Preprint Report: Collider simulation fixes, black-hole structure, and shift-robust learning

Roughly 8,800 arxiv preprints landed this week, and the center of gravity leaned more theoretical than the recent run of AI-heavy recaps. Approximately 19% touched machine learning research, while around 13% sat in gravitational physics research. High-energy physics simulation correction led the narrower themes at roughly 5% of machine learning research. Work on quantum gravity and black-hole structure forms about 8% of gravitational physics research, using reduced models to test which features are really structural. Work on robust learning under distribution shift forms about 5% of machine learning research, with calibration and uncertainty taking priority over clean-setting accuracy.

Correcting collider simulations

A lot of the collider-facing preprints are less interested in inventing new observables than in making existing pipelines fail less often when simulation shortcuts creep in. Learning Minimal-Deviation Corrections for Multi-Dimensional Mismodelling in HEP Simulations tackles the common problem that simulated events match data only along a few marginals, then fixes it with bounded residual corrections that stay close to the nominal generator. Transfer Learning Across Fast- and Full-Simulation Domains in High-Energy Physics goes after the gap between cheap and expensive detector simulation, then uses transfer learning to move models across that divide with less re-training. Uncovering Hidden Systematics in Neural Network Models for High Energy Physics addresses a subtler failure mode, where a network seems accurate but has absorbed detector-specific quirks, then probes those hidden biases directly instead of assuming the validation split is enough.

Black holes without excess machinery

Gravity preprints this week keep narrowing the question until the math becomes inspectable. Scalar memory from compact binary coalescences asks how beyond-GR signals might survive a merger calculation, then isolates a memory effect, a lasting waveform offset, in a form that can be compared to standard expectations. Gravitational Wave Memory in Beyond GR Theories studies a similar pressure point from another angle, treating memory as a clean diagnostic when fully realistic dynamics are hard to control. Matter Maps to Geometry in Gravitational Collapse takes the collapse problem, where intuition often outruns solvable models, and rebuilds it through an explicit matter-to-geometry relation that makes the causal structure easier to track. If you're working near holography or black-hole dynamics, the mood is to simplify first and generalize later.

Calibrating models under shift

Robustness work this week reads like a reaction against settings where a model looks fine until the test distribution moves. DRIVE-C: A Controlled Corruption Dataset for Autonomous Driving starts from the problem that real-world perception fails under weather, blur, or sensor damage, then builds controlled corruptions so failure modes can be measured instead of guessed. Uncertainty Quantification for Cardiac Shape Reconstruction with Deep Signed Distance Functions via MCMC methods addresses the usual point-estimate habit in medical reconstruction, then uses posterior sampling to show where shape predictions are stable and where they are not. Conformal Path Reasoning: Trustworthy Knowledge Graph Question Answering via Path-Level Calibration targets overconfident graph QA, then wraps path-level predictions in conformal calibration, a coverage guarantee method, so confidence scores mean more than model self-belief.