Research map / Large language models / Retrieval and knowledge
Toward Faithful Retrieval-Augmented Generation with Sparse Autoencoders
Guangzhi Xiong, Zhenghao He, Bohan Liu, Sanchit Sinha, Aidong Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RAGLens (SAE features + MI selection + GAM) instead of best internal-representation baseline ReDeEP | Llama2-7B · RAGTruth | AUC: 0.7458 -> 0.8413 | Improved |
| use RAGLens instead of ReDeEP | Llama2-13B · Dolly (Accurate Context) | AUC: 0.8420 -> 0.8568 | Improved |
| use pre-activation instead of post-activation SAE features (ablation) | RAGLens · TofuEval | accuracy: 0.5637 -> 0.6865 | Improved |
| use Transcoder instead of SAE as feature extractor (ablation) | RAGLens · RAGTruth | accuracy: 0.7810 -> 0.7778 | No clear effect |
| train detector on RAGTruth, test zero-shot on AggreFact vs CoT prompting | RAGLens · AggreFact | AUROC: 0.5741 -> 0.8019 | Improved |
| train on AggreFact and test on RAGTruth | RAGLens · RAGTruth | AUROC: 0.4842 -> 0.5330 | No clear effect |
| train on Data2txt, test on Summary | RAGLens · RAGTruth Summary | AUROC: 0.4924 -> 0.5386 | Mixed |
| add instance-level feedback from RAGLens for revision | Llama2-7B · RAGTruth mitigation (human) | hallucination rate: 71.11% -> 62.22% | Improved |
| use token-level instead of instance-level feedback (ablation) | Llama2-7B · RAGTruth mitigation (human) | hallucination rate: 62.22% -> 55.56% | Improved |
| use MI-based feature selection instead of random selection (ablation) | RAGLens · RAGTruth, Dolly | detection performance | Improved |
| use GAM predictor instead of LR, MLP or XGBoost (ablation) | RAGLens · multiple datasets | detection performance | Improved |
| use SAE features instead of raw hidden states at small K' (ablation) | RAGLens · AggreFact, TofuEval | detection performance | Improved |
Closest papers by meaning
- ReDeEP: Detecting Hallucination in Retrieval-Augmented Generation via Mechanistic Interpretability (ICLR 2025 · similarity 0.92)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.92)
- Robust Hallucination Detection in LLMs via Adaptive Token Selection (NeurIPS 2025 · similarity 0.91)
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.91)
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.91)
- OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation (CVPR 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
