Have I Seen Enough? Frozen Video-Language Models Encode Evidence Readiness

Key points

• Streaming video-language models must decide not only what to answer, but whether the evidence needed for the current question has arrived.1
• Existing systems learn that decision as a separate trigger; we ask whether an unmodified model already computes it.1
• Released streaming triggers are also linear readouts, yet a trained trigger read on its own base model's activations is approximately orthogonal to readiness and decodes it far less accurately than a probe.1
• We turn the readout into Readiness Gating, an answer-timing policy that improves accuracy by up to +9.75 pp at matched video duration with negligible computational overhead.1

Sources (1)

• [1]Have I Seen Enough? Frozen Video-Language Models Encode Evidence Readiness
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 6, 03:43 PM

  We show that frozen VideoLLMs carry a linearly readable evidence-readiness signal, labelled from timestamped evidence rather than from model output.

  Streaming video-language models must decide not only what to answer, but whether the evidence needed for the current question has arrived.

Extractive summary: sentences quoted from the sources.