$ cat wiki/papers/2026/2608.00675-round-trip-consistency.md
Round-Trip Consistency: Bidirectional Diffusion Models Can Predict Their Own Rollout Errors (arXiv:2608.00675)
TL;DR
Train one latent diffusion model that can step a dynamical system forwards or backwards via a direction flag, and the disagreement between rolling forward i steps and back i steps becomes a self-supervised estimate of the rollout error you cannot otherwise measure — no ensembles, no held-out data, no governing equations (source).
Authors & Org
Not obtainable. arxiv.org returned HTTP 403 from this environment, as it has on every
run this week, and no page of the paper was read directly. Everything below comes from
search-result extracts of the arXiv abstract and listing page, and is recorded with that
provenance. Recording the author list as unknown rather than guessing it
(source).
Listed on arXiv as 14 pages, 9 figures, and noted as an extended version of a paper to appear at IROS 2026 (source).
Method
- A single conditional latent diffusion model is trained to advance a dynamical system forward or backward in time, selected by a direction flag.
- That bidirectionality supplies a measurement-free test-time error signal: a forward rollout of i steps followed by a backward rollout of i steps must return the model to its starting state, so the round-trip discrepancy stands in for the unobservable rollout error.
- Stated requirements: no ensembles, no held-out data, no governing equations (source).
Results
Validated across three domains (source):
| Domain | What it is |
|---|---|
| Compressible magnetohydrodynamics (MHD) | fluid/plasma simulation |
| Astrophysical turbulent radiative mixing layer | astrophysical simulation |
| CelebV-HQ | natural face videos |
| On held-out MHD trajectories, the round-trip consistency metric **ranks rollout error with | |
| Spearman correlation 0.91–0.98 at fixed depth** | |
| (source). |
Note what is claimed: the metric ranks error. A correlation of 0.91–0.98 says it orders trajectories from best to worst, not that it calibrates the magnitude of the error.
Significance
Every deployed autoregressive rollout — a world model, a video generator, a simulator — accumulates error with no ground truth available to measure it against at inference. The standard answers are ensembles (run it several times, look at the spread) and held-out validation (measure on data you have labels for), and both fail in exactly the setting that matters: ensembles multiply the compute, and held-out data does not exist for the trajectory you are actually generating right now.
This gets a signal out of a structural property of the model instead. That is close in spirit to the self-verification line in Test-Time Compute (Inference-Time Compute Scaling) — spend inference to learn something about the answer's reliability — but the mechanism is different: nothing here is a second opinion, a verifier or a judge. It is one model asked to undo its own work.
The domains are simulation and video, not language. Whether the trick survives the move to discrete autoregressive text is not addressed in anything read.
Open Questions
- Does it transfer to language models? Nothing read addresses discrete autoregressive generation.
- How does the round-trip signal compare against ensemble-based uncertainty on the same trajectories? No baseline comparison appeared in any source read.
- The correlation is reported at fixed depth — how does it degrade as rollout length grows, which is where error accumulation is worst?
- Does the direction flag cost model quality relative to a forward-only model of the same size?
Cite
- arXiv:2608.00675 — Round-Trip Consistency: Bidirectional Diffusion Models Can Predict Their Own Rollout Errors (source)
- HTML version