What question did this study set out to answer?

The aim is to experimentally validate the Quantum Geometric Tensor framework as it relates to training algorithms.

May 10, 20260 citationsOpen Access

The Quantum Geometric Tensor Reveals That Training Is a Physical Process. II. Experimental Evidence

Key Points

The aim is to experimentally validate the Quantum Geometric Tensor framework as it relates to training algorithms.
Conducted a three-run experiment on random 3-SAT with matched real-valued and Clifford-weighted architectures.
Trained models under three conditions: physics-blind optimizer, physics-aware optimization, and tap-aware closed-loop control.
Analyzed the emergence of geometric events and topology in the training process.
Standard real-valued models maintained Ω = 0, while Clifford-weighted models activated geometric instruments.
17 confirmed I-bit events were reported across 10 of 18 transformer runs; 3 events in 3 of 18 TCN runs; none in 36 RBM runs.
Identified spectrally structured, Berry-active geometric events with no quantized topological transition on the tested surfaces.

Abstract

This preprint is the experimental companion to “The Quantum Geometric Tensor Reveals That Training Is a Physical Process. I. Framework and Diagnostic.” It tests the Quantum Geometric Tensor / Information-Geometry Balance Principle (QGT/IGBP) framework in a controlled three-run program on random 3-SAT, with the QGT understood here as a geometric object on parameterized state families rather than as a claim about quantum hardware. The experiment trains matched real-valued and Clifford-weighted architectures first under a physics-blind optimizer, then under physics-aware optimization, and finally under tap-aware closed-loop control with internal observability. The baseline result is exact: standard real-valued models preserve the algebraic flatline Ω = 0, while Clifford-weighted models activate the geometric instrument. Geometric events are already observed in Runs 1 and 2 under passive or lightly controlled training, while Run 3 shows that the critical regime can be steered and reveals the mature full-stack hierarchy. Across the full experiment, the paper reports 17 confirmed I-bit events across 10/18 transformer runs, 3 events across 3/18 TCN runs, and 0 events across all 36 RBM runs. A post hoc topology investigation then finds spectrally structured, Berry-active geometric events but no quantized topological transition on the tested surfaces. The experimental conclusion is that training is a measurable geometric process whose event capability depends jointly on encoding, dynamics, architecture, and observability. -Corrected April 2026 to fix bibliography entries, add public companion DOI references, and freeze the visible preprint date. No substantive scientific claims changed.-Correction May 2026 to fix berry curvature sign convention and subsequent experiment results.

AI에게 질문

Bookmark

View Full Paper