Modality Misalignment and Originality Attribution in Short-Form Video — Aditya Gautam, Meta
Meta uses multi-agent systems and specialized small VLMs to detect modality mismatch and attribute originality in short-form video at scale.
“So, the real data is very chaotic. We're talking about a scale of over 100 million and much more viral content.”
A Meta engineer describes building multi-agent pipelines and specialized small-scale vision-language models to solve modality-mismatch and originality-attribution problems across 100M+ chaotic, adversarial, multilingual short-form videos. The talk covers agent orchestration (LLMs, tools, MCP) and vision/metadata optimizations to selectively route only candidate videos through expensive workflows. It offers practical engineering insight into scaling VLM-based content understanding but is a single-company implementation talk rather than a broad industry signal.