- Who it helps
- Researchers building video understanding evaluations.
- Availability
- The October 25 paper and a public implementation are available. The repository does not bundle experimental score files or saved predictions.
Question format affects transfer
The paper calls specialization to a training interface “interface lock-in.” Its chronological mixed supervision aims to cover multiple question families; reported gains use matched observations.
Access to frames is a separate variable
The implementation includes configuration and software tests. Its CPU fixture checks do not reproduce the paper’s model scores. A useful comparison keeps frame access fixed before attributing differences to training.
What you can do
- Inspect the training setup and matched-observation evaluation before comparing reported video scores. Open reference ↗
Where the evidence stops
Reported benchmark improvements were not independently reproduced; changed observation access can confound comparisons.
Sources read. Public sources checked October 28, 2026. No installation, model inference or complete product workflow was executed.
Sources & further reading
Paper · paper
ChronoReplay: check what video training transfers ↗
Supports the subject identity, stated behavior and verification limits.
arxiv.org · page checked 2026-10-28
GitHub · implementation
ChronoReplay implementation ↗
Supports the subject identity, stated behavior and verification limits.
Yihao Hu / LINs Lab · page checked 2026-10-28
Coverage note: Original X posts were inaccessible; comparable independent demonstrations, discussion and complete historical coverage were not established.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
ChronoReplay studies whether video reasoning trained for one question format carries over to other spatial and temporal questions.