Searcharxiv⌕ Search

arXiv subjects

Philipp Alexander Kreer

Publications and source records attributed to Philipp Alexander Kreer.

7 recordsLinked to original sources

Bayesian Influence Functions for Hessian-Free Data Attribution

Classical influence functions face significant challenges when applied to deep neural networks, primarily due to non-invertible Hessians and high-dimensional parameter spaces. We propose the local Bayesian influence function (BIF), an extension of classical influence functions that replaces Hessian inversion with loss landscape statistics that can be estimated via stochastic-gradient MCMC sampling. This Hessian-free approach captures higher-order interactions among parameters and scales efficiently to neural networks with billions of parameters. We demonstrate state-of-the-art results on predicting retraining experiments.

cs.LG↗

Noise Injection Reveals Hidden Capabilities of Sandbagging Language Models

Capability evaluations play a crucial role in assessing and regulating frontier AI systems. The effectiveness of these evaluations faces a significant challenge: strategic underperformance, or ``sandbagging'', where models deliberately underperform during evaluation. Sandbagging can manifest either through explicit developer intervention or through unintended model behavior, presenting a fundamental obstacle to accurate capability assessment. We introduce a novel sandbagging detection method based on injecting noise of varying magnitudes into model weights. While non-sandbagging models show predictable performance degradation with increasing noise, we demonstrate that sandbagging models exhibit anomalous performance improvements, likely due to disruption of underperformance mechanisms while core capabilities remain partially intact. Through experiments across various model architectures, sizes, and sandbagging techniques, we establish this distinctive response pattern as a reliable, model-agnostic signal for detecting sandbagging behavior. Importantly, we find noise-injection is capable of eliciting the full performance of Mistral Large 120B in a setting where the model underperforms without being instructed to do so. Our findings provide a practical tool for AI evaluation and oversight, addressing a challenge in ensuring accurate capability assessment of frontier AI systems.

cs.AI↗

One-Loop QCD Corrections to $\bar{u}d \rightarrow t\bar{t}W$ at $\mathcal{O}(\varepsilon^2)$

We present a computation of the one-loop QCD corrections to top-quark pair production in association with a $W$ boson, including terms up to order $\varepsilon^2$ in dimensional regularization. Providing a first glimpse into the complexity of the corresponding two-loop amplitude, this result is a first step towards a description of this process at next-to-next-to-leading order (NNLO) in QCD. We perform a tensor decomposition and express the corresponding form factors in terms of a basis of independent special functions with compact rational coefficients, providing a structured framework for future developments. In addition, we derive an explicit analytic representation of the form factors, valid up to order $\varepsilon^0$, expressed in terms of logarithms and dilogarithms. For the complete set of special functions required, we obtain a semi-numerical solution based on generalized power series expansion.

hep-ph↗

The H-graph with unequal masses in quantum field theory

We compute the family of Feynman integrals related to the H-graph with unequal masses in relativistic quantum field theory. We present an $\varepsilon$-factorised differential equation for the 40 master integrals. The alphabet consists of 29 dlog-forms with algebraic arguments, involving six square roots. As these square roots are not simultaneously rationalizable, we express the master integrals in terms of iterated integrals. In addition, we express the H-graph with unit powers of the propagators up to weight four in terms of multiple polylogarithms in compact form.

hep-ph↗

The H-graph with equal masses in terms of multiple polylogarithms

The initial phase of the inspiral process of a binary system producing gravitational waves can be described by perturbation theory. At the third post-Minkowskian order a two-loop double box graph, known as H-graph contributes. We consider the case where the two objects making up the binary system have equal masses. We express all master integrals related to the equal-mass H-graph up to weight four in terms of multiple polylogarithms. We provide a numerical program which evaluates all master integrals up to weight four in the physical regions with arbitrary precision.

hep-ph↗

One loop QCD corrections to $gg \to t\overline{t}H$ at $\mathcal{O}(ε^2)$

We compute the one-loop corrections to \tth up to order $\mathcal{O}(ε^2)$ in the dimensional regularization parameter. We apply the projector method to compute polarized amplitudes, which generalize massless helicity amplitudes to the massive case. We employ a semi-numerical strategy to evaluate the scattering amplitudes. We express the form factors through scalar integrals analytically, and obtain separately integration by parts reduction identities in compact form. We integrate numerically the corresponding master integrals with an enhanced implementation of the Auxiliary Mass Flow algorithm. Using a numerical fit method, we concatenate the analytic and the numeric results, to obtain fast and reliable evaluation of the scattering amplitude. This approach improves numerical stability and evaluation time. Our results are implemented in the \texttt{Mathematica} package \texttt{TTH}.

hep-ph↗

Feynman integrals for binary systems of black holes

The initial phase of the inspiral process of a binary black-hole system can be described by perturbation theory. At the third post-Minkowskian order a two-loop double box graph, known as H-graph, contributes. In this talk we report how all master integrals of the H-graph with equal masses can be expressed up to weight four in terms of multiple polylogarithms. We also discuss techniques for the unequal mass case. The essential complication (and the focus of the talk) is the occurrence of several square roots.

hep-th↗