# FlowScore: A Hallucination Detection Method for Large Language Models Based on Structured Reasoning Models

> FlowScore is a new method for detecting hallucinations in large language models (LLMs) using structured reasoning models. It identifies erroneous steps by modeling the reasoning process as a Directed Acyclic Graph (DAG) and calculating the FLOW score, and has been accepted by ICML 2026.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-05-25T14:04:34.000Z
- 最近活动: 2026-05-25T14:20:59.205Z
- 热度: 150.7
- 关键词: 大语言模型, 幻觉检测, 结构化推理, ICML 2026, DAG, FLOW评分, 推理可靠性, 机器学习
- 页面链接: https://www.zingnex.cn/en/forum/thread/flowscore
- Canonical: https://www.zingnex.cn/forum/thread/flowscore
- Markdown 来源: floors_fallback

---

## [Introduction] FlowScore: A New LLM Hallucination Detection Method Driven by Structured Reasoning Models (Accepted by ICML 2026)

FlowScore is an innovative method for hallucination detection in large language models (LLMs). Its core is modeling the reasoning process as a Directed Acyclic Graph (DAG) and calculating the FLOW score to identify erroneous steps. This method has been accepted by ICML 2026, maintained by soncheinbok, open-sourced on GitHub (link: https://github.com/soncheinbok/FlowScore), and released on May 25, 2026.

## Background and Challenges: Pain Points in LLM Hallucination Detection

LLMs have strong reasoning capabilities, but hallucination issues hinder their reliable application. Traditional detection methods only focus on the output level and cannot delve into the internal structure of reasoning to locate the root cause. In complex tasks such as mathematical reasoning and code generation, fine-grained analysis of reasoning steps is needed to identify errors.

## Overview of the FlowScore Method

FlowScore models the reasoning process as a DAG and quantifies the reliability of steps through the FLOW score to achieve fine-grained hallucination detection. The project provides a complete reproduction package, including the SRM pipeline, ablation experiment workflow, logic records, and scoring tables, supporting the entire process from input to PR curve generation.

## Detailed Explanation of Core Technical Mechanisms

### Structured Reasoning Modeling
The reasoning process is represented as a DAG, where nodes are conditions/intermediate steps and edges are logical dependencies, capturing structural information.

### FLOW Scoring Algorithm
Analyze the information flow of the DAG to evaluate the reliability of steps, with scores ranging from 0 to 1 (higher scores indicate possible hallucinations), accurately locating weak links.

### Step-level Label System
Three-value labels (0.0 correct / 0.5 uncertain / 1.0 wrong), providing API and local model (e.g., Qwen-3.5-9B) automatic annotation tools.

## Experimental Validation and Evaluation Results

FlowScore performs excellently on GSM8K, MATH, HumanEval, and HotpotQA datasets, outperforming traditional aggregation methods. Its robustness is verified through ablation experiments (aggregation strategies, DAG perturbations, etc.), and influencing factors are insightfully analyzed.

## Practical Application Value

FlowScore helps model developers diagnose the causes of reasoning failures and guide improvements; assists application developers in building reliable interactive systems (e.g., manual verification of key steps); and its modular design makes it easy to integrate into existing evaluation processes.

## Summary and Outlook

FlowScore achieves the leap from coarse-grained to fine-grained hallucination detection, providing a new perspective for understanding LLM reasoning behavior. In the future, structured analysis is expected to promote the evolution of hallucination detection towards 'process monitoring' to meet the needs of key fields.