Papers

We introduce a systematic framework for constructing holonomic recurrences from meromorphic modular forms. Its central ingredient is a modular analogue of creative telescoping, which turns finite-dimensional quotients of spaces of meromorphic modular forms into explicit recurrence operators. This provides a common explanation for several previously isolated examples: it recovers Apéry's recurrence, Beukers' modular constructions and Zagier's seven sporadic recurrences. The framework also produces a large collection of new examples across different weights and groups. The modular origin of these sequences also confers them with non-trivial arithmetic properties, such as denominator bounds, irrationality statements, and $p$-adic congruences.

Via the Rindler method, we study holographic timelike entanglement entropy (TEE) for a general timelike interval in AdS$_3$/CFT$_2$ and flat$_3$/CCFT$_2$. The TEE is the pseudo entropy of a timelike interval. In the bulk, the imaginary part of the TEE comes from a timelike geodesic, but in the Rindler picture this geodesic has not been identified. After a Rindler transformation to a rotating BTZ black hole, the two spacelike extremal curves lie on the inner and outer horizons, and exchanging $l_u\leftrightarrow \ell l_\phi$ takes one to the other. We find the geometric meaning of the timelike extremal curve: a geodesic stretched between the two horizons. Its length, $i\pi\ell$, is independent of the interval, and it alone produces the imaginary part $i\pi c/6$. We follow the same procedure in flat space holography, where the spacelike curve sits on the cosmological horizon and the timelike curve lies between the cosmological horizon and timelike infinity. Evaluating the holographic TEE for a timelike interval in CCFT$_2$ on the plane, we find that, in contrast with the cylinder, the real and imaginary parts are equal.

We construct a three-parameter family of regular analytic supergravity solutions which are dual to (a) the $\mathcal{N}=2^*$ gauge theory on the four-sphere, and (b) $\mathrm{U}(1)^2$-invariant Janus interfaces in $\mathcal{N}=4$ SYM with or without the $\mathcal{N}=2^*$ mass turned on. To this end we employ a five-dimensional supergravity model coupled to six scalars which arises as a certain consistent truncation of type IIB supergravity on $S^5$. The curved domain wall solutions feature a hidden integral of motion which allows for their analytic treatment. We relate the three parameters of the solution to field theory sources and vevs and show how regularity of the solutions restricts their range. We briefly discuss the holographic evaluation of the one-point function of the field theory Lagrangian for our family of Janus solutions.

We give a new rounding technique for spreading-metric linear programs, called Greedy Region Growing, which adapts the classical technique of region growing to min-max graph layout problems and flow-expander decomposition. Using this technique, we obtain state-of-the-art approximation algorithms for several fundamental problems. (1) An $O(\log n)$-approximation for Cutwidth and Pathwidth, improving upon the previous bound of $O(\log n\exp(O(\sqrt{\log\log n})))$ due to Bansal, Katzelnick, and Schwartz (FOCS 2024). This matches the $\Omega(\log n)$ integrality gap of the standard LP relaxations. (2) An $O(\log n)$-approximation for their directed analogues, Minimum Storage and Minimum Register Sufficiency on directed acyclic graphs, improving the previous bound of $O(\log^{3/2} n)$. (3) A $\phi$-flow-expander decomposition that cuts at most an $O(\phi\log n\log\log n)$ fraction of the edges. This improves the $O(\phi\log n\exp(O(\sqrt{\log\log n})))$ bound of Bansal, Jambulapati, and Saranurak (ICALP 2026), and is within an $O(\log\log n)$ factor of the graph-theoretic lower bound. The vertex version of our algorithm also gives an $O(\log n)$-approximation for Treedepth. In particular, our results answer several open problems of Bansal, Katzelnick, and Schwartz (FOCS 2024). Technically, we also solve the conceptual question raised by their work: our technique shows how to adapt Seymour recursion to min-max objectives, and constructs $O(1)$-weak local partitions in spreading metrics.

