CautiousGen: Experimental Data — Entropy/JS Divergence Complementarity, FEP Head Evaluation, and LoRA Path B Results
收藏官方服务:
资源简介:
Speak only when certain, remain silent when in doubt. Lightweight inference-time hallucination suppression via entropy gating, JS divergence constraint, and confidence-guided filtering. No retraining required. Single 12GB GPU. Discretion wraps any HuggingFace causal LM with a three-gate hallucination filter. When the model is uncertain, it says so — rather than confidently generating wrong answers.
提供机构:
Zenodo创建时间:
2026-03-30



