Searcharxiv⌕ Search

arXiv subjects

Sven Krippendorf

Publications and source records attributed to Sven Krippendorf.

At least 19 recordsLinked to original sources

Towards a foundation model for Type IIB flux vacua

Progress in characterising the Type IIB flux landscape requires scalable methods for constructing flux vacua. In this work, we investigate generative models as a framework for sampling vacua across multiple Calabi-Yau geometries and for directing searches towards specified physical properties. We compare transformers and diffusion models trained to generate quantised fluxes together with approximate values of the complex structure moduli and axio-dilaton. We exploit a staged training procedure comprising pre-training, and post-training via reinforcement learning, and assess how each stage improves the generated distributions and enables targeted sampling. We further examine whether structures learned from geometries with smaller $h^{1,2}$ can be transferred to geometries with larger $h^{1,2}$ through fine-tuning on specialised datasets. Our results provide evidence that generative models can learn distributions of flux vacua across distinct compactifications, improve searches for vacua with selected properties and transfer information between geometries. This establishes a concrete basis for developing a foundation model of the Type IIB flux landscape.

hep-th↗

Faithful Uncertainties in ML-Based Inverse Problems: GPs, NNs and NTKs for the Triplet PDF

Machine-learning methods are increasingly used to solve inverse problems in physics, where a faithful estimate of the uncertainty on the reconstructed function is as important as its central value. We address this question in the simplest instance of a parton distribution function (PDF) determination: the reconstruction of the non-singlet triplet distribution $T_3(x,Q_0^2)$ from deep-inelastic scattering data. Within a closure-test framework, in which the underlying PDF is known and the pseudo-data carry correlated experimental uncertainties, we compare Gaussian processes with different kernels, neural-network ensembles trained with different loss functions, and empirical neural tangent kernel approximations. We find that methods yielding comparable central reconstructions can produce markedly different uncertainties, and we identify the ingredients that control them: a likelihood that includes the full experimental covariance, the functional prior encoded in the kernel or in the endpoint preprocessing, and, for tangent-kernel approximations, the stage of training at which the kernel is evaluated. A normalised-bias diagnostic shows that calibration depends on the correlations of the predictive covariance, not only on its pointwise width. Our results illustrate why closure tests and covariance-aware diagnostics are essential to validate uncertainties in ML-based approaches to inverse problems.

hep-ph↗

ENFR: Evidence Networks with Flow-Matched References for Amortized Simulation-Based Model Inference

The evidence, $\mathcal{Z} = p(x | \mathcal{M})$, of a data-generating model $\mathcal{M}$ is a powerful tool in Bayesian frameworks: it enables model inference, model misspecification tests, and goodness-of-fit assessment as an alternative to standard discrepancy statistics, e.g. $χ^2$. While versatile, the evidence is notoriously difficult to compute, even when the likelihood is available. In both likelihood-based and simulation-based inference, estimating $\mathcal{Z}$ provides an immediate solution to high-dimensional goodness-of-fit or model misspecification tests: using $\mathcal{Z}(x)$ as the test statistic induces a minimum-volume level-$α$ test. Such tests are essential to assess model fidelity, which is necessary to trust simulation-based parameter inference. Existing learned approaches largely target Bayes factors, which only measure evidence ratios. Instead, we provide accurate, fast, and amortized estimates of the absolute evidence. We present ENFR: Evidence Networks with Flow-matched References. Building on the Evidence Network framework of Jeffrey & Wandelt (2024), the proposed method first learns Bayes factors with respect to a flow-matched reference, from which the absolute evidence is recovered. We show improvement in estimation of $\mathcal{Z}$ over the flow-based baseline in all applications considered and the best performance of all density estimation methods considered in our benchmarks. We furthermore introduce and provide the proof for an architecture that jointly learns the individual evidences of a set of models $\{\mathcal{M}_i\}$. We demonstrate the performance and scaling behavior of these approaches on data from a variety of scientific applications.

stat.ML↗

The Future of AI-assisted Physics: Discovering, Verifying and Organizing Results via Formalization