Every 3-regular graph of girth at least $g$ has a cut containing at least a fraction $M_g$ of its edges, and the best rigorous bounds at large girth are $0.912\le \lim_{g\to\infty}M_g \le 0.92386$. Lower bounds on $M_g$ follow from the performance of local algorithms: on a graph of girth $g \ge 2p+2$ a depth-$p$ local algorithm sees the same tree around every edge, so its expected cut fraction on that tree bounds $M_g$ from below. We apply this strategy to the Variational Iterative Rotation Algorithm (VIRAL), a classical local algorithm, and evaluate it on the depth-$p$ tree for every even depth up to $p=32$. The resulting bounds exceed those recently obtained from the Quantum Approximate Optimization Algorithm (QAOA) for every $p\ge 4$, by $0.005$ to $0.011$, and they extend the range of girths covered from $g=36$ to $g=66$. Notably, at $g=66$ we obtain $M_{66} \ge 0.9127$ at $99.99999\%$ confidence level, which raises the asymptotic lower bound from $0.912$ to $0.9127$.

A teardrop geodesic on a hyperbolic surface is a closed curve that is geodesic except at exactly one point, called its vertex. One imagines a closed curve with one point held down as it deforms to a geodesic. For each teardrop geodesic, we give an explicit geometric construction of the closed geodesic in its free homotopy class. This construction also yields a relation between the length of the teardrop geodesic, the angle at its vertex, and the length of its geodesic representative. We then apply the construction to a self-intersecting closed geodesic at a chosen self-intersection point. Since, at that point, the geodesic splits into two directed teardrop geodesics, our construction gives a geometric description of Turaev's cobracket in terms of geodesic representatives. Using this description we prove that if two self-intersection points yield a pair of canceling terms in the cobracket, then the angle at these points and the lengths of the small loops determined are the same.

Video reward models are trained and evaluated on human preferences between videos that often differ simultaneously on properties such as prompt alignment, visual fidelity, motion, and generator identity. As a result, aggregate accuracy does not reveal which video properties a reward model can reliably distinguish. We construct preference pairs by controlled interventions that change one property of a generated video while holding the rest fixed. The resulting pairs, by construction, provide preferences for how the videos differ along a specific dimension, without requiring any human annotation. We propose IsoVRBench, a video generation reward modeling benchmark of 8,000 such pairs across eight interventions, and show that state-of-the-art models separate broad quality differences between generators but miss important features such as temporal order. The best reward model reaches only 60.8% on reordered frames against 100% for humans, and rankings on existing benchmarks do not predict this failure. Using the synthetic isolated intervation pairs as training data leads to 1) IsoVRM, a reward model outperforming its counterpart trained with a larger human labeled preference dataset by 3.8 points on VBench-2.0, 2) IsoWan, the video generator post-trained with IsoVRM, outperforming its human-data counterpart in human preferences, and 3) gains that stack when combined with human data. Because intervention pairs are automatically generated rather than collected, the benchmark and training data can be easily extended as generators evolve.

Contrastive vision-language (VL) pretraining in computational pathology has generally relied on curated image-caption datasets sourced from social media, educational videos, and research publications. These corpora are sparse relative to the diversity of tissue morphology and difficult to scale further, as pathologists do not routinely produce descriptive captions in clinical practice. We propose to instead leverage slide-level pathology reports which can be generated during routine diagnostic workflows in a scalable manner. A key challenge in this weakly-supervised setting is that pathology reports describe macroscopic, diagnostic findings while patch-level visual features describe microscopic findings. We introduce Locked Image Multiple Instance PreTraining (LIMIT), which addresses this gap by freezing a pretrained vision encoder and fine-tuning only the text encoder: report embeddings serve as cross-attention queries over bags of patch features, producing report-conditioned slide representations optimized via InfoNCE. Using LIMIT, we establish CONCH-Z, which aligns the text encoder of CONCH v1.5 to 335k clinical reports. Evaluated across 22 tasks spanning patch-level classification, tumor detection, and slide-level subtyping, CONCH-Z establishes state-of-the-art zero-shot performance over VL encoders trained on curated caption datasets, while simultaneously closing the gap with slide-level foundation models trained with substantially more parameters and compute. The code and model is available at https://github.com/mahmoodlab/LIMIT.

