. Using the group-theoretic method, a new exact solution is found, which is associated with the scaling symmetry transformation entering the conformal group.

We demonstrate giant nonlinearity of the kinetic inductance $L_k$ in a superconducting NbN/Al hybrid strip. In contrast to ordinary superconductors, where the maximum observed current-induced variation of $L_k$ does not exceed tens of percent, we observe a 50-fold change in $L_k$. Physically, this effect is connected with the proximity-induced superconductivity in the Al layer and the specific choice of parameters for which the dependence of the superconducting current $I$ on the Cooper pair momentum $\hbar q$ exhibits a plateau at a current $I^*$ well below the depairing current. Formally, this leads to $L_k \sim dq/dI \to \infty$ as $I \to I^*$. We believe that our results open a new route for creating the next generation of highly sensitive detectors of electromagnetic radiation, calorimeters, parametric amplifiers, and magnetometers.

"Don't tune hyperparameters on the test set" is often stated in machine learning textbooks. Violating it is considered a cardinal sin that produces misleadingly optimistic results, corrupts benchmark integrity, and thus can even be interpreted as scientific fraud. Yet evidence suggests that test set hyperparameter tuning does occur in practice, making it all the more important to understand its actual consequences. So how bad is it, really? In this work we question this dogma and put it to an empirical test. We systematically study the magnitude of the performance inflation caused by tuning the hyperparameters on the test set for MNIST-1D, CIFAR-10, and three tasks from the GLUE benchmark. Our experiments show that while the effect is real and significant, it is frequently small relative to other sources of noise. In many cases, we find that tuning on the test set recovers exactly the same model as when tuning on the validation set. Most importantly, we find that the rankings of models remain essentially preserved after tuning on the test set and therefore that consistent test-set tuning may not invalidate benchmarks or model selection. Our results call for a more nuanced view of tuning hyperparameters on the test set, stimulating researchers to openly report test tuning.

Maximum likelihood (ML) decoders perform quantum error correction by measuring a set of syndromes, classically computing an optimal recovery based on these syndromes and the noise model, and then performing that recovery. ML decoders are well understood for the toric code and generalize straightforwardly to other surface codes with Abelian anyons, where they can saturate information-theoretic decoding thresholds. On the other hand, codes with non-Abelian anyons have become increasingly relevant because they can support transversal or constant-depth non-Clifford gates and admit efficient preparation protocols in modern quantum devices. However, they present new challenges for decoding: non-Abelian anyons cannot in general be deterministically annihilated, and the logical operators cannot be classified by Abelian homology. In this work, we show that refined syndrome sets, including measurements that resolve non-Abelian anyon fusion outcomes, systematically improve the performance of ML decoders. We demonstrate this concretely by constructing explicit decoders for the $D_4$ quantum double model under charge noise, and obtaining numerical results for thresholds via mappings to classical statistical-mechanics models. We show that there is a nonzero noise threshold below which the entire logical qudit can be recovered with a single round of measurements.

