Research map / Graph learning / Graph self-supervised learning
Neural Message-Passing on Attention Graphs for Hallucination Detection
Fabrizio Frasca, Guy Bar-Shalom, Yftah Ziser, Haggai Maron
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 CHARM (GNN on attention graphs) instead of Lookback Lens | hallucination detector · NQ (token-level) | AUPR (%): 34.3 -> 40.3 | Improved |
| use CHARM instead of Lookback Lens | hallucination detector · CNN (token-level) | AUROC (%): 74.4 -> 75.4 | Improved |
| add layer-24 activations to node features (ablation) | CHARM · NQ (token-level) | AUPR (%): 40.3 -> 35.5 | Hurt |
| add layer-24 activations to node features (ablation) | CHARM · CNN (token-level) | AUROC (%): 75.4 -> 70.9 | Hurt |
| add layer-24 activations to node features (ablation) | CHARM · Winobias (response-level) | AUROC (%): 70.4 -> 77.8 | Improved |
| add layer-24 activations to node features (ablation) | CHARM · Math (response-level) | AUROC (%): 76.5 -> 80.8 | Improved |
| add layer-24 activations to node features (ablation) | CHARM · Movies (response-level) | AUROC (%): 80.3 -> 79.7 | No clear effect |
| use message passing on the attention graph instead of dense layers over nodes (ablation) | CHARM · CNN (token-level) | AUROC (%): 70.8 -> 75.4 | Improved |
| use message passing on the attention graph instead of dense layers (ablation) | CHARM · Math (response-level) | AUROC (%): 80.6 -> 81.7 | Improved |
| sparsify attention graph with higher threshold tau 0.5 instead of 0.05 (ablation) | CHARM · NQ (token-level) | AUPR (%): 40.3 -> 38.4 | Hurt |
| sparsify with tau=0.05 instead of 0.001 (ablation) | CHARM · NQ (token-level) | AUPR (%): 40.1 -> 40.3 | No clear effect |
| use CHARM for zero-shot transfer instead of Lookback Lens | hallucination detector · NQ -> CNN and CNN -> NQ | AUROC (%): NQ->CNN 68.6 -> 64.1; CNN->NQ 62.0 -> 65.5 | Mixed |
Closest papers by meaning
- Beyond the Global Scores: Fine-Grained Token Grounding as a Robust Detector of LVLM Hallucinations (CVPR 2026 · similarity 0.91)
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.90)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.90)
- Entropy Distribution as a Fingerprint for Hallucinations in Generative Models (NeurIPS 2026 · similarity 0.90)
- PAS: Prelim Attention Score for Detecting Object Hallucinations in Large Vision-Language Models (CVPR 2026 · similarity 0.90)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.89)
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.