Navigating humanoid robots through dynamic, non-convex environments requires tightly coupling onboard perception with footstep planning and dynamic locomotion control. Existing methods frequently discard complex free space through simplified convex obstacle bounds and fail to account for real-time dynamic occupancy predictions. This paper presents a real-time framework for humanoid navigation that directly constructs spatio-temporal safe corridors from LiDAR-based particle representations and couples them with Linear Inverted Pendulum (LIP)-based planning and Whole Body Control (WBC) to enable navigation around and within moving non-convex structures without restrictive geometric assumptions. Simulations on the Digit humanoid yield a 96% success rate in navigating around and inside moving non-convex obstacles in a 24 by 24 m^2 arena. Ablation studies confirm that the proposed iterative polytope generation significantly outperforms single-iteration baselines in narrow corridors, while evaluations demonstrate robustness to clutter, sensor noise, control delays, and force perturbations.

We study whether stock disclosures by recognizable investors generate abnormal market reactions beyond the reactions produced by comparable but non-recognizable managers. We define the Buffett effect as the incremental abnormal return to a disclosed new position that is attributable to the public recognition of the disclosing investor. We construct a recognition index from Wikipedia pageviews and ProQuest news mentions. Using data from 2000 to 2025 we compare market reaction to disclosures of 24 recognizable discretionary funds with disclosures of 24 matched control funds. New-position disclosures by recognizable managers generate a positive disclosure premium over the filing date and the following trading days, and the premium increases with investor recognition. We construct calendar-time portfolios that mechanically copy each group's disclosed purchases and find that copying the most recognized investors earns a significant alpha that persists for months and does not reverse. The evidence is most consistent with an entangled attention-information channel: investor recognition directs scarce investor attention toward selected disclosures, and the information revealed in those disclosures is then capitalized into prices.

Universal, equation-of-state (EOS) insensitive relations among the global properties of compact stars are well established for hadronic neutron stars and for self-bound strange stars. We ask whether analogous relations hold for the quark core of a hybrid star, and whether the core can be constrained from the surface. Describing quark matter with the three-flavor MIT bag model with repulsive vector interactions, matched to the hadronic DD2 EOS through a Maxwell construction, we scan the full $(B^{1/4},G_V)$ plane and obtain 1112 hybrid EOSs with maximum masses above $2\,M_\odot$, broadly compatible with current multimessenger constraints. For every stable configuration we evaluate the mass, radius, moment of inertia, and tidal deformability both at the stellar surface and at the boundary of the quark core. Global quantities follow the relations of hadronic stars—the $I$–Love relation to within $0.5\%$—while core quantities follow those of self-bound strange stars with median deviations of $6$–$9\%$. We explain the latter by expanding the structure of the core in powers of its compactness: to first order, the core mass, baryon mass, moment of inertia, and tidal deformability depend only on the compactness and on the properties of the hadron–quark interface, chiefly the ratio $\xi=p_{\rm tr}/\varepsilon_+$ of transition pressure to quark energy density. The strange-star relations are the $\xi\to0$ member of this family, and the deviations from them measure $\xi$. The fractional core mass and radius of all configurations follow the normalized mass profile of the star, and no global observable allows the core size to be inferred. Evanescent cores dominate the viable parameter space, while sizable cores require low bag constants and strong vector couplings.

Let $(M^n,g_0)$ be a complete Riemannian manifold with $\Ric_{g_0}\ge-(n-1)g_0$ whose local universal covers are uniformly $\delta$-Reifenberg-close to $\R^n$ below a fixed radius $\rho$. We show that for every $\alpha\in(0,1)$ there are $\delta(n,\alpha)>0$ and $T(n,\rho,\alpha)>0$ such that, if $\delta\le\delta(n,\alpha)$, a complete Ricci flow $g(t)$ with $g(0)=g_0$ and $|\Rm_{g(t)}|\le\alpha/t$ exists on $(0,T]$. On closed manifolds this is a uniform lower bound for the lifetime of the Ricci flow; on noncompact manifolds, where the initial curvature may be unbounded, it is an existence theorem, and the flow is unique among complete Ricci flows from $g_0$ with a curvature bound of the form $A/t$. The key step is a finite-buffer comparison of two Ricci flows through the Ricci–harmonic map heat flow, following Lee, which transfers a Type-I curvature bound from a complete selected flow to the incomplete flow lifted to a local cover. Consequently, below a dimensional Reifenberg threshold, collapsing closed manifolds admit singular nilpotent fibrations over their Gromov–Hausdorff limits. The Reifenberg hypothesis cannot be weakened to noncollapsing of the local covers: products of shrinking circles with Nifa's four-manifolds have $\Ric\ge0$ and uniformly noncollapsed local covers, but admit no continuous Gromov–Hausdorff approximations onto their limit.

