A streaming video memory test that keeps only questions needing evidence from earlier in the stream. Online models answer under 40% of them.
Editorial summary · not the author abstract. Read the linked paper for the complete abstract and claims.Figure 1: models fail even with the evidence available, memory necessity is checked against recent frames and text alone, and an evidence-controlled protocol builds the benchmark. Original figure · paper authors ↗
Release evidence
Evaluation adapters and scoring for about nine methods are public, with annotations in the repository and the videos (about 310 GB) on Hugging Face.