Research map / Large language models / Reasoning and chain of thought
Global-Local Confidence Fusion for Hallucination Detection in Mathematical Reasoning Task
Bo Zhang, Cong Gao, Linkang Yang, Bingxu Han, Minghao Hu, Zhunchen Luo, Guotong Geng, Xiaoying Bai, Jun Zhang, Wen Yao, Zhong Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a global-local confidence fusion approach for detecting hallucinations in mathematical reasoning by large reasoning models, considering both answer and reasoning process consistency.
Problem
Hallucination detection for Large Reasoning Models on complex mathematical reasoning; existing LLM-based detectors focus on final answers and overlook inconsistencies between the answer and the reasoning process, limiting detection during inference.
Approach
Fuses global and local confidence signals so that inconsistencies between the final answer and intermediate reasoning are captured, aiming to detect hallucinations during inference rather than only from the final answer; the abstract is truncated so specific mechanics are not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add generated hallucination analysis (reasoning) before the confidence score (ablation) | GHDM-1.7B · G+M (ID) | F1: 0.5904 -> 0.7686 | Improved |
| add generated hallucination analysis (reasoning) before the confidence score (ablation) | GHDM-1.7B · G+M (ID) | ACC: 0.6981 -> 0.8068 | Improved |
| add generated hallucination analysis (reasoning) before the confidence score (ablation) | GHDM-1.7B · ASDiv (OOD) | F1: 0.5335 -> 0.7790 | Improved |
| add generated hallucination analysis (reasoning) before the confidence score (ablation) | GHDM-8B · SVAMP (OOD) | F1: 0.8098 -> 0.8864 | Improved |
| add DPO training on strong-weak preference pairs | Qwen3-8B · G+M (ID) | AUC: 0.8779 -> 0.9244 | Improved |
| add PRM local confidence fusion (ablation) | GHDM-8B · G+M (ID) | F1: 0.8320 -> 0.8593 | Improved |
| add PRM local confidence fusion (ablation) | GHDM-8B · ASDiv (OOD) | F1: 0.8550 -> 0.8709 | Improved |
| add PRM local confidence fusion (ablation) | Qwen3-8B base (no DPO) · G+M (ID) | F1: 0.8047 -> 0.8050 | No clear effect |
| add PRM local confidence fusion (ablation) | Qwen3-8B base (no DPO) · ASDiv (OOD) | F1: 0.8363 -> 0.8346 | No clear effect |
| add PRM local confidence fusion (ablation) | Qwen3-1.7B base (no DPO) · G+M (ID) | F1: 0.6822 -> 0.7234 | Improved |
| add DPO training on strong-weak preference pairs | Qwen3-8B · G+M (ID) | PCC with PRM: 0.7429 -> 0.6911 | Improved |
| raise global weight lambda to 0.75 (vs lower) (ablation) | ConfFuse-8B · G+M (ID) | AUC: peak AUC 0.936 at lambda = 0.75 | Improved |
Closest papers by meaning
- Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models (AAAI 2026 · similarity 0.93)
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.92)
- MIRAGE: Assessing Hallucination in Multimodal Reasoning Chains of MLLM (NeurIPS 2025 · similarity 0.92)
- Attributive Reasoning for Hallucination Diagnosis of Large Language Models (AAAI 2025 · similarity 0.92)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.92)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 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.
