SALESFORCE AI RESEARCH

Intelligence with a sense of timing.

01 / THE EVIDENCE

Clear improvements over
state-of-the-art baselines.

We report PAUC-F, a proactive video question-answering metric that balances timeliness and precision.

Timely and selective responding

020406080100

PAUC-F ↑ · ω = 0.5 · GPT-4.1 judge · scale 0–100

02 / THE FRAMEWORK

Teach awareness.
Learn restraint.

Paper framework overview: a video stream with Silence, Response and Hold decisions; human-verified annotated data, a sliding visual window with textual memory and structured output, causal model training, and timeliness plus firing precision evaluation.
03 / DETAILED EVALUATIONS

Across benchmarks.

Four Common Models. Two Partial Models. Nine Evaluation Axes.

Radar chart across nine evaluation axes: four complete models plus GPT-5.6-Sol and Gemini-2.5-Pro with four ProactiveVideoQA PAUC-F scores each. Missing scores are left unplotted.