Research map / Detection and segmentation / CNN features
G2LDetect: A Global-to-Local Approach for Hallucination Detection
Read the paperCheck your idea against 108,967 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 G2LDetect, which converts a text into a hierarchical global tree with an LLM, extracts root-to-leaf path details, and detects each path with AND aggregation for hallucination detection.
Problem
Existing LLM-based hallucination detectors extract local factual units (sentences, sub-sentences, triples) directly from text without global context, causing isolated details, missed detections and over-detections, especially on longer texts.
Approach
Build a global hierarchical tree representation of the text first, then derive local details as root-to-leaf paths, detect each path against the reference, and aggregate with logical AND, so each detail is context-aware and coverage is complete.
Key results
- On HaluEval-Summary with GPT-4, G2LDetect reaches 72.98 F1 vs 70.93 for the best baseline (FacTool).
- On QAGS-CNNDM with GPT-4, G2LDetect reaches 84.44 F1 vs 79.00 for FactScore.
- On FEVER with GPT-4, G2LDetect reaches 75.55 F1 vs 72.76 for REFCHECKER; on SCIFACT with GPT-4 it reaches 94.62 F1 vs 93.55 for standard prompting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use G2LDetect instead of the strongest listed baseline | LLaMA3-8B-Instruct · HaluEval-Summary | F1: 60.76 -> 64.34 | Helped |
| use G2LDetect instead of the strongest listed baseline | ChatGPT · HaluEval-Summary | F1: 67.88 -> 70.97 | Helped |
| use G2LDetect instead of the strongest listed baseline | GPT-4 · HaluEval-Summary | F1: 70.93 -> 72.98 | Helped |
| use G2LDetect instead of the strongest listed baseline | LLaMA3-8B-Instruct · QAGS-CNNDM | F1: 64.23 -> 67.24 | Helped |
| use G2LDetect instead of the strongest listed baseline | ChatGPT · QAGS-CNNDM | F1: 64.00 -> 68.75 | Helped |
| use G2LDetect instead of the strongest listed baseline | GPT-4 · QAGS-CNNDM | F1: 79.01 -> 84.44 | Helped |
| use G2LDetect instead of the strongest listed baseline | LLaMA3-8B-Instruct · FEVER | F1: 68.50 -> 69.27 | Helped |
| use G2LDetect instead of the strongest listed baseline | ChatGPT · FEVER | F1: 65.91 -> 68.42 | Helped |
| use G2LDetect instead of the strongest listed baseline | GPT-4 · FEVER | F1: 72.76 -> 75.55 | Helped |
| use G2LDetect instead of the strongest listed baseline | LLaMA3-8B-Instruct · SCIFACT | F1: 86.02 -> 86.16 | Helped |
| use G2LDetect instead of the strongest listed baseline | ChatGPT · SCIFACT | F1: 88.15 -> 88.12 | No clear effect |
| use G2LDetect instead of the strongest listed baseline | GPT-4 · SCIFACT | F1: 93.55 -> 94.62 | Helped |
Surprising findings
- Chain-of-thought prompting underperforms standard prompting on QAGS-CNNDM with ChatGPT and GPT-4, suggesting CoT struggles with long texts.
- Rule-based sentence splitting methods (SelfCheckGPT, FActScore) are generally less effective than LLM-based extraction methods such as FacTool.
Limitations
- Evaluation uses sampled subsets (e.g., one-tenth of HaluEval-Summary) due to LLM resource limits.
- Relies on LLM-generated tree structures, and the number of path-wise local details can grow with tree size, which is not analyzed for cost.
- Human evaluation covers only 50 results per method from two annotators.
Experimental setup
Datasets: HaluEval-Summary QAGS-CNNDM FEVER SCIFACT
Benchmarks: HaluEval-Summary QAGS-CNNDM FEVER SCIFACT
Metrics: precision recall F1 (macro-average)
Baselines: Standard prompting Chain-of-thought SelfCheckGPT FActScore FacTool REFCHECKER
Builds on: SelfCheckGPT FActScore FacTool REFCHECKER chain-of-thought
Closest papers by meaning
- GLSim: Detecting Object Hallucinations in LVLMs via Global-Local Similarity (NeurIPS 2025 · similarity 0.90)
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.90)
- Zero-resource Hallucination Detection for Text Generation via Graph-based Contextual Knowledge Triples Modeling (AAAI 2025 · similarity 0.90)
- HalLoc: Token-level Localization of Hallucinations for Vision Language Models (CVPR 2025 · similarity 0.90)
- Beyond the Global Scores: Fine-Grained Token Grounding as a Robust Detector of LVLM Hallucinations (CVPR 2026 · similarity 0.90)
- LLM-Check: Investigating Detection of Hallucinations in 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.