New diffusion model by Alexander Scheinker predicts its own errors
Bidirectional diffusion models achieve 91-98% error prediction with round-trip consistency
Alexander Scheinker's new paper introduces a breakthrough approach called *round-trip consistency* for bidirectional diffusion models, enabling them to self-diagnose rollout errors without requiring ground truth data or ensembles. The key insight is that rolling forward *i* steps and then backward *i* steps should return the model to its starting point; discrepancies in this round-trip serve as a self-supervised error signal.
The method was validated on three challenging datasets: compressible magnetohydrodynamics (MHD), turbulent Navier-Stokes flows, and natural face videos (CelebV-HQ). On held-out MHD trajectories, the round-trip consistency metric (denoted as *Cᵢ*) achieved Spearman correlations of 0.91–0.98 in ranking rollout errors, and a simple calibrator predicted error magnitudes within 1.14× (68% confidence) and 1.29× (95% confidence). Notably, the signal flagged out-of-distribution data (e.g., the Orszag-Tang vortex) with near-perfect accuracy (AUROC 0.98) and reduced incurred error by 15% at 80% coverage—a threefold improvement over depth-only baselines. The bidirectional training incurred no additional cost and even improved performance in both forward and backward directions, with the backward pass doubling as a fast inverse solver.
- Round-trip consistency in bidirectional diffusion models predicts rollout errors without ground truth, achieving 91-98% Spearman correlation on MHD datasets
- Error signal reduces incurred error by 15% at 80% coverage and flags out-of-distribution data with 98% AUROC
- Single bidirectional model matches accuracy of a 10-model ensemble in turbulent Navier-Stokes benchmarks at 10% training cost
Why It Matters
Enables trustworthy long-horizon generative modeling by letting models self-diagnose errors in real time, critical for scientific simulations and video generation.