Rapid advances in AI have opened a new era for physics, enabling unprecedented capabilities in research and discovery. However, these advances also bring challenges in verifying, trusting, and reusing the results they so rapidly produce. Formalization is a necessary part of the solution to these challenges. In this paper, as well as giving an overview of the role of AI and formalization in physics, we make the case that: (1) formalization is an essential layer for AI-driven discoveries in both computational and theoretical physics; (2) AI is lowering the barrier to, and increasing the efficiency of, formalization in physics, though it remains weaker at choosing definitions and designing libraries than at writing proofs; (3) the speed AI brings to formalization in physics currently comes at the cost of trust, credit, and community; (4) shared, community-reviewed formal libraries such as Physlib should be central in making progress in this field, both as a trusted, reusable foundation for physics results and as the environment in which AI systems can reason about physics; (5) formalizing physics raises challenges beyond those in mathematics, most importantly that a checked proof does not establish that its definitions faithfully capture the intended physics. We also give our perspectives on near-term and long-term directions in these areas.

hep-ph↗

Spectral Reach: Understanding Neural Scaling as Progress into the Spectral Tail

Neural scaling laws describe predictable power-law relationships between model size, dataset size, compute, and performance. While these laws guide the development of modern foundation models, the mechanisms underpinning them remain poorly understood, in part due to the absence of scalable analysis tools. To close this gap, we introduce "spectral position": a scalable measure of which eigenvalues of the empirical neural tangent kernel (eNTK) currently drive loss reduction. Applying this measure to scaling experiments, we find that spectral position decreases throughout training: learning shifts from dominant eigenmodes into the spectral tail. Larger models reach further into the tail than smaller models, revealing a size-dependent capacity we call "spectral reach". This suggests why larger models achieve lower losses: they sustain learning on weak spectral signals inaccessible to smaller models. We further identify feature learning as a key enabler of spectral reach. It adaptively amplifies gradient magnitudes as learning advances, sustaining progress where frozen representations stall. This points to concrete interventions through architecture and optimizer design.

cs.LG↗

Scattering Amplitudes as Programs: Self-Evolving Search for Theory and Event Generation

By viewing scattering amplitudes as computer programs, we connect two goals: exposing useful analytic structure and constructing efficient numerical evaluators for collider phenomenology. Equivalent programs can differ sharply in multiplicity scaling, arithmetic complexity, cancellation, and runtime. Amplitude calculation therefore defines a structured search problem over analytic representations, recursive algorithms, colour and helicity organisation, and reuse of intermediate objects. We embed repository-scale coding agents inside an external generate-evaluate-select loop with frozen evaluators and objectives, and study three optimisation targets. A scaling search moves from BCFW recursion to a specialised fixed-k split-helicity transfer algorithm, reaching an 805x geometric-mean speed-up on the scored grid. A structural search reorganises NMHV terms into R-invariant-style cells and glued supercells, reducing inter-term cancellation. Across twenty QCD and electroweak processes, an exact-operation search reduces counted arithmetic by 47.7x and gives a 5.9x post-hoc Python runtime improvement. In matched pure-gluon component tests from n=4 to n=6, the evolved engine's speed advantage over the tested Sherpa-Comix exact-sum call grows from 10x to 277x, while its gap to process-specialised MadGraph5_aMC@NLO Fortran compiled at -O3 narrows from 77.7x to 13.7x. The searches move between mathematical representations and combine recursion, symmetry, basis reduction, dynamic programming, and shared computation into hybrid amplitude programs. They provide initial evidence that parts of amplitude optimisation can be made systematic through self-evolving program search, while leaving native generator integration and end-to-end event throughput as future tests. The generator comparison is an isolated exact matrix-element call, not a modification of MadGraph or Sherpa.

