PEARL lets vision-language models register user-defined concepts mid-stream in video
Researchers formalize Personalized Streaming Video Understanding, release the PSVU-Bench benchmark and a training set, and show a plug-and-play baseline that needs no per-concept training.
In plain words
Today's video AI models can only recognize concepts fixed during training and work on finished clips, not live streams. PEARL is a method that lets a user name a new concept while a video is still playing, and the system then tracks it and recalls when it appeared earlier. The authors also released a benchmark and training data to measure this task.
What was announced
In a paper on arXiv, the authors argue that existing personalized vision-language models learn a fixed concept set during training and process static images or offline videos, so they cannot register a concept a user defines partway through a stream or recall earlier moments when it appeared. They formalize this as Personalized Streaming Video Understanding (PSVU) and introduce PSVU-Bench, described as the first benchmark for the task, annotating 132 videos with timestamped questions across Concept-Definition, Real-Time and Past-Time QA.
Their proposed baseline, PEARL, stores user-defined concepts separately from archived video clips and rewrites each query with concept descriptions to retrieve relevant history. The team built PSVU-10K, a training set of verified concept questions from public internet videos, and trained PEARL's retriever and answering model once; concepts defined at test time need no further training. Applied without training to three architectures, the authors report average gains of 15.05% on Real-Time QA and 12.60% on Past-Time QA over corresponding offline models, with all three surpassing existing online video models on both QA types. Additional training on Qwen3-VL-8B reportedly added 8.66% and 9.39% respectively. Code, data and models are on GitHub.
What it does not show
- The evidence is the paper's abstract only; the reported accuracy figures and benchmark claims come from the authors and are not independently confirmed.
For security and IT teams
- Scope
- The source is the paper's abstract, covering the task definition, benchmark composition and headline results, not the full paper's experimental detail.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.