State-of-the-art medical multi-modal LLMs (med-MLLMs), such as LLaVA-Med and BioMedGPT, primarily depend on scaling model size and data volume, with training driven largely by autoregressive objectives. However, we reveal that this approach can lead to weak vision-language alignment, making these models overly dependent on costly instruction-following data. To address this, we introduce ExGra-Med, a novel multi-graph alignment framework that jointly aligns images, instruction responses, and extended captions in the latent space, advancing semantic grounding and cross-modal coherence. To scale to large LLMs (e.g., LLaMA-7B), we develop an efficient end-to-end training scheme using black-box gradient estimation, enabling fast and scalable optimization. Empirically, ExGra-Med matches LLaVA-Med's performance using just 10% of the pre-training data, achieving a 20.13% gain on VQA-RAD and approaching full-data performance. It also outperforms strong baselines like BioMedGPT and RadFM on visual chatbot and zero-shot classification tasks, demonstrating its promise for efficient, high-quality vision-language integration in medical AI.

Whole slide pathology image classification presents challenges due to gigapixel image sizes and limited annotation labels, hindering model generalization. This paper introduces a prompt learning method to adapt large vision-language models for few-shot pathology classification. We first extend the Prov-GigaPath vision foundation model, pre-trained on 1.3 billion pathology image tiles, into a vision-language model by adding adaptors and aligning it with medical text encoders via contrastive learning on 923K image-text pairs. The model is then used to extract visual features and text embeddings from few-shot annotations and fine-tunes with learnable prompt embeddings. Unlike prior methods that combine prompts with frozen features using prefix embeddings or self-attention, we propose multi-granular attention that compares interactions between learnable prompts with individual image patches and groups of them. This approach improves the model's ability to capture both fine-grained details and broader context, enhancing its recognition of complex patterns across sub-regions. To further improve accuracy, we leverage (unbalanced) optimal transport-based visual-text distance to secure model robustness by mitigating perturbations that might occur during the data augmentation process. Empirical experiments on lung, kidney, and breast pathology modalities validate the effectiveness of our approach; thereby, we surpass several of the latest competitors and consistently improve performance across diverse architectures, including CLIP, PLIP, and Prov-GigaPath integrated PLIP.

Diffusion models are central to generative modeling and have been adapted to graphs by diffusing adjacency matrix representations. The challenge of having up to $n!$ such representations for graphs with $n$ nodes is only partially mitigated by using permutation-equivariant learning architectures. Despite their computational efficiency, existing graph diffusion models struggle to distinguish certain graph families, unless graph data are augmented with ad hoc features. This shortcoming stems from enforcing the inductive bias within the learning architecture. In this work, we leverage random matrix theory to analytically extract the spectral properties of the diffusion process, allowing us to push the inductive bias from the architecture into the dynamics. Building on this, we introduce the Dyson Diffusion Model, which employs Dyson's Brownian Motion to capture the spectral dynamics of an Ornstein-Uhlenbeck process on the adjacency matrix while retaining all non-spectral information. We demonstrate that the Dyson Diffusion Model learns graph spectra accurately and outperforms existing graph diffusion models.

Path-wise observables--functionals of stochastic trajectories--are at the heart of time-average statistical mechanics and are central to thermodynamic inequalities such as uncertainty relations, speed limits, and correlation-bounds. They provide a means of thermodynamic inference in the typical situation, when not all dissipative degrees of freedom in a system are experimentally accessible. So far, theories focusing on path-wise observables have been developing in two major directions, diffusion processes and Markov-jump dynamics, in a virtually disjoint manner. Moreover, even the respective results for diffusion and jump dynamics were derived with a patchwork of different approaches that are predominantly indirect. Stochastic calculus was recently shown to provide a direct approach to path-wise observables of diffusion processes, while a corresponding framework for jump dynamics remained elusive. In our work we develop, in an exact parallelism with continuous-space diffusion, a complete stochastic calculus for path-wise observables of Markov-jump processes. We formulate a "Langevin equation" for jump processes, define general path-wise observables, and establish their covariation structure, whereby we fully account for transients and time-inhomogeneous dynamics. We prove the known kinds of thermodynamic inequalities in their most general form and discus saturation conditions. We determine the response of path-wise observables to general (incl. thermal) perturbations and carry out the continuum limit to achieve the complete unification of diffusion and jump dynamics. Our results open new avenues in the direction of discrete-state analogs of generative diffusion models and the learning of stochastic thermodynamics from fluctuating trajectories.

