Accurate LiDAR-camera calibration is crucial for multi-sensor systems. However, traditional methods often rely on physical targets, which are impractical for real-world deployment. Moreover, even carefully calibrated extrinsics can degrade over time due to sensor drift or external disturbances, necessitating periodic recalibration. To address these challenges, we present a Targetless LiDAR-Camera Calibration (TLC-Calib) that jointly optimizes sensor poses with a neural Gaussian-based scene representation. Reliable LiDAR points are frozen as anchor Gaussians to preserve global structure, while auxiliary Gaussians prevent local overfitting under noisy initialization. Our fully differentiable pipeline with photometric and geometric regularization achieves robust and generalizable calibration, consistently outperforming existing targetless methods on KITTI-360, Waymo, and FAST-LIVO2, and surpassing even the provided calibrations in rendering quality.

High-resolution optical microscopy has transformed biological imaging, yet its resolution and contrast deteriorate with depth due to multiple light scattering. Conventional correction strategies typically approximate the medium as one or a few discrete layers. While effective in the presence of dominant scattering layers, these approaches break down in thick, volumetric tissues, where accurate modeling would require an impractically large number of layers. To address this challenge, we introduce an inverse-scattering framework that represents the entire volume as a superposition of angular deflectors, each corresponding to scattering at a specific angle. This angular formulation is particularly well suited to biological tissues, where narrow angular spread due to the dominant forward scattering allow most multiple scattering to be captured with relatively few components. Within this framework, we solve the inverse problem by progressively incorporating contributions from small to large deflection angles. Applied to simulations and in vivo reflection-mode imaging through intact mouse skull, our method reconstructs up to 121 angular components, converting ~80% of multiply scattered light into signal. This enables non-invasive visualization of osteocytes in the skull that remain inaccessible to existing layer-based methods. These results establish the scattering-angle basis as a deterministic framework for imaging through complex media, paving the way for high-resolution microscopy deep inside living tissues.

This study investigates the loss of generalization ability in neural networks, revisiting warm-starting experiments from Ash & Adams. Our empirical analysis reveals that common methods designed to enhance plasticity by maintaining trainability provide limited benefits to generalization. While reinitializing the network can be effective, it also risks losing valuable prior knowledge. To this end, we introduce the Hare & Tortoise, inspired by the brain's complementary learning system. Hare & Tortoise consists of two components: the Hare network, which rapidly adapts to new information analogously to the hippocampus, and the Tortoise network, which gradually integrates knowledge akin to the neocortex. By periodically reinitializing the Hare network to the Tortoise's weights, our method preserves plasticity while retaining general knowledge. Hare & Tortoise can effectively maintain the network's ability to generalize, which improves advanced reinforcement learning algorithms on the Atari-100k benchmark. The code is available at this https URL

In vision-language models (VLMs), prompt tuning has shown its effectiveness in adapting models to downstream tasks. However, learned prompts struggle to generalize to unseen classes, as they tend to overfit to the classes that are targeted during prompt tuning. Examining failure cases, we observed that learned prompts disrupt the semantics of unseen classes, generating text embeddings with incorrect semantic relationships among classes. To address this, we propose Similarity Alignment Regularization (SAR), which regularizes learnable prompts to preserve the semantic relationships among classes captured by hand-crafted prompts. Specifically, we first obtain novel classes related to base classes using ChatGPT-4o and utilize them as potential unseen classes during prompt tuning. Then, by targeting both base and novel classes, SAR aligns the similarity relationships among text embeddings generated by learnable prompts with the similarity relationships from hand-crafted prompts. Extensive experiments applying SAR to existing prompt tuning methods demonstrate its effectiveness in improving generalization to unseen classes.

Inverse problems involving partial differential equations (PDEs) can be seen as discovering a mapping from measurement data to unknown quantities, often framed within an operator learning approach. However, existing methods typically rely on large amounts of labeled training data, which is impractical for most real-world applications. Moreover, these supervised models may fail to capture the underlying physical principles accurately. To address these limitations, we propose a novel architecture called Physics-Informed Deep Inverse Operator Networks (PI-DIONs), which can learn the solution operator of PDE-based inverse problems without labeled training data. We extend the stability estimates established in the inverse problem literature to the operator learning framework, thereby providing a robust theoretical foundation for our method. These estimates guarantee that the proposed model, trained on a finite sample and grid, generalizes effectively across the entire domain and function space. Extensive experiments are conducted to demonstrate that PI-DIONs can effectively and accurately learn the solution operators of the inverse problems without the need for labeled data.

Federated learning (FL) often suffers from performance degradation due to key challenges such as data heterogeneity and communication constraints. To address these limitations, we present a novel FL framework called FedWSQ, which integrates weight standardization (WS) and the proposed distribution-aware non-uniform quantization (DANUQ). WS enhances FL performance by filtering out biased components in local updates during training, thereby improving the robustness of the model against data heterogeneity and unstable client participation. In addition, DANUQ minimizes quantization errors by leveraging the statistical properties of local model updates. As a result, FedWSQ significantly reduces communication overhead while maintaining superior model accuracy. Extensive experiments on FL benchmark datasets demonstrate that FedWSQ consistently outperforms existing FL methods across various challenging FL settings, including extreme data heterogeneity and ultra-low-bit communication scenarios.

