Verified, then logged.
Sun 22 Nov 2026 · 00:20 KST / 15:20 UTC한국어 | ENSearch /

Breaking01 / AI

CineMME reveals gaps in linking dialogue to speakers

Researchers found that visual information helped some video questions but could hinder dialogue-based reasoning.

In a dim security operations room at night, a lone analyst seen from behind studies a wall of monitors showing blurred logs and network graphs among empty desks. AI-generated image

Researchers introduced CineMME, a video understanding benchmark, on 13 Nov, reporting a gap between locating dialogue and linking it to speakers.

According to the paper's abstract, CineMME contains 975 short narrative video clips. Its tasks test when dialogue occurs, what is said and who speaks. It also tests how models use dialogue to interpret the plot.

The abstract reports Gemini-3 Pro's dialogue localisation score as 68.34%. Its speaker-linking score is 7.74%. These measure overlap with the correct time intervals and character regions, respectively.

The researchers found that visual information had different effects across tasks. It improved performance on questions focused on perception. But it could reduce performance on dialogue-based reasoning.

According to the abstract, earlier benchmarks also cover video question answering and audio-visual reasoning. They rarely test whether evidence connects dialogue to plot within the same clips. CineMME is designed to examine that connection.

The researchers did not disclose in the abstract how much visual information reduced performance on dialogue-based reasoning.