混合棄權提升 LLM 可靠性

• Architecture: Five parallel detectors combined through hierarchical cascade mechanism for progressive filtering and computational efficiency • Detection Dimensions: Multi-axis risk assessment including safety signals, confidence scores, knowledge boundary detection, contextual signals, and repetition patterns • Inference-Time Operation: Functions as detachable abstention layer operating entirely at inference time without model retraining • Cascade Design: Reduces unnecessary computation by progressively filtering queries, achieving substantial latency improvements over non-cascaded models • Threshold Calibration: Context-aware thresholds dynamically adjust based on real-time signals such as domain and user history • Performance Metrics: Achieves precision >0.95 and recall >0.98 in production settings; reduces false positives by 80% while maintaining high acceptance rates for benign queries • Computational Efficiency: Most queries handled on fast path with only small fraction incurring full cost of deep detection and validation • Generalization: Architecture generalizes across diverse model configurations and domain-specific workloads as demonstrated through expanded benchmark results

This research addresses a critical production deployment challenge for LLMs by decoupling safety mechanisms from model architecture, enabling organizations to retrofit existing systems with adaptive safety layers without retraining. The model-agnostic approach and demonstrated Pareto improvements (simultaneous gains in safety and utility) suggest potential industry-wide adoption patterns, particularly in regulated domains like healthcare and finance where false positives create significant operational costs. The inference-time deployment model positions this as a practical solution for enterprises managing heterogeneous LLM deployments. The emphasis on calibration and context-awareness indicates a broader industry shift toward dynamic, user-aware safety systems rather than static filtering rules. The latency optimization through cascade design addresses a key barrier to safety system adoption in latency-sensitive applications, potentially enabling safer LLM deployment in real-time interactive systems.