We study how humans learn from AI, leveraging an introduction of an AI-powered Go program (APG) that unexpectedly outperformed the best professional player. We compare the move quality of professional players to APG's superior solutions around its public release. Our analysis of 749,190 moves demonstrates significant improvements in players' move quality, especially in the early stages of the game where uncertainty is highest. This improvement was accompanied by a higher alignment with AI's suggestions and a decreased number and magnitude of errors. Young players show greater improvement, suggesting potential inequality in learning from AI. Further, while players of all skill levels benefit, less skilled players gain higher marginal benefits. These findings have implications for managers seeking to adopt and utilize AI in their organizations.

This paper presents our contributions to the Speech Emotion Recognition in Naturalistic Conditions (SERNC) Challenge, where we address categorical emotion recognition and emotional attribute prediction. To handle the complexities of natural speech, including intra- and inter-subject variability, we propose Multi-level Acoustic-Textual Emotion Representation (MATER), a novel hierarchical framework that integrates acoustic and textual features at the word, utterance, and embedding levels. By fusing low-level lexical and acoustic cues with high-level contextualized representations, MATER effectively captures both fine-grained prosodic variations and semantic nuances. Additionally, we introduce an uncertainty-aware ensemble strategy to mitigate annotator inconsistencies, improving robustness in ambiguous emotional expressions. MATER ranks fourth in both tasks with a Macro-F1 of 41.01% and an average CCC of 0.5928, securing second place in valence prediction with an impressive CCC of 0.6941.

We discuss how the surface gravity can be classically defined for dynamical black holes. In particular we focus on defining the surface gravity for locally defined horizons and compare a number definitions proposed in the literature. We illustrate the differences between the various proposals in the case of an arbitrary dynamical, spherically symmetric black hole spacetime. We also discuss how the trapping horizon formalism of Hayward can be related to other constructions.

We interpret the recent excess in a rare decay of the Higgs boson, $H\to Z\gamma$, using a light axion-like particle (ALP) in the massrange$0.05 - 0.1$ GeV.The dominant decay of such a light ALP is into a pair of collimated photons, whose decay is required to happen before reaching the ECAL detector, such that it mimics a single photon in the detector. It can explain the excess with a coupling $C^{\rm eff}_{aZH} / \Lambda \sim 4 \times 10^{-5}\;{\rm GeV}^{-1}$, while the decay of the ALP before reaching the ECAL requires the diphoton coupling $C^{\rm eff}_{\gamma\gamma}/ \Lambda \ge 0.35 \,{\rm TeV}^{-1} (0.1\,{\rm eV}/m_a)^2$. A potential test would be the rare decay of the $Z$ boson $Z \to a H^* \to a (b \bar b)$ at the Tera-$Z$ option of the future FCC and CEPC. However, it has a branching ratio of only $O(10^{-12})$, and thus barely testable. The production cross section for $pp \to Z^* \to a H$ via the same coupling $C^{\rm eff}_{aZH} / \Lambda$ at the LHC is too small for detection.

Large Language Models store extensive factual knowledge acquired during large-scale pre-training. However, this knowledge is inherently static, reflecting only the state of the world at the time of training. Knowledge editing has emerged as a promising solution for updating outdated or incorrect facts without full retraining. However, most existing locate-and-edit methods primarily focus on token-level likelihood optimization without addressing semantic coherence. Our analysis reveals that such edited knowledge is often encoded as isolated residual streams in the model's latent space, distinct from pre-existing knowledge and bypassing natural reasoning process. To address this, we propose \textsc{Steam}, a semantic-level knowledge editing framework that enhances integration of updated knowledge into the model's knowledge structure. \textsc{Steam} first identifies target representations as semantic anchors for the updated factual association, then guides the internal representation of the edited fact towards these anchors through an alignment loss during optimization. Experimental results demonstrate that \textsc{Steam} improves model's ability to reason with edited knowledge and enhances semantic coherence, underscoring the importance of latent-space alignment for reliable and coherent knowledge editing. The code is available at this https URL.

Charged lepton flavor violation is forbidden in the Standard Model but possible in several new physics scenarios. In many of these models, the radiative decays $\tau^{\pm}\rightarrow\ell^{\pm}\gamma$ ($\ell=e,\mu$) are predicted to have a sizeable probability, making them particularly interesting channels to search at various experiments. An updated search via $\tau^{\pm}\rightarrow\ell^{\pm}\gamma$ using full data of the Belle experiment, corresponding to an integrated luminosity of 988 fb$^{-1}$, is reported for charged lepton flavor violation. No significant excess over background predictions from the Standard Model is observed, and the upper limits on the branching fractions, $\mathcal{B}(\tau^{\pm}\rightarrow \mu^{\pm}\gamma)$ $\leq$ $4.2\times10^{-8}$ and $\mathcal{B}(\tau^{\pm}\rightarrow e^{\pm}\gamma)$ $\leq$ $5.6\times10^{-8}$, are set at 90\% confidence level.

