Research map / Large language models / Language, safety and interpretability
Lyapunov Probes for Hallucination Detection in Large Foundation Models
Bozhi Luan, Gen Li, Yalan Qin, Jifeng Guo, Yun Zhou, Faguo Wu, Hongwei Zheng, Wenjun Wu, Zhaoxin Fan
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 |
|---|---|---|---|
| add Lyapunov constraint loss (ablation) | Lyapunov Probe · TriviaQA | AUPRC: Qwen-3-4B 74.19 -> 79.47; Llama-2-7B 78.13 -> 83.09 | Improved |
| use multi-layer hidden states instead of single layer (ablation) | Lyapunov Probe · TriviaQA | AUPRC: Llama-2-7B 77.34 -> 83.09; Falcon-7B 60.50 -> 65.52 | Improved |
| add perturbation data in training (ablation) | Lyapunov Probe · TriviaQA | AUPRC: Llama-3-8B 82.35 -> 86.46; Qwen-3-4B 79.92 -> 79.47 | Mixed |
| add two-stage training (ablation) | Lyapunov Probe · TriviaQA | AUPRC: Llama-3-8B 84.80 -> 86.46; Qwen-3-4B 80.58 -> 79.47 | Mixed |
| replace standard probe with Lyapunov Probe | Llama-3-8B · TriviaQA | AUPRC: 78.82 -> 86.46 | Improved |
| replace standard probe with Lyapunov Probe | Qwen-3-4B · PopQA | AUPRC: 64.41 -> 64.02 | No clear effect |
| replace standard probe with Lyapunov Probe | Llama-2-7B · MMLU | AUPRC: 33.60 -> 33.79 | No clear effect |
| replace standard probe with Lyapunov Probe | LLaVA-1.5-7B · VizWiz-VQA | AUPRC: 77.02 -> 83.18 | Improved |
| replace standard probe with Lyapunov Probe | LLaVA-1.5-7B · POPE | AUPRC: 98.08 -> 99.13 | No clear effect |
| train probe on TriviaQA and test on other datasets instead of in-domain | Lyapunov Probe · CoQA (cross-domain, Llama-3-8B) | AUPRC: 81.28 -> 67.65 | Hurt |
Closest papers by meaning
- Entropy Distribution as a Fingerprint for Hallucinations in Generative Models (NeurIPS 2026 · similarity 0.92)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.91)
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.91)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.90)
- Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering (ICLR 2025 · similarity 0.90)
- Robust Hallucination Detection in LLMs via Adaptive Token Selection (NeurIPS 2025 · 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.