hep-ph↗

Physics as Code: From Scans to Theorems with ITP APIs in $SU(5)$ Model Building

A recurring challenge in theoretical physics is to make reliable global statements about bounded but combinatorially large model spaces. Exhaustive scans quickly become opaque or impractical, while statistical exploration does not by itself provide theorem-backed guarantees. This motivates workflows in which the model-building problem itself is formalized inside an interactive theorem prover (ITP). In this paper we develop an API-based methodology for formalizing such bounded model-building questions inside Lean, an interactive theorem prover. The central step is to represent the relevant charge spectra, predicates, and reduction moves as reusable ITP definitions, and then to derive the classification from proved reduction theorems rather than from an ad hoc scan. We demonstrate the strategy in a concrete $SU(5)$ case study motivated by F-theory model building with additional Abelian symmetries. At the charge-spectrum layer, we classify bounded spectra that admit a top-quark Yukawa coupling, avoid a selected set of dangerous operators, and satisfy a minimal charge-spectrum completeness condition. Our main result shows that every such spectrum in the bounded search space arises from finitely many minimal top-Yukawa witnesses together with controlled completions and certified closure steps. This classification represents a formally verified description of the full viable class in the charge-spectrum setting studied here. The development is implemented inside PhysLib as reusable infrastructure rather than as a one-off verification script. It provides a proof of principle for how interactive theorem provers can turn combinatorially difficult model-building problems into correctness-first, reusable workflows, and we discuss how the resulting certified classification can serve as reliable input for downstream analyses.

hep-th↗

Parameter compression in the flux landscape

We present a data-driven investigation of the exhaustive ensemble of no-scale type IIB flux vacua constructed in \cite{Chauhan:2025rdj}. Using a combination of linear and non-linear dimensionality-reduction techniques, we analyse both flux and moduli spaces and demonstrate that the effective dimensionality of the underlying 12-dimensional flux space is substantially reduced. A central component of our study is a physics-informed autoencoder, which provides a non-linear compression of the flux and moduli data into a low-dimensional latent space. The learned latent representation organises vacua according to desired features and, in particular, isolates distinguished regions associated with small values of the flux superpotential $|W_0|$, revealing non-trivial correlations that are not captured by linear methods. In parallel, we apply tools from topological data analysis, specifically persistent homology, to probe the global structure of the vacuum distribution. This allows us to identify robust, long-lived topological features in both moduli and flux subspaces. This work is a necessary step for developing foundation models in string phenomenology.

hep-th↗

Spinodal Gravitational Waves

We uncover a new gravitational-wave production mechanism in cosmological, first-order, thermal phase transitions. These are usually assumed to proceed via the nucleation of bubbles of the stable phase inside the metastable phase. However, if the nucleation rate is sufficiently suppressed, then the Universe may supercool all the way down the metastable branch and enter the spinodal region. In this case the transition proceeds via the exponential growth of unstable modes and the subsequent formation, merging and relaxation of phase domains. We use holography to follow the real-time evolution of this process in a strongly coupled, four-dimensional gauge theory. The resulting gravitational wave spectrum differs qualitatively from that in transitions mediated by bubble nucleation. We discuss the possibility that the spinodal dynamics may be preceded by a period of thermal inflation.

hep-th↗

Open Source Planning & Control System with Language Agents for Autonomous Scientific Discovery

We present a multi-agent system for automation of scientific research tasks, cmbagent (https://github.com/CMBAgents/cmbagent). The system is formed by about 30 Large Language Model (LLM) agents and implements a Planning & Control strategy to orchestrate the agentic workflow, with no human-in-the-loop at any point. Each agent specializes in a different task (performing retrieval on scientific papers and codebases, writing code, interpreting results, critiquing the output of other agents) and the system is able to execute code locally. We successfully apply cmbagent to carry out a PhD level cosmology task (the measurement of cosmological parameters using supernova data) and evaluate its performance on two benchmark sets, finding superior performance over state-of-the-art LLMs. The source code is available on GitHub, demonstration videos are also available, and the system is deployed on HuggingFace and will be available on the cloud.

cs.AI↗

Beyond Scaling Curves: Internal Dynamics of Neural Networks Through the NTK Lens

Scaling laws offer valuable insights into the relationship between neural network performance and computational cost, yet their underlying mechanisms remain poorly understood. In this work, we empirically analyze how neural networks behave under data and model scaling through the lens of the neural tangent kernel (NTK). This analysis establishes a link between performance scaling and the internal dynamics of neural networks. Our findings of standard vision tasks show that similar performance scaling exponents can occur even though the internal model dynamics show opposite behavior. This demonstrates that performance scaling alone is insufficient for understanding the underlying mechanisms of neural networks. We also address a previously unresolved issue in neural scaling: how convergence to the infinite-width limit affects scaling behavior in finite-width models. To this end, we investigate how feature learning is lost as the model width increases and quantify the transition between kernel-driven and feature-driven scaling regimes. We identify the maximum model width that supports feature learning, which, in our setups, we find to be more than ten times smaller than typical large language model widths.