Photons emitted by incoherent sources exhibit temporal bunching, a fundamental manifestation of Bose-Einstein statistics first observed by Hanbury Brown and Twiss (HBT). While intensity correlation measurements originated in astronomy and are now central across quantum technologies, biophysics, and spectroscopy, observing temporal bunching in broadband thermal light remains limited by detector temporal resolution. Conventional approaches - such as streak cameras or optical upconversion - introduce severe trade-offs in efficiency, bandwidth, or setup versatility. Here, we propose and demonstrate a third route: pre-detection temporal magnification of the optical field prior to measuring its intensity autocorrelation function in a conventional HBT setup. Using an electro-optic time lens with a temporal magnification of 660, we resolve the temporal intensity autocorrelation of broadband thermal light generated via non-degenerate spontaneous parametric down-conversion with a bandwidth of 250 GHz. This enables direct observation of thermal photon bunching on sub-3.5 picosecond timescales without narrow spectral filtering, establishing pre-detection temporal imaging as a general platform for broadband statistical optics, with applications spanning ultrafast quantum optics, biophysical imaging, and potentially stellar intensity interferometry.

Generating images beyond the training resolution of Diffusion Transformers (DiTs) remains challenging, as the number of tokens grows rapidly with resolution, full attention becomes expensive, and outputs often suffer from structural distortions and degraded detail. We identify uniform full attention as the common source of both problems. Attending to all tokens at every step causes quadratic cost and exposes the model to more tokens and relative distances beyond the training range, pushing the model out of distribution. We observe that attention's effective receptive field narrows during denoising, from global in early steps that form structure to local in later steps that refine details. We propose AReA, a training-free attention mechanism whose receptive field evolves from coarse to fine, following the denoising progression. In early steps, AReA applies cross-scale coarse attention over downsampled keys and values, capturing global context, while in later steps it applies local attention within progressively shrinking windows. Without additional parameters or training, experiments show that AReA consistently outperforms state-of-the-art training-free baselines across target resolutions and different models, while also being faster and more efficient.

Most machine-learning interatomic potentials (MLIPs) are trained on gas-phase data, which limits their use for chemistry in solution. Here, we introduce AnySolv, a transferable implicit solvation potential that provides solvent-dependent energy and force corrections to existing gas-phase potentials without retraining them. It is trained on 1.4 million solvation energies and forces computed with density functional theory (DFT) and SMD in 36 solvents, spanning charged, open-shell, and transition-state structures. Because solvents are represented by continuous physical descriptors, AnySolv supports 179 solvents, including those absent from training. On external benchmarks, AnySolv reproduces DFT-SMD solvation energies to 2-4 meV mean absolute error at more than three orders of magnitude lower cost, and it matches the accuracy of DFT-SMD against experimental solvation free energies. This enables screening of the mechanisms of 120 nucleophilic aromatic substitution reactions in six phases as well as molecular dynamics simulations that reproduce the experimentally observed solvent-dependent conformational preferences of alanine dipeptide. By providing transferable solvation corrections to existing MLIPs, AnySolv enables scalable simulations of solvent-dependent chemical reactivity and molecular dynamics.

We adapt to the context of quantum theory the treatment, familiar in relativity theory, of a reference frame as a tetrad. Since a tetrad is a set of four orthonormal vectors at a spacetime point, it encodes a timelike direction, three orthogonal spatial directions, and units; and so it gives a configuration in spacetime (a spatial orientation and state of motion) of an idealized infinitesimal apparatus (“rulers and a clock"). So a natural “first guess" is that the state of a quantum reference frame should be an assignment of complex amplitudes to a set of tetrads. We develop this idea, in terms of bundles over spacetime of bases of the tangent spaces. This paper is a companion to two other papers. It emphasises the main conceptual aspects of our proposal, while the other two papers give physical details. It also explains the main contrast between those two papers. Namely: (i) the first considers a single spacetime geometry, so that a quantum reference frame, on our proposal, gives a superposition of perspectives on that geometry; while (ii) the second considers the spacetime manifold with no geometry given a priori, which will imply that a quantum reference frame, on our proposal, can describe a superposition of perspectives on a superposition of geometries—not “just" a superposition of perspectives on a single geometry. The main mathematical idea for representing this contrast will be choosing between (i) the orthonormal frame bundle, for a given spacetime geometry; (ii) the frame bundle in the usual sense of containing all bases of all spacetime points' tangent spaces.