Nuclear magnetic resonance instruments are becoming available to the do-it-yourself community. The challenges encountered in the endeavor to build a magnetic resonance imaging instrument from scratch were confronted in a four-day hackathon at Singapore University of Technology and Design in spring 2024. One day was devoted to educational lectures and three days to system construction and testing. Seventy young researchers from all parts of the world formed six teams focusing on magnet, gradient coil, RF coil, console, system integration, and design, which together produced a working MRI instrument in three days. The different steps, encountered challenges, and their solutions are reported.

Progress of the COVID-19 pandemic was quantified, in the first instance, using the daily number of positive cases recorded by the national public health authorities. Averaged over a seven-day window, the daily incidence of COVID-19 in Germany reveals clear sections of exponential growth or decay in propagation of infection. Comparing with incidence profiles according to onset-of-symptoms shows that reporting of cases involves variable delays. Observed changes in exponential rates come from growing public awareness, governmental restrictions and their later relaxation, annual holidays, seasonal variation, emergence of new viral variants, and from mass vaccination. Combining the measured rates with epidemiological parameters established for SARS-CoV-2 yields the dynamics of change in disease transmission. Combined with the distribution of serial intervals (or generation times), the rate gives basic and instantaneous values of the reproduction number that govern development and ultimate outcome of the epidemic. Herd immunity requires vaccination of approximately seventy percent of the population, but this increases to circa eighty percent for the more transmissible Alpha-variant. Beyond this point, progressive vaccination reduces the susceptible population, and competes with the emergence of new variants. By the first Omicron wave, circa seventy percent were doubly vaccinated, with the target then standing at circa eighty percent. Combined with the distribution of times-to-death, incidence rates from onset of symptoms predict the daily profile of COVID-associated deaths and estimated case-fatality ratio. Cases are under-reported in the first wave and reflect age heterogeneity in fatalities at the second wave. In periods of low incidence, COVID mortality was one percent or less of detected infection.

We construct the dynamic models governing two nonreciprocally coupled fields for cases with zero, one, and two conservation laws. Starting from two microscopic nonreciprocally coupled Ising models, and using the mean-field approximation, we obtain closed-form evolution equations for the spatially resolved magnetization in each lattice. For single spin-flip dynamics, the macroscopic equations in the thermodynamic limit are closely related to the nonreciprocal Allen-Cahn equations, i.e. conservation laws are absent. Likewise, for spin-exchange dynamics within each lattice, the thermodynamic limit yields equations similar to the nonreciprocal Cahn-Hilliard model, i.e. with two conservation laws. In the case of spin-exchange dynamics within and between the two lattices, we obtain two nonreciprocally coupled equations that add up to one conservation law. For each of these cases, we systematically map out the linear instabilities that can arise. Our results provide a microscopic foundation for a broad class of nonreciprocal field theories, establishing a direct link between non-equilibrium statistical mechanics and macroscopic continuum descriptions.

