Research map / Large language models / Retrieval and knowledge
Graph of Verification: Structured Verification of LLM Reasoning with Directed Acyclic Graphs
Jiwei Fang, Bin Zhang, Changwei Wang, Jin Wan, Zhiwei Xu
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 |
|---|---|---|---|
| decompose verification into atomic nodes with minimal context (GoV) instead of holistic verification | Qwen2.5-72B-Instruct verifier · Number Triangle Summation (N=8) | F1: 49.5 -> 98.1 | Improved |
| decompose verification into atomic nodes (GoV) instead of holistic | Qwen2.5-32B-Instruct verifier · Number Triangle Summation (N=2) | F1: 100.0 -> 99.8 | No clear effect |
| decompose verification into atomic nodes (GoV) instead of holistic | Qwen2.5-72B-Instruct verifier · Number Triangle Summation (N=2) | F1: 99.6 -> 98.7 | Hurt |
| paragraph-level node blocks with inclusive context (GoV) instead of holistic | Qwen2.5-32B-Instruct verifier · ProcessBench OlympiadBench | F1: 35.9 -> 65.2 | Improved |
| paragraph-level node blocks (GoV) instead of holistic | Qwen2.5-7B-Instruct verifier · ProcessBench GSM8K | F1: 47.3 -> 58.0 | Improved |
| paragraph-level node blocks (GoV) instead of holistic | Llama3.3-70B-Instruct verifier · ProcessBench GSM8K | F1: 78.6 -> 78.6 | No clear effect |
| paragraph-level node blocks (GoV) instead of holistic | Llama3.3-70B-Instruct verifier · ProcessBench MATH | F1: 54.4 -> 71.5 | Improved |
| use block granularity with inclusive context (GoV) instead of PARC premise extraction | Qwen2.5-32B-Instruct verifier · ProcessBench GSM8K | F1: 70.0 -> 78.1 | Improved |
| use block granularity with inclusive context (GoV) instead of PARC | Qwen2.5-72B-Instruct verifier · ProcessBench MATH | F1: 66.5 -> 75.2 | Improved |
| training-free GoV with Qwen2.5-7B versus trained PRMs | ProcessBench · ProcessBench OlympiadBench | F1: 22.9 -> 42.9 | Improved |
| training-free GoV with Qwen2.5-7B versus Skywork-PRM-7B | ProcessBench · ProcessBench GSM8K | F1: 70.8 -> 58.0 | Hurt |
| iterative per-node verification (GoV) instead of single-pass CoT | Qwen2.5-32B-Instruct verifier · Number Triangle Summation (500 instances) | total cost ($): 0.140 -> 1.329 | Hurt |
Closest papers by meaning
- Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning (ICLR 2024 · similarity 0.89)
- Variation in Verification: Understanding Verification Dynamics in Large Language Models (ICLR 2026 · similarity 0.88)
- Inference-Time Conformal Reasoning with Valid Factuality Control for Large Language Models (ICML 2026 · similarity 0.88)
- Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models (ICML 2025 · similarity 0.88)
- Putnam-AXIOM: A Functional & Static Benchmark for Measuring Higher Level Mathematical Reasoning in LLMs (ICML 2025 · similarity 0.88)
- Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph (ICLR 2024 · similarity 0.88)
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.