Bilateral coronary tree (LCA/RCA) reconstruction in cardiac CT angiography (CCTA) faces $64^3$ discretization, territory ambiguity, and ostium drift. Within GT-foreground $128^3$ ROIs ($\pm 15$-voxel margin, 9.8–12.0 mm; seed-free in ROI; whole-volume localization upstream), we propose a canonical $(Z_{IS},Y_{PA},X_{LR})$ multi-task network and scalar-potential arborescence solver. Volumes, centerlines, and thick-stem ostia map to $(Z_{IS},Y_{PA},X_{LR})$, extending a 3D pyramid encoder and tubular decoder with a $128^3\to 64^3$ skip, stop-gradient $128^3$ residual branch, 3D sub-voxel displacement head $\Delta\hat{u}(x)$, ipsilateral head $(P_{LCA},P_{RCA})$, and $32^3$ ostium head. From a recall-balanced anchor, 327,385 active parameters (398,843 total) are fine-tuned in a 30-epoch AdamW loop ($\epsilon=10^{-8}$, 7 groups) with L2-SP/anchor blending. Our solver couples Valsalva-STJ calibration, $128^3$ rooting, contralateral inhibition, radius-ratio pruning, normal-plane pullback, and collision reduction. On OOS-50 (Cases 251–300, 50 volumes, 100 trees; 201–250 isolated) vs. 3D U-Net (skeleton/$M_3$–$M_5$) and 5 proxies, it reaches $2.7671\pm 1.1385$ mm $CD_{CL}^{union}$ (SD 0.0415), $3.2338\pm 1.3557$ mm $CD_{CL}^{arc1mm}$, and $3.2926\pm 2.6253$ mm $CD_{ipsi}$ (arc1mm 3.7460 mm, node $p_{Holm}\le 0.0400$ vs. Ext-0b [arc1mm $p=0.1629/0.2187$]; 76.33% junction [23.04% 1-to-1], 79.65% branch F1). On Dev-14 (14 volumes, 28 trees), it reaches $2.1362\pm 0.5041$ mm $CD_{CL}^{union}$ (SD 0.0369), $2.3798\pm 0.5725$ mm $CD_{CL}^{arc1mm}$, and $2.2480\pm 0.6835$ mm $CD_{ipsi}$ (arc1mm 2.4951 mm, node $CD_{CL}^{union}$ $p_{Holm}=0.0494$ vs. Ext-0b [$CD_{ipsi}$/arc1mm $p_{Holm}=0.1353/0.0676$–$0.0906$]; 56.76%/77.74% F1@1/2mm; 79.95% junction [21.06% 1-to-1], 83.80% branch F1, 4.6880/4.9298 mm ostia), and $N=64$ $CD_{CL}^{union}=2.6291\pm 1.0629$ mm (arc1mm 3.0470, $CD_{ipsi}$ 3.0641).

Instance-level roof-to-footprint offset (RFO) prediction is central to extracting building footprints from off-nadir imagery. Query-based pipelines commonly use high-dimensional instance tokens to predict signed two-dimensional RFOs. We investigate whether RFO prediction can instead use a compact offset token. Under local pinhole projection and vertical-extrusion assumptions, the idealized RFO map admits a five-parameter sufficient descriptor comprising intrinsic shape, composite amplitude, and relative geometry. This factorization provides a structural prior for a five-dimensional offset token, whose channels learn task-relevant latent representations through end-to-end training. Based on this design, we propose LoDEOT, which retains high-dimensional instance tokens for detection and segmentation but maps instance-token, concentration-gated roof, and box-mask evidence to a five-dimensional offset token followed by an independent two-dimensional readout. Known denoising-query target indices further align each supervised decoder-layer estimate with the same clean instance RFO, organizing successive predictions as target-aligned recovery under perturbed query conditions. Experiments on five real-world building datasets demonstrate the effectiveness of LoDEOT for building footprint extraction. Experiments on real-world building datasets demonstrate that a five-dimensional offset token can support accurate RFO prediction. On BONAI, LoDEOT achieves the best roof-detection bAP and bAP50 and leads all five offset-corrected footprint metrics among the evaluated end-to-end methods, with FAP50 of 54.58 and mEPE of 5.23 pixels. Its FAP50 exceeds those of the evaluated end-to-end baselines by 7.56-16.85 percentage points.

Real-world users often exhibit highly heterogeneous preferences over multiple objectives for LLM responses. A lightweight aligner can tailor these responses to individual preferences, but scarce user-specific feedback makes personalized training difficult. Learning shared initializations across users can support few-shot adaptation. However, heterogeneous preferences and competing objectives cause gradient conflicts across users and within each user, hindering effective initialization learning. This raises a central question: how can we collaboratively learn aligner initializations that support few-shot adaptation to diverse user preferences? To answer this question, we propose Approximate Pareto Optimality (APO). We first group users whose updates are compatible, so that their information can be combined with less interference. Within each group, we combine gradient descent with controlled ascent to coordinate competing objectives and move towards preference-specific points on the Pareto front. This produces an initialization that is close to the optima of the users in the group. We then iteratively refine it using updates from few-shot local adaptation, making it more effective for personalization. Furthermore, we establish conditional suboptimality bounds for a one-local-step collaborative update and characterize how initialization error affects subsequent stochastic adaptation. Experiments on Fed-ChatbotPA and UltraFeedback show consistent improvements over existing methods using only 20 local examples.