The structural dynamics of biological macromolecules, such as proteins, DNA/RNA, or their complexes, are strongly influenced by protonation changes of their typically many titratable groups, which explains their pH sensitivity. In turn, conformational and environmental changes in the biomolecule affect the protonation state of these groups. With a few exceptions, conventional force field-based molecular dynamics (MD) simulations do not account for these effects, nor do they allow for coupling to a pH buffer. The $\lambda$-dynamics method implements this coupling and thus allows for MD simulations at constant pH. It uses separate Hamiltonians for the protonated and deprotonated states of each titratable group, with a $\lambda$ variable that continuously interpolates between them. However, rigorous implementations of Hamiltonian Interpolation (HI) $\lambda$-dynamics are prohibitively slow when used with Particle Mesh Ewald (PME). To circumvent this problem, it has been proposed to interpolate the charges instead of the Hamiltonians (QI). Here, we propose a rigorous yet efficient Multipole-Accelerated Hamiltonian Interpolation (MAHI) method to perform $\lambda$-dynamics in GROMACS. Starting from a charge-scaled Hamiltonian, precomputed with the Fast Multipole Method (FMM) or with PME, the correct HI forces are calculated with negligible computational overhead. We compare HI with QI and show that HI leads to more frequent transitions between protonation states, resulting in better sampling and accuracy. Our performance benchmarks show that introducing, e.g., 512 titratable sites to a one million atom MD system increases runtime by less than 20% compared to a regular FMM-based simulation. We have integrated the scheme into our GPU-FMM code for the simulation software GROMACS, allowing an easy and effortless transition from standard force field simulations to constant pH simulations.

Advancing quantum information, communication and sensing relies on the generation and control of quantum correlations in complementary degrees of freedom. Here, we demonstrate the preparation of electron-photon pair states using the phase-matched interaction of free electrons with the evanescent vacuum field of a photonic-chip-based optical microresonator. Spontaneous inelastic scattering produces intracavity photons coincident with energy-shifted electrons. Harnessing these pairs for correlation-enhanced imaging, we achieve a two-orders of magnitude contrast improvement in cavity-mode mapping by coincidence-gated electron spectroscopy. This parametric pair-state preparation will underpin the future development of free-electron quantum optics, providing a pathway to quantum-enhanced imaging, electron-photon entanglement, and heralded single-electron and Fock-state photon sources.

Nanoelectromechanical resonators provide an ideal platform for investigating the interplay between electron transport and nonlinear mechanical motion. Externally driven suspended carbon nanotubes, containing an electrostatically defined quantum dot are especially promising. These devices possess two main sources of nonlinearity: the electromechanical coupling and the intrinsic contributions of the resonator that induce a Duffing-like nonlinear behavior. In this work, we observe the interplay between the two sources across different driving regimes. The main nonlinear feature we observe is the emergence of arch-like resonances in the electronic transport when the resonator is strongly driven. We show that our model is in good agreement with our experimental electron transport measurements on a suspended carbon nanotube. This characterization paves the way for the exploration of nonlinear phenomena using mesoscopic electromechanical resonators.

Experiments, in particular on biological systems, typically probe lower-dimensional observables which are projections of high-dimensional dynamics. In order to infer consistent models capturing the relevant dynamics of the system, it is important to detect and account for the memory in the dynamics. We develop a method to infer the presence of hidden states and transition pathways based on observable transition probabilities conditioned on history sequences for projected (i.e. observed) dynamics of Markov processes. Histograms conditioned on histories reveal information on the transition probabilities of hidden paths locally between any specific pair of observed states. The convergence rate of these histograms towards a stationary distribution provides a local quantification of the duration of memory, which reflects how distinct microscopic paths projecting onto the same observed transition decorrelate in path space. This provides insight about the hidden topology of microscopic paths in a holography-like fashion. The method can be used to test for the local Markov property of observables. The information extracted is also helpful in inferring relevant hidden transitions which are not captured by a Markov-state model.

