Retrieval-Augmented Generation on Natural Questions and HotpotQA (45% adversarial noise)
88RAGAS FaithfulnessHeavy-CRAG
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Heavy-CRAGEvaluator=gpt-4o-mini2026.03 | 88 | 93.4 | 1,235 | 3 | |
| Tiny-Critic RAGBackbone=Qwen3-1.7B, Fine-tuning=LoRA (r = 16, α = 32)2026.03 | 86 | 91.2 | 492 | 0.06 | |
| Naive RAGRetriever=BGE-M3, Generator=Llama-3-8B-Instruct2026.03 | 44 | — | 450 | 0 |