Research map / Large language models / Language, safety and interpretability
From Out-of-Distribution Detection to Hallucination Detection: A Geometric View
Litian Liu, Reza Pourreza, Yubing Jian, Yao Qin, Roland Memisevic
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 analytical decision-neutral point as feature mean instead of empirical training mean (ablation) | NCI · CSQA (Llama-3.2-3B-Instruct) | AUROC: 62.79 -> 66.07 | Improved |
| use NCI with analytical proxy instead of Perplexity | hallucination detector · CSQA (Llama-3.2-3B-Instruct) | AUROC: 63.23 -> 66.07 | Improved |
| restrict fDBD to top-1000 alternative tokens instead of all tokens (ablation) | fDBD · CSQA (Llama-3.2-3B-Instruct) | AUROC: 68.15 -> 69.24 | Improved |
| restrict fDBD to top-100000 alternative tokens instead of 1000 (ablation) | fDBD · CSQA (Llama-3.2-3B-Instruct) | AUROC: 69.24 -> 68.26 | Hurt |
| restrict fDBD to top-1 alternative token (ablation) | fDBD · CSQA (Llama-3.2-3B-Instruct) | AUROC: 69.24 -> 68.64 | No clear effect |
| perturb selected k by 20% (ablation) | fDBD · GSM8K (Llama-3.2-3B-Instruct) | AUROC: 76.36 -> 76.32 | No clear effect |
| use fDBD with selected k instead of all tokens (ablation) | fDBD · GSM8K (Llama-3.2-3B-Instruct) | AUROC: 75.59 -> 76.36 | Improved |
| use fDBD with selected k instead of all tokens (ablation) | fDBD · CSQA (Qwen-2.5-7B-Instruct) | AUROC: 71.50 -> 72.47 | Improved |
| apply stochastic decoding at temperature 1.0 instead of 0.2 | fDBD · CSQA (Llama-3.2-3B-Instruct) | AUROC: 69.30 -> 69.19 | No clear effect |
| apply stochastic decoding at temperature 1.0 instead of 0.2 | NCI · CSQA (Llama-3.2-3B-Instruct) | AUROC: 67.07 -> 67.93 | No clear effect |
| use fDBD instead of Perplexity on larger model | fDBD (Qwen-3-32B) · GSM8K | AUROC: 76.86 -> 80.60 | Improved |
| use fDBD instead of NCI on MoE model | fDBD (Qwen-3-30B-A3B) · AQuA | AUROC: 70.09 -> 69.60 | Hurt |
Closest papers by meaning
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.93)
- Beyond the Global Scores: Fine-Grained Token Grounding as a Robust Detector of LVLM Hallucinations (CVPR 2026 · similarity 0.92)
- HalLoc: Token-level Localization of Hallucinations for Vision Language Models (CVPR 2025 · similarity 0.91)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.91)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.91)
- Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback (AAAI 2025 · similarity 0.91)
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.
