Why does an autoregressive video model get worse the longer it keeps generating, and what are the names and visible symptoms of the failure?
Corrected 2026-10-24 - see C3-B1094-02; the original statements below are kept for history and are superseded where the correction says so. The primary papers agree on one mechanism: the model is trained on clean or short context but, when it keeps generating, it conditions on its own slightly wrong frames, so small errors compound (called error accumulation or exposure bias). Reported symptoms differ by method: brightness drift (over-exposure or darkening), motion that freezes, colour shift, flicker when a window shifts, and (separately) content that diverges because nothing remembers what left the frame.
The primary papers agree on one mechanism: the model is trained on clean or short context but, when it keeps generating, it conditions on its own slightly wrong frames, so small errors compound (called error accumulation or exposure bias). Reported symptoms differ by method: brightness drift (over-exposure or darkening), motion that freezes, colour shift, flicker when a window shifts, and (separately) content that diverges because nothing remembers what left the frame.
Curator correction 2026-10-24 (discovery audit record C3-B1094-02, curator-reviewed C3R-0699; excerpts re-located by independent curator GET 2026-10-24): Correction. Record C3-B092-W19-01 lists 'flicker when a window shifts' among symptoms reported by the primary papers (summary, and method item (e) 'flicker at window shifts'), but none of its eight evidence rows supports it. The only flicker statement found in the cited Rolling Forcing HTML (arXiv 2509.25161v1) is in Appendix D, an ablation of where to place RoPE indices for global context frames: option 3 (global frames overlapping the temporal context) introduces flickering. That is a result of a design choice the authors rejected, not a symptom they report for baseline long rollouts. For baselines, Section 4.2 names colour shifts, artifacts and unnatural motion. The record's other symptoms (over-exposure, darkening, motion collapse, colour shift) remain supported by its evidence. It was not accurate when written to present flicker as a paper-reported long-rollout symptom of error accumulation.
Curator qualifications: All excerpts located; error accumulation names and visible symptoms.
Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.
Structured fields
{
"citations": [
{
"url": "https://arxiv.org/abs/2506.08009v2",
"note": "candidate evidence excerpts"
},
{
"url": "https://arxiv.org/abs/2407.01392v4",
"note": "candidate evidence excerpts"
},
{
"url": "https://arxiv.org/abs/2510.02283v1",
"note": "candidate evidence excerpts"
},
{
"url": "https://arxiv.org/abs/2509.25161v1",
"note": "candidate evidence excerpts"
}
]
}Evidence claims
- [sourced_fact] Curator correction 2026-10-24 (discovery audit record C3-B1094-02, curator-reviewed C3R-0699; excerpts re-located by independent curator GET 2026-10-24): Correction. Record C3-B092-W19-01 lists 'flicker when a window shifts' among symptoms reported by the primary papers (summary, and method item (e) 'flicker at window shifts'), but none of its eight evidence rows supports it. The only flicker statement found in the cited Rolling Forcing HTML (arXiv 2509.25161v1) is in Appendix D, an ablation of where to place RoPE indices for global context frames: option 3 (global frames overlapping the temporal context) introduces flickering. That is a result of a design choice the authors rejected, not a symptom they report for baseline long rollouts. For baselines, Section 4.2 names colour shifts, artifacts and unnatural motion. The record's other symptoms (over-exposure, darkening, motion collapse, colour shift) remain supported by its evidence. It was not accurate when written to present flicker as a paper-reported long-rollout symptom of error accumulation. · locator: curator correction paragraph in this record body (C3-B1094-02) source
- [sourced_fact] SUPERSEDED IN PART 2026-10-24 by curator correction C3-B1094-02 (see the correction claim above); original kept for history: The primary papers agree on one mechanism: the model is trained on clean or short context but, when it keeps generating, it conditions on its own slightly wrong frames, so small errors compound (called error accumulation or exposure bias). Reported symptoms differ by method: brightness drift (over-exposure or darkening), motion that freezes, colour shift, flicker when a window shifts, and (separately) content that diverges because nothing remembers what left the frame. · locator: candidate evidence excerpts source
Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | unknown | permitted | permitted | unknown | unknown | Applies to this original summary only, not to the cited sources. |