cs.LG↗

Solving inverse problems of Type IIB flux vacua with conditional generative models

We address the inverse problem in Type IIB flux compactifications of identifying flux vacua with targeted phenomenological properties such as specific superpotential values or tadpole constraints using conditional generative models. These machine learning techniques overcome computational bottlenecks in traditional approaches such as rejection sampling and Markov Chain Monte Carlo (MCMC), which struggle to generate rare, finely-tuned vacua. As a proof of concept, we demonstrate that conditional generative models provide a more efficient alternative, specifically using conditional variational autoencoders (CVAEs). We introduce a CVAE framework tailored to flux compactifications, incorporating physical constraints directly into the loss function - enabling the generation of physically consistent vacua beyond the training set. Our experiments on conifold and symmetric torus background geometries show that the CVAE achieves a speedup of about $O(10^3)$ compared to Metropolis sampling, particularly in narrow target ranges for superpotential values. Additionally, the CVAE generates novel, distinct flux configurations beyond the training data, highlighting its potential for probing computationally challenging regions of the string landscape. Our results establish conditional generative models as a powerful and scalable tool for targeted flux vacua generation, opening new pathways for model building in regions of the landscape previously inaccessible by traditional model building techniques.

hep-th↗

Deep observations of the Type IIB flux landscape

We present deep observations in targeted regions of the string landscape through a combination of analytic and dedicated numerical methods. Specifically, we devise an algorithm designed for the systematic construction of Type IIB flux vacua in finite regions of moduli space. Our algorithm is universally applicable across Calabi-Yau orientifold compactifications and can be used to enumerate flux vacua in a region given sufficient computational efforts. As a concrete example, we apply our methods to a two-modulus Calabi-Yau threefold, demonstrating that systematic enumeration is feasible and revealing intricate structures in vacuum distributions. Our results highlight local deviations from statistical expectations, providing insights into vacuum densities, superpotential distributions, and moduli mass hierarchies. This approach opens pathways for precise, data-driven mappings of the string landscape, complementing analytic studies and advancing the understanding of the distribution of flux vacua. This allows us to obtain different types of solutions with hierarchical suppressions, e.g.~vacua with small values of the Gukov-Vafa-Witten superpotential $|W_0|$. We find an example with $|W_0| = 5.547 \times 10^{-5}$ at large complex structure, without light directions and the use of non-perturbative effects.

hep-th↗

Learning Optimal and Interpretable Summary Statistics of Galaxy Catalogs with SBI

How much cosmological information can we reliably extract from existing and upcoming large-scale structure observations? Many summary statistics fall short in describing the non-Gaussian nature of the late-time Universe in comparison to existing and upcoming measurements. In this article we demonstrate that we can identify optimal summary statistics and that we can link them with existing summary statistics. Using simulation based inference (SBI) with automatic data-compression, we learn summary statistics for galaxy catalogs in the context of cosmological parameter estimation. By construction these summary statistics do not require the ability to write down an explicit likelihood. We demonstrate that they can be used for efficient parameter inference. These summary statistics offer a new avenue for analyzing different simulation models for baryonic physics with respect to their relevance for the resulting cosmological features. The learned summary statistics are low-dimensional, feature the underlying simulation parameters, and are similar across different network architectures. To link our models, we identify the relevant scales associated to our summary statistics (e.g. in the range of modes between $k= 5 - 30 h/\mathrm{Mpc}$) and we are able to match the summary statistics to underlying simulation parameters across various simulation models.

