# GeoDecider: Hierarchical Agent Workflow Based on Large Language Models Enables Interpretable Rock Classification

> GeoDecider proposes a training-free hierarchical agent workflow that reconstructs the rock classification task into an expert-like structured reasoning process. Through three stages—coarse-grained classification, tool-augmented reasoning, and geological consistency refinement—it achieves better results than traditional baselines on four benchmark datasets while maintaining good interpretability and reasoning efficiency.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-05-05T05:42:51.000Z
- 最近活动: 2026-05-06T02:26:17.974Z
- 热度: 121.3
- 关键词: 岩石分类, 智能体工作流, 大语言模型, 可解释AI, 地质学, 测井数据, 多阶段推理, 工具增强
- 页面链接: https://www.zingnex.cn/en/forum/thread/geodecider
- Canonical: https://www.zingnex.cn/forum/thread/geodecider
- Markdown 来源: floors_fallback

---

## Introduction: GeoDecider—Hierarchical Agent Workflow Based on Large Language Models Enables Interpretable Rock Classification

This article proposes GeoDecider, a training-free hierarchical agent workflow that reconstructs rock classification into an expert-like structured reasoning process, consisting of three stages: coarse-grained classification, tool-augmented reasoning, and geological consistency refinement. This method outperforms traditional baselines on four benchmark datasets while maintaining good interpretability and reasoning efficiency.

## Background: Challenges in Rock Classification and Limitations of Traditional Methods

Rock classification is a core task in petroleum exploration and geological research, which requires inferring rock types from logging signals. Traditional machine learning methods treat it as a one-time classification problem, ignoring the multi-round reasoning of experts and the application of geological principles; although existing deep learning methods have good accuracy, they lack interpretability and it is difficult to ensure that results conform to geological laws.

## Core Idea: From One-Time Classification to Agent Workflow

The core innovation of GeoDecider lies in redefining rock classification as an expert-like structured reasoning process. It draws on the ideas of agent systems to design a multi-stage workflow from coarse to fine, leveraging the reasoning capabilities of large language models (LLMs) without the need for domain-specific training. Key insight: Although LLMs cannot directly process raw logging data, they excel in knowledge integration, logical reasoning, and tool usage. When combined with tool interfaces, they can build an efficient classification system that understands geological principles.

## Detailed Explanation of the Three-Stage Workflow Architecture

GeoDecider consists of three connected stages:
1. **Coarse-grained classification**: Use a pre-trained traditional classifier for initial classification, output a natural language description with confidence as the LLM context to reduce subsequent reasoning costs.
2. **Tool-augmented fine-grained reasoning**: The LLM calls tools such as context analysis, neighborhood retrieval, and geological rule checking, collects evidence through chain-of-thought, and makes precise judgments.
3. **Geological consistency refinement**: Check the sequence of classification results, correct outliers based on geological priors, and ensure the correlation and continuity of rock types in adjacent strata.

## Experimental Verification and Performance

Evaluation on four public benchmark datasets shows:
1. **Accuracy improvement**: Outperforms representative traditional baselines on all datasets;
2. **Interpretability advantage**: Each decision can be traced through the reasoning chain, with clear visibility of tool calls, evidence, and judgment basis;
3. **Efficiency balance**: The coarse classification stage significantly reduces the frequency of LLM reasoning and computational overhead, maintaining high performance while reducing costs.

## Technical Insights and Future Outlook

GeoDecider provides a paradigm for the intelligent processing of domain-specific tasks: combining the general reasoning capabilities of LLMs with domain tools and knowledge to build a structured agent workflow, which is superior to end-to-end deep learning. It reveals the application trend of LLMs: evolving from prompt engineering to agent architecture. For the petroleum geology industry, this method focuses on the interpretability of results and geological rationality, making it suitable for high-risk decision-making scenarios. In the future, similar paradigms can be explored in more vertical fields.