Training a small projector between a frozen vision encoder and language model is an established approach to multimodal learning. As the parameter count of language models scales dramatically, we revisit which vision capabilities this approach can add while keeping their pretrained weights fixed. Here we train a 50M parameter projector from the vision encoder of Kimi K2.6 to GLM 5.2 and 5.3, both models without native vision capabilities, and further present a reproducible recipe for training these adapters at scale. We study the following: (a) how vision capabilities of multimodal models scale as purely the language model side scales, and (b) what specific vision capabilities are able to be imbued into a pure language model at scale, and which ones remain limited. We evaluate on MMMU-Pro and BLINK, examining both overall performance and results on individual visual tasks.

Rapid progress in video generation has led to a plethora of models that differ substantially in capability and generation cost. This raises a natural question: can each request be efficiently routed to an appropriate model? We find that even when the consensus of the other annotators is used as an oracle, it agrees with each annotator's own favorite only 34-55% of the time. Motivated by this observation, we introduce TasteRoute, a personalized video-generation router that selects a generator jointly based on the input request, user preferences, and available generation budget. Across text-to-video and image-to-video settings, TasteRoute is competitive with strong simple baselines on preference routing while reducing average generation cost. The cost saving increases under higher budget caps. Finally, we release TasteRoute-3k, a human-annotated dataset containing multi-model video comparisons, quality judgments, preference rankings, and user-profile signals to facilitate future research on personalized and cost-aware video routing.

We show that an intense low-frequency laser field can qualitatively modify the local extremal structure of an interband energy gap in a crystal. Using an instantaneous laser-dressed basis obtained from the nonperturbative diagonalization of the multiband crystal Hamiltonian, we derive a quartic normal form for the dressed gap near a high-symmetry point. At a critical field, the quadratic part of this normal form changes sign, and the gap experiences a bifurcation: a single symmetric minimum splits into two symmetry-related minima separated by a local maximum. We identify the microscopic origin of this bifurcation by decomposing the quadratic part into a homogeneous Stark contribution and a field-modified band-curvature contribution. The latter originates from the multiband redistribution of interband polarizations and provides the mechanism for the field-induced change of the gap extrema. A realization for the Kronig–Penney model demonstrates the same bifurcation at multiple high-symmetry points and for different interband gaps. As an illustration of its spectral consequences, we show that the bifurcation gives rise to an additional van Hove feature in the instantaneous interband susceptibility above the absorption edge. Our results establish a general mechanism by which strong-field dressing modifies the local geometry of the electronic spectrum.

Masked diffusion language models (dLLMs) generate text by iteratively denoising masked positions, re-predicting each token multiple times before it is committed. An autoregressive decoder exposes an answer's distribution once, at the step that commits it; a dLLM exposes it at every denoising step before commitment, and we show that an adversary can exploit this. Since an answer remains open to revision over many denoising steps, an adversary with access to internal activations can watch how likely the model is to produce a chosen answer and adjust the intervention accordingly. Building on this observation, we study targeted bias injection, an attack that steers a frozen dLLM toward a demographic answer selected by the adversary. The attack uses a simple proportional-integral (PI) controller that tracks the target-answer probability during denoising and adapts the strength of a steering vector on the fly. On ambiguous BBQ questions where the correct answer is abstention, our attack raises LLaDA-8B-Instruct's preference for the targeted group from 1.8 to 16.7 percentage points, more than three times the strongest fixed-strength steering baseline, and on SocialStigmaQA it raises the selection of stigmatizing answers from 17.6% to 58.1%. Fitted to other demographic targets, the same attack shifts answers by up to 37 percentage points, and each attack takes about 40 minutes on one GPU. On the primary target, feedback is what makes the attack work: constant steering at the same average strength over the token-committing steps produces a far smaller shift while corrupting nearly three times as many outputs, and a constant strength set separately for each example still falls well short. Our findings identify the denoising trajectory as a new control channel in dLLMs and call for bias audits that examine the serving stack rather than the frozen model alone.