The pass-all-test binary reward has become a standard recipe for training software engineering agents, but its sparsity eliminates difference of partially correct patches. The test pass-rate is a natural alternative. However, we find that it is not a reliable indicator of patch quality. A functionality being tested can be covered by either one challenging test or one hundred trivial tests, where the majority might be already satisfied pass-to-pass tests. We introduce CodeLift, a reinforcement learning method that dynamically calibrates pass-rate rewards, turning the test suite into an adaptive curriculum. At each iteration, CodeLift groups tests that kill the same patches and patches that pass the same tests, then gives more weight to test groups that fewer patch groups pass. This reduces the impact of majority simple tests and focuses more on what is still challenging. Within each rollout, CodeLift applies the same scoring rule to repository snapshots, assigning credit to useful actions even when later ones introduce regressions. Experiments on SWE benchmarks show that CodeLift consistently improves pass@1/2/4/8 over binary and pass-rate baselines. Analyses show that CodeLift provides dense, stable learning signals and unlocks the inherent curriculum in the test suite.

We present a new method to estimate the spectral resolution of a radio data cube from its noise properties. We assume that the data cube is the result of the convolution of a spectrally-uncorrelated cube with a constant convolution kernel K. We demonstrate that K and the resolution, defined as K's integral, can be estimated thanks to the analytical relation between the autocorrelation of the noise spectra extracted from the cube and the power spectrum of the kernel. We discuss a number of algorithms that make this method robust in the realistic case of a finite number of spectra and channels, real signal and artefacts, and we present the results obtained for a number of datasets from the literature. We further highlight that this method is applicable to any data cube that meets our assumptions independently of the observing frequency or the exact nature of the cube's "third" axis, e.g., it could be used to measure the time resolution of transient-search data cubes. We make this method available as part of the open-source software tool spectre.

Motivated by the observation of superconductivity in twisted bilayer WSe$_2$ over an extended range of displacement fields, reaching from the correlated insulator down to zero field, we analyze the correlated electron phases of this system at hole filling $\nu = -1$ and twist angle $3.65^\circ$ using Hartree–Fock, random phase approximation, and functional renormalization group calculations. We derive a continuum model with gate-screened Coulomb interaction from first principles and compare its instabilities to those of a realistic three-orbital Wannier model with a Hubbard interaction and an artificial single-orbital Hubbard model. In both cases, we find a superconducting instability of mixed $d$-wave singlet and $p$-wave triplet character, arising adjacent to inter-valley-coherent spin-density-wave order in the phase diagram spanned by displacement field and interaction strength. Our additional analysis of the pairing mechanism points to inter-valley coherent spin fluctuations as the pairing glue. The extent of the pairing regime and the associated critical temperatures depend on the model input, and we trace differences to the interaction range entering the projected interaction. Within the continuum model, the gate-screened Coulomb interaction yields pairing down to zero displacement field, bringing theory into agreement with experiment.

Neural surrogates for Partial Differential Equations (PDEs) are usually trained on synthetic data from numerical simulators. As models grow and scaling laws demand larger datasets, the cost of data generation often outweighs the training cost and becomes the bottleneck in the scaling quest. Unlike data in other domains, however, each PDE sample has a tunable cost that depends on its numerical fidelity, with the resulting cost-accuracy trade-off governed by known convergence rates. We exploit this property by formalizing neural surrogate training as a function of both data quantity and fidelity under a fixed data generation compute budget, treating fidelity as a fine-grained discrete spectrum rather than a binary distinction. Building on this, we show that multi-fidelity training enables Pareto-optimal error scaling against data generation compute. By using simulations at lower fidelities as an additional training signal, training on the optimal fidelity mixture matches the accuracy of high-fidelity-only training, with up to $6.3\times$ less data generation compute across our settings. We validate these findings across six PDE problems spanning structured grids and unstructured meshes, two distinct error sources (discretization and iterative-solver error), and four surrogate architectures, totaling over 7,000 training runs.

