Research map / Large language models / Language, safety and interpretability
Enhancing Hallucination Detection through Noise Injection
Litian Liu, Reza Pourreza, Sunny Panchal, Apratim Bhattacharyya, 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 |
|---|---|---|---|
| add MLP-activation noise injection to sampling | answer-entropy detection, Llama-2-7B-chat · GSM8K | AUROC: 71.56 -> 76.14 | Improved |
| add noise injection across six LLMs | answer-entropy detection · GSM8K | AUROC: e.g. Gemma-2B 51.36 -> 57.11; Llama-3.2-3B 76.53 -> 82.70 | Improved |
| add noise injection | answer-entropy detection, Phi-3-mini · TriviaQA | AUROC: 82.00 -> 82.02 | No clear effect |
| add noise injection without hurting generation accuracy | Llama-2-7B-chat · GSM8K | accuracy: 23.64 -> 24.09 | No clear effect |
| add noise injection to five uncertainty metrics (ablation) | Predictive entropy, lexical sim., semantic entropy, EigenScore, SelfCheckGPT · TriviaQA | AUROC: e.g. 75.70 -> 77.21 (semantic entropy) | Improved |
| inject noise in lower/middle/upper/all layers (ablation) | Llama-2-7B-chat · CSQA | AUROC: 67.55 -> 69.10 to 70.68 | Improved |
| add input perturbation (shuffled ICL examples) on top of noise (ablation) | Llama-3.2-3B-Instruct · GSM8K | AUROC: 82.70 -> 82.84 | Improved |
| add input perturbation alone (ablation) | Llama-3.2-3B-Instruct · GSM8K | AUROC: 76.53 -> 76.10 | Hurt |
| use shared noise across layers rather than independent noise (ablation) | Llama-3.2-3B-Instruct · GSM8K | AUROC: 80.20 -> 82.70 | Improved |
| use positive-mean rather than zero-mean noise (ablation) | Llama-3.2-3B-Instruct · CSQA | AUROC: 72.27 -> 72.83 | No clear effect |
| use uniform instead of bounded Gaussian noise (ablation) | Llama-3.2-3B-Instruct · CSQA | AUROC: alpha 0.09: 72.64 vs 72.83 | No clear effect |
| replace additive noise with inference-time dropout (ablation) | Llama-2-7B-chat · CSQA | accuracy: 70.83 -> 67.50 | Hurt |
Closest papers by meaning
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.94)
- Entropy Distribution as a Fingerprint for Hallucinations in Generative Models (NeurIPS 2026 · similarity 0.93)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.93)
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.92)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.92)
- 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.