astro-ph.CO↗

PICZL: Image-based Photometric Redshifts for AGN

Computing photo-z for AGN is challenging, primarily due to the interplay of relative emissions associated with the SMBH and its host galaxy. SED fitting methods, effective in pencil-beam surveys, face limitations in all-sky surveys with fewer bands available, lacking the ability to capture the AGN contribution to the SED accurately. This limitation affects the many 10s of millions of AGN clearly singled out and identified by SRG/eROSITA. Our goal is to significantly enhance photometric redshift performance for AGN in all-sky surveys while avoiding the need to merge multiple data sets. Instead, we employ readily available data products from the 10th Data Release of the Imaging Legacy Survey for DESI, covering > 20,000 deg$^{2}$ with deep images and catalog-based photometry in the grizW1-W4 bands. We introduce PICZL, a machine-learning algorithm leveraging an ensemble of CNNs. Utilizing a cross-channel approach, the algorithm integrates distinct SED features from images with those obtained from catalog-level data. Full probability distributions are achieved via the integration of Gaussian mixture models. On a validation sample of 8098 AGN, PICZL achieves a variance $σ_{\textrm{NMAD}}$ of 4.5% with an outlier fraction $η$ of 5.6%, outperforming previous attempts to compute accurate photo-z for AGN using ML. We highlight that the model's performance depends on many variables, predominantly the depth of the data. A thorough evaluation of these dependencies is presented in the paper. Our streamlined methodology maintains consistent performance across the entire survey area when accounting for differing data quality. The same approach can be adopted for future deep photometric surveys such as LSST and Euclid, showcasing its potential for wide-scale realisation. With this paper, we release updated photo-z (including errors) for the XMM-SERVS W-CDF-S, ELAIS-S1 and LSS fields.

astro-ph.GA↗

Collective variables of neural networks: empirical time evolution and scaling laws

This work presents a novel means for understanding learning dynamics and scaling relations in neural networks. We show that certain measures on the spectrum of the empirical neural tangent kernel, specifically entropy and trace, yield insight into the representations learned by a neural network and how these can be improved through architecture scaling. These results are demonstrated first on test cases before being shown on more complex networks, including transformers, auto-encoders, graph neural networks, and reinforcement learning studies. In testing on a wide range of architectures, we highlight the universal nature of training dynamics and further discuss how it can be used to understand the mechanisms behind learning in neural networks. We identify two such dominant mechanisms present throughout machine learning training. The first, information compression, is seen through a reduction in the entropy of the NTK spectrum during training, and occurs predominantly in small neural networks. The second, coined structure formation, is seen through an increasing entropy and thus, the creation of structure in the neural network representations beyond the prior established by the network at initialization. Due to the ubiquity of the latter in deep neural network architectures and its flexibility in the creation of feature-rich representations, we argue that this form of evolution of the network's entropy be considered the onset of a deep learning regime.

cs.LG↗

New non-supersymmetric flux vacua in string theory

In this note we construct large ensembles of supersymmetry breaking solutions arising in the context of flux compactifications of type IIB string theory. This class of solutions was previously proposed in arXiv:hep-th/0402135 for which we provide the first explicit examples in Calabi-Yau orientifold compactifications with discrete fluxes below their respective tadpole constraint. As a proof of concept, we study the degree 18 hypersurface in weighted projective space $\mathbb{CP}_{1,1,1,6,9}$. Furthermore, we look at 10 additional orientifolds with $h^{1,2}=2,3$. We find several flux vacua with hierarchical suppression of the vacuum energy with respect to the gravitino mass. These solutions provide a crucial stepping stone for the construction of explicit de Sitter vacua in string theory. Lastly, we also report the difference in the distribution of $W_0$ between supersymmetric and non-supersymmetric minima.

hep-th↗

W0_sample = np.random.normal(0,1)?

In this note we explore the distribution of vacuum expectation values of the superpotential $W_0$ in explicit Type IIB flux compactifications. We show that the distribution can be approximated, universally across geometries, by a two-dimensional Gaussian with a model dependent standard deviation. We identify this behaviour in 20 Calabi-Yau orientifold compactifications with between two and five complex structure moduli by constructing a total of $\mathcal{O}(10^7)$ flux vacua. We observe a characteristic scaling behaviour of the width $σ$ of our distributions with respect to the D3-charge contributions $N_{\text{flux}}$ from fluxes which can be approximated by $σ\sim \sqrt{N_{\text{flux}}}$. This $W_0$ distribution implies that locating small values of $|W_0|$ as a preferred regime associated with classes of string theory solutions typically featuring hierarchies, simplifies to the basic statement of finding small Euclidean norms of normally distributed values. We do also identify small modifications to this Gaussian behaviour in our samples which might be seen as indications for the breakdown of the continuous flux approximation commonly used in the context of statistical analyses of the flux landscape.

hep-th↗