We present the first high-resolution transmission spectroscopy of the ultrahot Jupiter KELT-18b, probing its escaping atmosphere by analyzing a single transit obtained with the CARMENES spectrograph. Transmission spectroscopy and Doppler tomography were utilized to isolate the planetary signal. To mitigate contamination from the strong telluric OH lines overlapping the region of interest, two distinct data analysis procedures were applied: a standard Molecfit and OH-masking approach as well as a PCA-based method. By combining the two complementary approaches, we successfully detect a He\,I triplet absorption feature of $1.10 \pm 0.22\%$ ($5.1\sigma$). The signal exhibits a velocity shift consistent with zero and is attributed to the exoplanetary atmosphere. An injection test reveals that the PCA-based method significantly suppresses the retrieved signal. Additionally, using the ATES and TPM codes, we modeled the atmospheric photoevaporation and the helium triplet profile for different He/H ratios and stellar XUV fluxes, accounting for instrument and planetary rotational broadening. From our simulations we derive a mass-loss rate and an evaporation efficiency for KELT-18b of $10.47 \leq \log_{10}(\dot{M}) \leq 11.68$ and $0.153 \leq \eta \leq 0.321$, demonstrating that the atmosphere of KELT-18b is undergoing photoevaporation. Furthermore, our simulations show good agreement with the observed transmission spectrum, allowing us to exclude very low He/H ratios at a $3\sigma$ confidence level for XUV fluxes up to approximately twice the nominal reconstructed value. Our result expands the known parameter space of exoplanets with confirmed helium absorption, establishing KELT-18b as the most massive planet orbiting the hottest host star with such a detection.

Interacting fermions owe both their non-trivial correlations and their computational power to their departure from Gaussianity. Yet, quantifying that departure geometrically faces two challenges: the Gaussian manifold is non-convex, and perturbative treatments are confined to weak interactions. Here, we introduce a geometric measure of fermionic non-Gaussianity, the fidelity distance of a target state to the manifold of fermionic Gaussian states. We construct an iteration formalism to numerically access it, applicable also at strong coupling. Further, we prove a sharpened operator Jensen inequality, of interest in its own right. We use it to derive a witness that involves only two- and four-point fermionic correlators. Numerical benchmarks on the one-dimensional Fermi–Hubbard model show that the quality of the witness improves with system size and, for strong coupling, is tighter than antiflatness. Our results open a scalable route to charting non-Gaussianity of strongly-interacting fermionic quantum simulators.

Population III (Pop III) stars link the chemical enrichment of the early Universe to the formation of the first massive black holes, yet their evolution remains poorly understood at extreme masses. Using GENEC, we evolve rotating zero metallicity models from $500$ to $1000\,M_\odot$ at $v/v_{\rm crit}=0.1$ through core Si burning, and a $1000\,M_\odot$, $v/v_{\rm crit}=0.4$ model evolved through He exhaustion. In the fiducial $1000\,M_\odot$ model, convection and rotational transport carry freshly synthesized C and O into H burning layers, producing $62.1\,M_\odot$ of primary N by He exhaustion. Surface mass loss removes $20.2\,M_\odot$, including $1.8\times10^{-3}\,M_\odot$ of N, with $\log(\mathrm{N}/\mathrm{O})=0.68$, $\log(\mathrm{N}/\mathrm{C})=1.48$, and $\log(\mathrm{C}/\mathrm{O})=-0.80$. By the final hydrostatic model, the inner $100\,M_\odot$ is confined within $0.088\,R_\odot$ while the photosphere remains several thousand solar radii away. As C, Ne, O, and Si burning accelerate, chemical mixing freezes out and advanced products remain buried beneath the envelope. At core Si exhaustion, the central angular velocity reaches $4\times10^{-3}\,{\rm s^{-1}}$ while the surface rotates at $5.25\times10^{-8}\,{\rm s^{-1}}$, a contrast of $7.6\times10^4$. Contraction drives this spin up, while inner specific angular momentum changes by less than $1\%$. The maximum $j_{\rm eq}/j_{\rm ISCO}$ is $0.313$, limiting disk formation in our inside out estimate. These models show that Pop~III stars can produce strong primary N signatures and rapidly rotating black hole progenitors while their envelopes determine which products enrich the early Universe.

