- Who it helps
- Researchers building assistants that watch video as it arrives.
- Availability
- November 4 paper; November 5 training-code release with linked checkpoint and prepared data. The checkpoint card lists 4B BF16 parameters and no hosted inference provider.
Credit each decision
The method scores prefixes with later replies replaced by silence, then separates each turn’s remaining reward from earlier success. It weights waiting and answering turns equally regardless of reply length.
TurnWise trains video assistants to answer or keep watching ↗
Artifacts have specific dependencies
The released training setup targets Qwen2.5-VL with pinned verl and SGLang revisions, Python 3.10 and PyTorch 2.6.0. The paper also evaluates Qwen3-VL; the supplied overrides do not establish support for that backbone.
TurnWise trains video assistants to answer or keep watching ↗TurnWise training code and compatibility notes ↗TurnWise checkpoint card ↗Prepared streaming-video training data ↗
What you can do
- Read the linked primary evidence and its access limits before planning a trial. Open reference ↗
Where the evidence stops
Training, checkpoints and reward judging were not run. Model and upstream dataset reuse terms need review before deployment.
Sources read. Sources read November 5, 2026 UTC. No third-party project, model inference or complete product workflow was executed.
Sources & further reading
Paper · paper
TurnWise trains video assistants to answer or keep watching ↗
Primary page read November 5. Claims were checked against the page; the workflow was not executed.
Shuai Dong and collaborators · page checked 2026-11-05
GitHub · implementation
TurnWise training code and compatibility notes ↗
Primary page read November 5. Claims were checked against the page; the workflow was not executed.
Shuai Dong and collaborators · page checked 2026-11-05
Hugging Face · implementation
Prepared streaming-video training data ↗
Primary page read November 5. Claims were checked against the page; the workflow was not executed.
Shuai Dong · page checked 2026-11-05
Hugging Face · implementation
TurnWise checkpoint card ↗
Primary page read November 5. Claims were checked against the page; the workflow was not executed.
Shuai Dong · page checked 2026-11-05
6 likes · observed 2026-11-05 13:19 UTC
Coverage note: Original YouTube demonstrations and independent cross-platform attention were not established. No end-to-end workflow was executed.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
TurnWise gives each reply or waiting decision its own training signal, aiming to reduce repeated answers during streaming video.