Research map / Large language models / Language, safety and interpretability
From Detection to Diagnosis: Advancing Hallucination Analysis with Automated Data Synthesis
Yanyi Liu, Qingwen Yang, Tiezheng Guo, Feiyu Qu, Jun Liu, Yingyou Wen
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 |
|---|---|---|---|
| GRPO training with HDG-synthesized data (reasoning mode) | Qwen3-4B · detection average of seven | macro F1: 76.74 -> 79.65 | Improved |
| GRPO training with HDG data (non-reasoning mode) | Qwen3-4B · detection average of seven | macro F1: 64.92 -> 70.13 | Improved |
| enable reasoning mode (ablation) | Qwen3-4B · detection average of seven | macro F1: 64.92 -> 76.74 | Improved |
| enable reasoning mode (ablation) | HDM-4B-RL · detection average of seven | macro F1: 70.13 -> 79.65 | Improved |
| HDM-4B-RL vs specialized detector Lynx-8B | HDM-4B-RL · detection average of seven | macro F1: 77.25 -> 79.65 | Improved |
| HDM-4B-RL vs Lynx-8B on HaluEval | HDM-4B-RL · HaluEval | macro F1: 86.96 -> 83.96 | Hurt |
| HDM-4B-RL vs GPT-4.1 prompting | HDM-4B-RL · detection average of seven | macro F1: 75.37 -> 79.65 | Improved |
| HDM-4B-RL vs o4-mini prompting | HDM-4B-RL · detection average of seven | macro F1: 82.62 -> 79.65 | Hurt |
| multi-step pipeline instead of single prompt (ablation) | Qwen3-32B · SummEval diagnosis | localization hit rate: 64.85 -> 76.97 | Improved |
| single-prompt HDM-4B-RL vs single-prompt GPT-4.1 and o4-mini | HDM-4B-RL · SummEval diagnosis | detection accuracy: 77.54 (GPT-4.1) -> 92.28 | Improved |
| single-prompt HDM-4B-RL vs Qwen3-32B single prompt | HDM-4B-RL · SummEval diagnosis | mitigation AlignScore: 70.98 -> 69.16 | Mixed |
Closest papers by meaning
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.95)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.94)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.94)
- Entropy Distribution as a Fingerprint for Hallucinations in Generative Models (NeurIPS 2026 · similarity 0.92)
- Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback (AAAI 2025 · similarity 0.92)
- Attributive Reasoning for Hallucination Diagnosis of Large Language Models (AAAI 2025 · similarity 0.92)
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.