The decay of dark matter (DM) is a generic prediction of many particle physics models, producing distinctive monochromatic signatures in environments such as the Galactic halo. Searches for decaying DM in the elusive 'MeV gap' have remained conspicuously under-explored, despite well-motivated models including axion-like particles, sterile neutrinos, and hidden-sector scenarios predicting signatures in this energy range. The upcoming Compton Spectrometer and Imager (COSI) telescope will help bridge this gap. Designed to probe this energy band with high spectral resolution, low background, and wide field of view, COSI is well-suited for DM line searches at the MeV scale. In this work, we make projections for COSI's 2-year sensitivity for DM decay in the Milky Way halo using dedicated all-sky simulations incorporating COSI's predicted instrument response. We conclude that COSI will improve the sensitivity to the DM lifetime by more than an order of magnitude over the DM mass range 0.4 to 10 MeV.

Over the last decade, magnetic fields (B-fields) have become a common ingredient in galaxy formation simulations. Although seemingly unimportant in matching observational constraints such as the $M_*-M_{\rm halo}$ relations, the B-field influences the circumgalactic medium (CGM) gas and determines cosmic ray (CR) propagation, making it an indispensable non-thermal component. Nevertheless, the intricate physics of B-fields, particularly how they evolve to the observed $\mu$G-level field, remains largely elusive. In this work, we systematically address the evolution of B-fields over cosmic time using the FIRE-2 (Feedback In Realistic Environments) magnetohydrodynamics (MHD) simulations. We present results for both Milky Way (MW)-mass and dwarf galaxies, with and without CRs. We find that CRs have little impact on B-field properties, while the halo mass controls the saturated B-field strength at $z\sim 0$. However, all runs reach a halo-independent magnetic-to-turbulent ratio of $f_{\rm B}\sim 1-10\%$, though the methods used to measure turbulence could affect this value. We identify a three-stage evolutionary course for the B-field in tandem with galaxy formation: i. expansion, ii. amplification, and iii. saturation. The feature is universal, though the timings depend on the initial location and star formation history of the gas. Flux-freezing is the predominant driver for the first and last stages, exhibiting power relations $B\sim \rho^{2/3}$ and $\rho^{1/2}$, although the latter could also be explained by dynamical and energetic equilibria. The intermediate amplification stage, potentially driven by “dynamos”, is fuelled by structure formation, feedback, and gas turbulence. We conjecture a numerically triggered, turbulence-driven amplification process to explain this phenomenon.

For the two-point function of the massless, minimally-coupled, real scalar field in Euclidean de Sitter space with $\lambda\phi^4$-interaction, we compute the leading-logarithmic terms in the invariant distance of the two-point function to $O(\lambda^2)$, which counts as the fifth subleading order (N$^5$LO) in the $\sqrt{\lambda}$-expansion. We show that the analytic continuation to Lorentzian signature coincides with the expansion of the analogous quantity computed in the stochastic formalism. We determine the large-order behaviour of the stochastic series expansion and prove its Borel summability, which suggests that the Euclidean theory can be used to extract the late-time behaviour of massless correlation functions in Lorentzian de Sitter space.

What information does a reinforcement learning policy omit? At each state, a standard policy stores one probability per action. It has no place to record how actions relate to each other, such as which pairs go together and which conflict. We ask when this missing structure matters, and we give exact answers. We store the policy at each state as a density matrix, a positive semidefinite, unit-trace matrix borrowed from quantum information theory. We use it here as a purely classical data structure, computed on ordinary classical hardware. The diagonal holds the usual action probabilities. The off-diagonal entries, called coherences, are new parameters. This gives density-matrix reinforcement learning (DMRL). Coherences change outcomes in two settings. First, suppose a policy is regularized toward a coherent reference policy using matrix relative entropy. The optimum then has coherences that no diagonal reference policy produces, and its action probabilities can depend on the reference policy's coherences. This is relevant to fine-tuning of large language models (LLMs) with a Kullback–Leibler penalty. Second, suppose the agent chooses a unit-vector action and the reward is a quadratic function of that vector. This could arise when an LLM judge scores a weighted consolidation of reasoning paths, or when coupled robotic sensors each respond to a mix of actions. Then some coherent policies earn strictly more expected reward than every classical policy with the same action probabilities, and we give the margin in closed form. In a two-action example, both policies assign probability $0.5$ to each action, yet the classical one earns expected reward $0$ while the coherent one earns $1$, the maximum. Outside these two settings, DMRL matches or transfers the classical results. This is a theory paper. All results are proved, and we report no empirical evaluation.

