Renwei Meng
Back

Normal Fréchet Drift: Geometric Diagnostics for Post-Training

Separating normal departure from redistribution along a pretrained reference manifold, with a sample-based estimator and error analysis. Under review at ICLR 2027.

Normal Fréchet Drift: Geometric Diagnostics for Post-Training
Jul–Oct 2026ICLR 2027 · Under reviewCo-first author
Post-trainingGeometryGenerative Models

Research overview

Separating normal departure from redistribution along a pretrained reference manifold, with a sample-based estimator and error analysis. Under review at ICLR 2027.

Conducted with Prof. Tao Lin at Westlake University, July–October 2026. The work distinguishes different forms of distributional change rather than collapsing every post-training change into one score.