Research map / Large language models / Language, safety and interpretability
Beyond Next Token Probabilities: Learnable, Fast Detection of Hallucinations and Data Contamination on LLM Output Distributions
Guy Bar-Shalom, Fabrizio Frasca, Derek Lim, Yoav Gelberg, Yftah Ziser, Ran El-Yaniv, Gal Chechik, 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 |
|---|---|---|---|
| add the full token distribution sequence (LOS-Net) instead of only actual token probabilities and rank (ATP+R-Transf.) (ablation) | LOS-Net · IMDB hallucination detection, Mistral-7b-instruct | test AUC: 89.64 -> 94.73 | Improved |
| add the full token distribution sequence instead of ATP and rank only (ablation) | LOS-Net · BookMIA data contamination detection, Pythia-6.9b | AUC: 79.59 -> 90.71 | Improved |
| add the full token distribution sequence instead of ATP and rank only (ablation) | LOS-Net · HotpotQA hallucination detection, Llama3-8b-instruct | test AUC: 66.72 -> 72.60 | Improved |
| increase top-K retained probabilities from 10 to 1000 (ablation) | LOS-Net · HotpotQA, Mistral-7b-instruct | test AUC: 71.82 -> 72.92 | Improved |
| use only K=10 top probabilities (API-limited access) (ablation) | LOS-Net · HotpotQA, Mistral-7b-instruct | test AUC: 72.92 -> 71.82 | No clear effect |
| learned LOS-Net instead of heuristic gray-box scores and prompting-based methods | LOS-Net · IMDB hallucination detection, Mistral-7b-instruct | test AUC: around 31 AUC units over best baseline | Improved |
| learned LOS-Net instead of reference-free heuristics (Min-K%, Min-K%++, Loss) | LOS-Net · BookMIA, Pythia-6.9b | AUC: 68.78 -> 90.71 | Improved |
| zero-shot transfer of LOS-Net to unseen LLMs without fine-tuning | LOS-Net · BookMIA cross-LLM | AUC: best among reference-free in 10/12 cases | Improved |
| zero-shot cross-LLM application for hallucination detection | LOS-Net · hallucination detection | AUC | Hurt |
| fine-tune a pretrained LOS-Net on other datasets/LLMs instead of training from scratch | LOS-Net · hallucination detection transfer | test AUC: better in 16/18 cases | Improved |
| train on Llama3-8b and test on Mistral-7b with fine-tuning | LOS-Net · IMDB hallucination detection, cross-LLM | test AUC: around 27 AUC units over best probability-based baseline | Improved |
| learned LOS-Net vs white-box Activation Probe | LOS-Net · hallucination detection, 6 dataset/LLM combos | test AUC: better in 2/6, similar in 3/6 | Mixed |
Closest papers by meaning
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.93)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.93)
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.91)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.91)
- Robust Hallucination Detection in LLMs via Adaptive Token Selection (NeurIPS 2025 · similarity 0.91)
- ANAH-v2: Scaling Analytical Hallucination Annotation of Large Language Models (NeurIPS 2024 · 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.