A supermassive black hole (SMBH) at the core of an active galactic nucleus (AGN) provides room for the elusive ultra-light scalar particles (ULSP) to be produced through a phenomenon called \textit{superradiance}. This phenomenon produces a cloud of scalar particles around the black hole by draining its spin angular momentum. In this work, we present a study of the superradiant instability due to a scalar field in the vicinity of the central SMBH in an AGN. We begin by showing that the time-evolution of the gravitational coupling $\alpha$ in a realistic ambiance created by the accretion disk around the SMBH in AGN leads to interesting consequences such as the amplified growth of the scalar cloud, enhancement of the gravitational wave emission rate, and appearance of higher modes of superradiance within the age of the Universe. We then explore the consequence of superradiance on the characteristics of the AGN. Using the Novikov-Thorne model for an accretion disk, we divide the full spectrum into three wavelength bands- X-ray ($10^{-4}-10^{-2}~\mu$m), UV (0.010-0.4~$\mu$m), and Vis-IR (0.4-100~$\mu$m) and observe sudden drops in the time-variations of the luminosities across these bands and Eddington ratio ($f_{\textrm{Edd}}$) with a characteristic timescale of superradiance. Using a uniform distribution of spin and mass of the SMBHs in AGNs, we demonstrate the appearance of depleted regions and accumulations along the boundaries of these regions in the planes of different band-luminosities and $f_{\textrm{Edd}}$. Finally, we discuss some possible signatures of superradiance that can be drawn from the observed time-variation of the AGN luminosities.

Under the assumption that the various evidences of a `95 GeV' excess, seen in data at the Large Electron Positron (LEP) collider as well as the Large Hadron Collider (LHC), correspond to actual signals of new physics Beyond the Standard Model (BSM), we characterise the underlying particle explaining these anomalies in terms of its Charge/Parity (CP) quantum numbers. In doing so, we use $\chi^2$ fits to test the CP-even (scalar) and CP-odd (pseudoscalar) hypotheses and superpositions of these, thus under the assumption of a spin-0 resonance. This is done through the exploitation of $\tau^+\tau^-$ decays, in both their fully hadronic and semi-leptonic modes, in a model-independent way, so that our approach enables one to test a variety of BSM hypotheses, having proven here that the High-Luminosity LHC (HL-LHC) will be in a position to disentangle the CP nature of such a new particle within $\pm(0.27-0.47)$ radians of the true hypothesis at $90\%$ Confidence Level (CL), depending on the assumed background systematics.

We report an extensive study of the properties of carbyne using first-principles calculations. We investigate carbyne's mechanical response to tension, bending, and torsion deformations. Under tension, carbyne is about twice as stiff as the stiffest known materials and has an unrivaled specific strength of up to 7.5*10^7 Nm/kg, requiring a force of ~10 nN to break a single atomic chain. Carbyne has a fairly large room-temperature persistence length of about 14 nm. Surprisingly, the torsional stiffness of carbyne can be zero but can be 'switched on' by appropriate functional groups at the ends. Further, under appropriate termination, carbyne can be switched into a magnetic-semiconductor state by mechanical twisting. We reconstruct the equivalent continuum-elasticity representation, providing the full set of elastic moduli for carbyne, showing its extreme mechanical performance (e.g. a nominal Young's modulus of 32.7 TPa with an effective mechanical thickness of 0.772 A). We also find an interesting coupling between strain and band gap of carbyne, which is strongly increased under tension, from 3.2 to 4.4 eV under a 10% strain. Finally, we study the performance of carbyne as a nanoscale electrical cable, and estimate its chemical stability against self-aggregation, finding an activation barrier of 0.6 eV for the carbyne-carbyne cross-linking reaction and an equilibrium cross-link density for two parallel carbyne chains of 1 cross-link per 17 C atoms (2.2 nm).

DCE-MRI provides information about vascular permeability and tissue perfusion through the acquisition of pharmacokinetic parameters. However, traditional methods for estimating these pharmacokinetic parameters involve fitting tracer kinetic models, which often suffer from computational complexity and low accuracy due to noisy arterial input function (AIF) measurements. Although some deep learning approaches have been proposed to tackle these challenges, most existing methods rely on supervised learning that requires paired input DCE-MRI and labeled pharmacokinetic parameter maps. This dependency on labeled data introduces significant time and resource constraints, as well as potential noise in the labels, making supervised learning methods often impractical. To address these limitations, here we present a novel unpaired deep learning method for estimating both pharmacokinetic parameters and the AIF using a physics-driven CycleGAN approach. Our proposed CycleGAN framework is designed based on the underlying physics model, resulting in a simpler architecture with a single generator and discriminator pair. Crucially, our experimental results indicate that our method, which does not necessitate separate AIF measurements, produces more reliable pharmacokinetic parameters than other techniques.