CS-VAR: Cross-Session Evidence-Aware Retrieval-Augmented Detector for Live Streaming Risk Assessment
Researchers propose CS-VAR (Cross-Session Evidence-Aware Retrieval-Augmented Detector), a system for live streaming risk assessment that combines a lightweight domain-specific model with a Large Language Model (LLM). The LLM reasons over retrieved cross-session behavioral evidence and transfers insights to the small model during training, enabling it to recognize recurring harmful patterns across streams. This design maintains efficiency for real-time deployment. The system addresses challenges in detecting scams and coordinated malicious behaviors that accumulate gradually across seemingly unrelated streams. Extensive offline experiments on large-scale datasets validate the approach.
Key facts
- CS-VAR stands for Cross-Session Evidence-Aware Retrieval-Augmented Detector.
- It uses a lightweight domain-specific model for fast session-level risk inference.
- A Large Language Model (LLM) guides the small model by reasoning over cross-session evidence.
- The system is designed for live streaming risk assessment.
- It targets scams and coordinated malicious behaviors.
- Harmful actions often accumulate gradually across streams.
- The approach maintains efficiency for real-time deployment.
- Extensive offline experiments were conducted on large-scale datasets.
Entities
—