Multi-Agent Systems for Video Content Moderation at Scale

The Multi-Agent Architecture

Meta manages content integrity at scale by decomposing video analysis into three specialized agents, avoiding the overhead of a single monolithic frontier model. The system operates as follows:

• Perceiver Agent: Acts as the vision expert. It performs temporal decomposition by identifying points of semantic change rather than using fixed frame rates, then compresses similar frames. It emits clip-level embeddings, OCR, and natural language descriptions as a structured JSON object.
• Reviewer Agent: Functions as the central orchestrator. It consumes the perceiver's output, analyzes temporal anomalies (e.g., a sports video suddenly shifting to political content), and integrates real-time user sentiment and engagement signals to make final moderation decisions.
• Retriever Agent: Manages the corpus index. It maps metadata, embeddings, and entities into inverted, vector, and graph databases to facilitate fast similarity lookups during inference.

Model Optimization and Training

Because the system processes billions of frames, it relies on specialized, small-scale Vision Language Models (VLMs) rather than general-purpose frontier models. The development process includes:

• Domain-Specific Pretraining: Vision encoders are pretrained on in-house user-generated content rather than clean web data to better handle adversarial uploads and multilingual screen text.
• Instruction Tuning: Models are tuned to output structured JSON schemas, enabling consistent chain-of-thought reasoning across the agentic pipeline.
• DPO and Human-in-the-Loop: The team uses Direct Preference Optimization (DPO) on production samples. When the system fails, a human queue traces the error to specific tool calls or retrieval steps, creating a continuous feedback loop for retraining.
• Distillation and Quantization: The team performs knowledge distillation and quantization to reduce model size, creating a performance-to-cost table to ensure the pipeline remains scalable and cost-effective.

System Efficiency and Evaluation

To reduce compute load, the pipeline employs three primary pruning strategies:

• Spatial-Temporal Reduction: Compresses redundant frames before processing.
• Verdict Caching: Skips the full pipeline for viral content by reusing previously computed moderation verdicts.
• Metadata Pruning: Uses creator authenticity scores and historical quality metrics to whitelist trusted content, allowing it to bypass the agentic pipeline entirely.

Evaluation is conducted holistically, measuring per-node latency, reasoning budgets, token costs, and judge drift against human-labeled ground truth.