Every physical prediction of an event generator is a statement about the hadronic fluxes that can be recorded by particle detectors. By contrast, the formal backbone of modern event-generator precision is provided by the unobservable partonic final states of the parton shower, with hadronisation included as a model. We construct the probability distribution of the hadronic fluxes non-perturbatively, directly from detector operators acting on asymptotic hadronic states, and encode it in an event generator source functional. This distribution contains every measurement a detector can make, and is therefore what a generator must compute with precision, including the effects of hadronisation. Restricting the functional to the energy flow yields the parton shower source functional, which is infrared and collinear safe and contains every calorimetric observable, making it a natural object against which to assess parton-shower accuracy. We compute its logarithmic structure at one loop in $e^+e^-$ annihilation, finding it directly sensitive to kinematic recoil and hence to the recoil scheme of a parton shower. The parton shower functional depends on a continuum of low scales, one for each resolved direction, and in a toy model its logarithms cannot be entirely resummed within a factorisation whose evolution derives only from the high-scale physics. Our operator-based formulation opens the door to the study of event-generator-like systems in theories beyond QCD, including conformal theories such as $N=4$ super-Yang-Mills theory, and thereby offers a new route to non-perturbative insight into event generation.

Our understanding of gas torques on massive black hole binaries (MBHBs) relies on hydrodynamical (HD) simulations of circumbinary disks (CBDs) in which an $\alpha$-viscosity models angular momentum transport. However, CBDs around MBHBs near merger are hot, magnetized plasmas where transport is driven by magnetohydrodynamical (MHD) turbulence sustained by the magnetorotational instability (MRI). We test how well an $\alpha$-viscosity captures MRI turbulence with 3D global Athena++ simulations of a locally isothermal CBD around an equal-mass, circular MBHB, run in HD and MHD from the same initial torus on identical grids, so that only the transport mechanism differs. Excising the region within twice the binary separation, $r<2~a_{\rm bin}$, allows $1000$ orbits of evolution, several times longer than previous MHD simulations. Both disks deliver the same accretion rate, and the gas beyond $2~a_{\rm bin}$ exerts a positive torque in both, yet the MHD torque is $3.5$ times larger and $3$ times more variable. Near the midplane, transport is similar in the disk body but three times weaker in MHD at the cavity edge, where $2$-$3$ times more mass piles up in a more eccentric cavity, raising the torque. The accretion rate is modulated mainly at twice the binary frequency in HD, but by a persistent lump at the cavity edge and its beat with the binary in MHD. The outer-disk torque therefore implies a contribution to the gas-induced phase shift in the gravitational-wave signal of LISA binaries $3.5$ times larger than in $\alpha$-disk models, and a different variability signature for LSST and Roman.

We revisit the scenario of critical Higgs inflation, in which inflation is driven by the Standard Model Higgs boson with parameters chosen close to a critical point, giving rise to an inflection point in its potential. After reviewing this model, we perform a systematic study of its viability centered around three criteria: (1) consistency with electroweak-scale measurements, (2) ability to reproduce the cosmic microwave background (CMB) observables, and (3) sensitivity of the model predictions to unknown UV physics, based on a breakdown of perturbative unitarity. We first show that compliance with electroweak-scale measurements is controlled by the parameters in the Higgs potential at inflationary scales, also updating the literature values. We then demonstrate how for any choice of potential parameters, agreement with the CMB observables can be achieved through the choice of the nonminimal coupling, in both metric and Palatini formulations of gravity. However, we also find that there exist no regions of parameter space, in either the metric or Palatini formulations, in which compliance with all three physical criteria is possible. While this does not exclude the possibility of the Higgs field driving inflation, it means that predictions for any observable are in principle sensitive to the details of the model's UV completion.