Light-matter interactions are of fundamental scientific and technological interest. Ultrafast electron microscopy and diffraction with combined femtosecond-nanometer resolution elucidate the laser-induced dynamics in structurally heterogeneous systems. These measurements, however, remain challenging due to the brightness limitation of pulsed electron sources, leading to an experimental trade-off between resolution and contrast. Larger signals can most directly be obtained by higher repetition rates, which, however, are typically limited to a few kHz by the thermal relaxation of thin material films. Here, we combine nanometric electron-beam probing with sample support structures tailored to facilitate rapid specimen cooling. Optical cycling of a charge-density wave transformation enables quantifying the mean temperature increase induced by pulsed laser illumination. Varying the excitation fluence and repetition rate, we gauge the impact of excitation confinement and efficient dissipation on the heat diffusion in different sample designs. In particular, a thermally optimized support can dissipate average laser intensities of up to 200 $\mu W/\mu m^2$ within a few nanoseconds, allowing for reversible driving and probing of the CDW transition at a repetition rate of 2 MHz. Sample designs tailored to ultrafast measurement schemes will thus extend the capabilities of electron diffraction and microscopy, enabling high-resolution investigations of structural dynamics.

The discovery of the kagome metal CsV$_3$Sb$_5$ sparked broad interest, due to the coexistence of a charge density wave (CDW) phase and possible unconventional superconductivity in the material. In this study, we use low-energy electron diffraction (LEED) with a $\mu$m-sized electron beam to explore the periodic lattice distortion at the antimony-terminated surface in the CDW phase. We recorded high-quality backscattering diffraction patterns in ultrahigh vacuum from multiple cleaved samples. Unexpectedly, we did not find superstructure reflexes at intensity levels predicted from dynamical LEED calculations for the reported $2 \times 2 \times 2$ bulk structure. Our results suggest that in CsV$_3$Sb$_5$ the periodic lattice distortion accompanying the CDW is less pronounced at Sb-terminated surfaces than in the bulk.

Adaptation refers to the ability to recover and maintain ``normal'' function upon perturbations of internal of external conditions and is essential for sustaining life. Biological adaptation mechanisms are dissipative, i.e. they require a supply of energy such as the coupling to the hydrolysis of ATP. Via evolution the underlying biochemical machinery of living organisms evolved into highly optimized states. However, in the case of adaptation processes two quantities are optimized simultaneously, the adaptation speed or accuracy and the thermodynamic cost. In such cases one typically faces a trade-off, where improving one quantity implies worsening the other. The solution is no longer unique but rather a Pareto set -- the set of all physically attainable protocols along which no quantity can be improved without worsening another. Here we investigate Pareto fronts in adaptation-dissipation trade-offs for a cellular thermostat and a minimal ATP-driven receptor-ligand reaction network. We find convex sections of Pareto fronts to be interrupted by concave regions, implying the coexistence of distinct optimization mechanisms. We discuss the implications of such ``compromise-optimal'' solutions and argue that they may endow biological systems with a superior flexibility to evolve, resist, and adapt to different environments.

We investigate path-wise observables in experiments on driven colloids in a periodic light field to dissect selected intricate transport features, kinetics, and transition-path time statistics out of thermodynamic equilibrium. These observables directly reflect the properties of individual paths in contrast to the properties of an ensemble of particles, such as radial distribution functions or mean-squared displacements. In particular, we present two distinct albeit equivalent formulations of the underlying stochastic equation of motion, highlight their respective practical relevance, and show how to interchange between them. We discuss conceptually different notions of local velocities and interrogate one- and two-sided first-passage and transition-path time statistics in and out of equilibrium. Our results reiterate how path-wise observables may be employed to systematically assess the quality of experimental data and demonstrate that, given sufficient control and sampling, one may quantitatively verify subtle theoretical predictions.