We study environment-dependent annihilation of quasi-Dirac Higgsino dark matter coupled to an axion-like particle with a mass slightly below the Higgsino-pair threshold. Axion-mediated annihilation becomes more important relative to electroweak annihilation in the Sun and Galactic halo than during cosmological reannihilation following late nonthermal production. This separation arises from both the low-velocity enhancement of the subthreshold propagator and the change from a coannihilating Higgsino multiplet to the lightest neutral state. In solar capture-annihilation equilibrium, a dominant annihilation channel into massless dark photons suppresses the high-energy neutrino signal while leaving capture efficient. Higgsinos can consequently account for a few percent of the total dark matter abundance while satisfying solar-neutrino constraints and retaining an appreciable terrestrial inelastic-scattering rate, with implications for the LZ high-recoil candidate. In the Galactic halo, the enhanced cross section instead produces correlated continuum and line gamma-ray signals, with the lines potentially close to the present sensitivity for a cuspy Galactic dark-matter profile. We discuss supersymmetric realizations based on generalized mirage mediation and hidden strong dynamics, as well as an extended Higgs-sector alternative. The dark photons also contribute to dark radiation, providing a complementary cosmological probe.

We show that, in equal characteristic $p$, the Hodge-Witt polygon can only move down and domino numbers can only increase under specialization in smooth proper families. We also show that the Hodge polygon of the generic fiber of a proper smooth $p$-adic formal scheme over $\mathcal{O}_C$ lies on or above the Hodge-Witt polygon of its special fiber, where $C$ is a complete algebraically closed extension of $\mathbb{Q}_p$. We further show that, in this mixed-characteristic setting, if the crystalline cohomology of the special fiber is torsion-free in degrees $n$ and $n+1$, then the degree-$n$ Hodge polygon of the generic fiber lies on or below the abstract Hodge polygon of the special fiber. We illustrate these comparisons with explicit examples.

Post-hoc class activation maps (CAMs) are a standard tool for inspecting the evidence behind an image classifier's predictions, yet nothing in ordinary training encourages these maps to be spatially appropriate. We study whether inexpensive spatial supervision can improve a classifier's own predicted-class Grad-CAM without ever optimizing an attribution map. Box-Guided Evidence Routing (BGER) trains a lightweight gate on the final feature map under box or mask supervision and routes classification through the gated features, while Grad-CAM is computed separately at the pre-gate representation, so the evaluated map never enters the training objective. With a BCE routing loss, BGER raises MaxBoxAccV2 from $0.584$ to $0.715$ on CUB-200-2011 and from $0.757$ to $0.832$ on Stanford Dogs at comparable accuracy. Matched controls attribute most of the ResNet-50 gain to the spatial supervision reshaping the backbone rather than to routing itself: when classification bypasses the gate, most of the improvement remains, and detaching gradients through the gate leaves the ResNet-50 result nearly unchanged. The same detachment preserves most of the gain in two DenseNet-121 chest X-ray settings but removes the apparent gain on Swin-T, and directly supervising the CAM reaches stronger localization at a larger accuracy cost. Overall, spatial supervision can improve separately evaluated post-hoc CAMs, but both the mechanism and the size of the benefit depend on the architecture and the evaluation setting.

We study continuous Brascamp–Lieb inequalities associated with isotropic measures on the sphere $S^{n-1}$ and on the Grassmannian $G(n,k)$. We give two proofs of the continuous Grassmannian inequality. The first proof approximates the continuous isotropic measure by discrete isotropic measures and then applies the equal-rank geometric Brascamp–Lieb inequality. In the rank-one case, the construction of the discrete isotropic measures follows from Barthe's work, which is standard, whereas in the higher-rank case the construction of finite approximate isotropic measures is given by a Tchakaloff-type theorem. Weak convergence, geometric Brascamp–Lieb inequality and Fatou's lemma then yield the continuous inequalities. The second proof generalizes the heat-semigroup preservation method of Barthe and Huet to isotropic measures on $G(n,k)$. We prove a heat-semigroup preservation theorem in the Grassmannian setting and show that this theorem yields a Grassmannian Brascamp–Lieb inequality. At the end, we also give a sketch for the mixed-rank case of the continuous Brascamp–Lieb inequality.