Research map / Detection and segmentation / Open-set and few-shot detection
Beyond In-Domain Detection: SpikeScore for Cross-Domain Hallucination Detection
Yongxin Deng, Zhen Fang, Sharon Li, Ling Chen
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 SpikeScore (max second-order difference over multi-turn trajectory) instead of single-turn SAPLMA score | SAPLMA probe · Cross-domain average, Llama-3.1-8B | mean AUROC: 0.5764 -> 0.7860 | Improved |
| use SpikeScore instead of single-turn SEP score | SEP probe · Cross-domain average, Llama-3.1-8B | mean AUROC: 0.5282 -> 0.7684 | Improved |
| SpikeScore vs ICR Probe, strongest cross-domain baseline | ICR Probe · Cross-domain average, Qwen3-14B | mean AUROC: 0.7435 -> 0.7874 | Improved |
| SpikeScore vs ICR Probe on smaller model | ICR Probe · Cross-domain average, Llama-3.2-3B | mean AUROC: 0.7463 -> 0.7474 | No clear effect |
| use SAPLMA backbone instead of Perplexity for SpikeScore (ablation) | SpikeScore · Backbone average, Llama-3.1-8B | mean AUROC: 0.7160 -> 0.7860 | Improved |
| use SAPLMA backbone instead of Reasoning score for SpikeScore (ablation) | SpikeScore · Backbone average, Llama-3.1-8B | mean AUROC: 0.7595 -> 0.7860 | Improved |
| use SAPLMA backbone instead of In-Context Sharpness (ablation) | SpikeScore · Backbone average, Llama-3.1-8B | mean AUROC: 0.7311 -> 0.7860 | Improved |
| use SAPLMA backbone instead of SEP for SpikeScore (ablation) | SpikeScore · Backbone average, Qwen3-14B | mean AUROC: 0.7718 -> 0.7874 | Improved |
| train-based backbones (SEP) instead of training-free (Perplexity) under SpikeScore (ablation) | SpikeScore · Backbone average, Qwen3-8B | mean AUROC: 0.6742 -> 0.7293 | Improved |
| use SpikeScore in retrieval-augmented pipeline vs ICR Probe | ICR Probe · RAGTruth, Llama-3.1-8B | AUROC: 0.7405 -> 0.8154 | Improved |
| use SpikeScore in RAG vs ICR Probe | ICR Probe · TriviaQA RAG, Qwen3-8B | AUROC: 0.7495 -> 0.8413 | Improved |
| use second-order difference instead of coefficient of variation to measure fluctuation (ablation) | SpikeScore · CoQA-trained, Llama-3.1-8B | AUROC | Improved |
Closest papers by meaning
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.92)
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.92)
- 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 0.90)
- ANAH-v2: Scaling Analytical Hallucination Annotation of Large Language Models (NeurIPS 2024 · similarity 0.90)
- Entropy Distribution as a Fingerprint for Hallucinations in Generative Models (NeurIPS 2026 · 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.
