Bidirectional Diffusion Models Self-Assess Rollout Errors Without Ground Truth

Clean-Hovercraft5825 · reddit · 2026-08-06

Autoregressive models (like latent diffusion models) tend to accumulate errors during long rollouts, and deploying them lacks ground truth for measurement.

This research proposes a single conditional latent diffusion model that steps a dynamical system forward or backward in time via a direction flag. This bidirectionality provides a measurement-free test-time error signal: rolling forward and then backward must return the model to its start, making the round-trip discrepancy a self-supervised proxy for the unobservable rollout error.

The method requires no ensembles, held-out data, or governing equations, at the cost of just one extra rollout. Experiments show that training both directions in a single network outperforms using two separate specialist models. Full code for data generation, training, and analysis is open-sourced.

Original post →

More from Research

Research channel →