The Low Frequency Trap: Video Language Models Fail at Simple Event Bookkeeping

In 30 Sekunden

Was passiert ist

Researchers found that video language models like Gemini 3.6 Flash fail to count events reliably in videos, especially frequent or numerous events, despite appearing accurate on final answers.

Warum es zählt

Engineers building video understanding systems should care, particularly those relying on models for temporal reasoning or event tracking in production applications.

Achtung

Higher sampling rates and different prompts show modest improvements in raw accuracy but do not fix the underlying problem of faithful event recovery from video sequences.