It has been proposed that an observed inverse power-law dependence of the Markovian estimate for the steady-state dissipation rate on the coarse-graining scale in self-similar networks reflects a scale-dependent energy dissipation. By explicit examples, it is demonstrated here that there are in general no relations between such an apparent power-law dependence and the actual dissipation on different length scales. We construct fractal networks with a single dissipative scale and networks with a true inverse energy-dissipation cascade, and show that they display the same scaling behavior. Moreover, we show that a self-similar network structure does not imply an inverse power-law scaling but may be mistaken for one in practice. When no dissipative cycles become hidden by the coarse graining, any scale dependence of the dissipation estimate vanishes if the memory is correctly accounted for in the time-reversal operation. A $k$-th order estimator is derived and necessary and sufficient conditions are proved for a guaranteed lower bound on dissipation. These higher-order estimators saturated in the order are proved to provide sharper lower bounds on dissipation and their scale dependence signifies hidden dissipative cycles. It is shown that estimators not saturated in the order may erroneously overestimate the microscopic dissipation. Our results underscore the still underappreciated importance of correctly accounting for memory in analyzing coarse observations.

A broken time-reversal symmetry, i.e. broken detailed balance, is central to non-equilibrium physics and is a prerequisite for life. However, it turns out to be quite challenging to unambiguously define and quantify time-reversal symmetry (and violations thereof) in practice, that is, from observations. Measurements on complex systems have a finite resolution and generally probe low-dimensional projections of the underlying dynamics, which are well known to introduce memory. In situations where many microscopic states become "lumped" onto the same observable "state" or when introducing "reaction coordinates" to reduce the dimensionality of data, signatures of a broken time-reversal symmetry in the microscopic dynamics become distorted or masked. In this perspective we highlight why in defining and discussing time-reversal symmetry, and quantifying its violations, the precise underlying assumptions on the microscopic dynamics, the coarse graining, and further reductions, are not a technical detail. These assumptions decide whether the conclusions that are drawn are physically sound or inconsistent. We summarize recent findings in the field and reflect upon key challenges.

The short de Broglie wavelength and strong interaction empower free electrons to probe scattering and excitations in materials and resolve the structure of biomolecules. Recent advances in using nanophotonic structures to mediate bilinear electron-photon interaction have brought novel optical manipulation schemes to electron beams, enabling high space-time-energy resolution electron microscopy, quantum-coherent optical modulation, attosecond metrology and pulse generation, transverse electron wavefront shaping, dielectric laser acceleration, and electron-photon pair generation. However, photonic nanostructures also exhibit nonlinearities, which have to date not been exploited for electron-photon interactions. Here, we report the interaction of electrons with spontaneously generated Kerr nonlinear optical states inside a continuous-wave driven photonic chip-based microresonator. Optical parametric processes give rise to spatiotemporal pattern formation, or dissipative structures, corresponding to coherent or incoherent optical frequency combs. By coupling such microcombs in situ to electron beams, we demonstrate that different dissipative structures induce distinct fingerprints in the electron spectra and Ramsey-type interference patterns. In particular, using spontaneously formed femtosecond temporal solitons, we achieve ultrafast temporal gating of the electron beam without the necessity of a pulsed laser source or a pulsed electron source. Our work elucidates the interaction of free electrons with a variety of nonlinear dissipative states, demonstrates the ability to access solitons inside an electron microscope, and extends the use of microcombs to unexplored territories, with ramifications in novel ultrafast electron microscopy, light-matter interactions driven by on-chip temporal solitons, and ultra-high spatiotemporal resolution sampling of nonlinear optical dynamics and devices.

We derive general bounds on the probability that the empirical first-passage time $\overline{\tau}_n\equiv \sum_{i=1}^n\tau_i/n$ of a reversible ergodic Markov process inferred from a sample of $n$ independent realizations deviates from the true mean first-passage time by more than any given amount in either direction. We construct non-asymptotic confidence intervals that hold in the elusive small-sample regime and thus fill the gap between asymptotic methods and the Bayesian approach that is known to be sensitive to prior belief and tends to underestimate uncertainty in the small-sample setting. We prove sharp bounds on extreme first-passage times that control uncertainty even in cases where the mean alone does not sufficiently characterize the statistics. Our concentration-of-measure-based results allow for model-free error control and reliable error estimation in kinetic inference, and are thus important for the analysis of experimental and simulation data in the presence of limited sampling.