AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

LUMINA: Detecting Hallucinations in RAG System with Context–Knowledge Signals

Samuel Yeh, Sharon Li, Tanwi Mallick

ICLR 2026 · Large language models › Retrieval and knowledge · approach: RAG robustness and evaluation

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 changedWhereMeasuredEffect
use LUMINA instead of ReDeEPReDeEP · HalluRAG, Llama2-13BAUROC: 0.7645 -> 0.9166Improved
use LUMINA instead of ReDeEPReDeEP · RAGTruth, Llama2-13BAUROC: 0.8055 -> 0.8569Improved
use LUMINA instead of ReDeEPReDeEP · RAGTruth, Llama3-8BAUROC: 0.7495 -> 0.7446Hurt
add entropy normalization to layer weighting in information processing rate (ablation)LUMINA · RAGTruth, Mistral-7BAUROC: 0.6111 -> 0.7679Improved
add layer weighting to entropy normalization (ablation)LUMINA · RAGTruth, Llama2-7BAUROC: 0.7164 -> 0.7652Improved
combine external context and internal knowledge scores (ablation)LUMINA · RAGTruth, Llama2-13BAUPRCImproved
use RBF kernel instead of cosine kernel in MMD (ablation)LUMINA external score · RAGTruthAUPRCNo clear effect
vary top-k tokens for MMD approximation (50, 100, 500) (ablation)LUMINA · RAGTruthAUROC: <0.02% differenceNo clear effect
vary random documents used for d' (ablation)LUMINA · RAGTruthAUROC: std <0.0025No clear effect
set lambda to 0.5 instead of tuned 0.2 (ablation)LUMINA · RAGTruth, Llama2-7B and Llama3-8BAUPRC: drop <0.025No clear effect
inject noise in retrieved and random documents (ablation)LUMINA · RAGTruth, Llama2-13BAUPRC: >0.1 dropHurt
inject noise in retrieved and random documents (ablation)LUMINA · RAGTruth, other LLMsAUPRCNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy