# ALS Atomic Logic Sheet Empirical Study: How Structured Prompting Boosts Code Generation Accuracy from 6.9% to 100%

> An empirical study by the Korea University of Science and Technology team verified the effectiveness of the Atomic Logic Sheet (ALS) in LLM code generation through 720 Claude Sonnet 4.5 inference calls, showing that ALS increased accuracy in key scenarios from 6.9% to 100%.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-05-24T07:12:06.000Z
- 最近活动: 2026-05-24T07:25:33.505Z
- 热度: 157.8
- 关键词: ALS, 原子逻辑表, 提示工程, 代码生成, Claude, 实证研究, WMS
- 页面链接: https://www.zingnex.cn/en/forum/thread/als-6-9-100
- Canonical: https://www.zingnex.cn/forum/thread/als-6-9-100
- Markdown 来源: floors_fallback

---

## Introduction / Main Floor: ALS Atomic Logic Sheet Empirical Study: How Structured Prompting Boosts Code Generation Accuracy from 6.9% to 100%

An empirical study by the Korea University of Science and Technology team verified the effectiveness of the Atomic Logic Sheet (ALS) in LLM code generation through 720 Claude Sonnet 4.5 inference calls, showing that ALS increased accuracy in key scenarios from 6.9% to 100%.

## Original Authors and Source

- **Original Authors/Maintainers**: Mansu Kim, Museong Choi, Eunyoung Wang, Sungtaek Chung (Korea University of Science and Technology)
- **Source Platforms**: GitHub / Zenodo
- **Original Project Name**: ALS-WMS-Research-Experiment-N20
- **Original Link**: https://github.com/tank9567/ALS-WMS-Research-Experiment-N20
- **Publication Date**: May 24, 2026
- **DOI**: 10.5281/zenodo.20363098
- **License**: CC BY 4.0

---

## Research Background: Scientific Exploration of Prompt Engineering

In the field of LLM code generation, a long-standing challenge is: **How to convert vague natural language requirements into precise instructions that models can execute accurately?**

Traditional prompt engineering often relies on empirical trial and error, lacking systematic methodological guidance. The research team from Korea University of Science and Technology proposed a structured prompting method called **Atomic Logic Sheet (ALS)** and verified its effectiveness through rigorous empirical research.

This study focused on the WMS (Warehouse Management System) code generation scenario, using the Claude Sonnet 4.5 model for 720 inference calls, providing valuable quantitative data for the field of prompt engineering.

---

## Experimental Design

Research used a three-group control design, with 20 runs per group and 12 task scenarios in total:

| Group | Input Conditions | Description |
|------|---------|------|
| Group A | Requirement description only | Baseline control group |
| Group B | Requirement + Natural language design | Conventional best practice |
| Group C | Requirement + Natural language design + ALS | Experimental group |

## Key Results

| Metric | Group A (Requirement Only) | Group B (+NL Design) | Group C (+ALS) |
|------|--------------|---------------|------------|
| LCR (Line Coverage Rate) | 93.4% | 96.8% | 96.9% |
| **CDR (Critical Scenario Detection Rate)** | **6.9%** | **41.9%** | **100.0%** |
| SDC (Safety Defect Count) | 62.9 | 1.1 | 7.0 |

**Core Finding**: In the core metric of Critical Scenario Detection (CDR), ALS increased accuracy from 6.9% (requirement only) and 41.9% (+natural language design) to **100%**.

---

## Concept Definition

Atomic Logic Sheet is a structured requirement representation method that decomposes business rules into atomic-level logical units and organizes them in a table format. Each atomic logic includes:

- **Condition**: Premise that triggers the rule
- **Action**: Operation to execute when the condition is met
- **Priority**: Conflict resolution mechanism between rules

## Comparison with Traditional Methods

| Method | Expression Form | Precision | Verifiability |
|------|---------|--------|---------|
| Pure natural language | Text paragraphs | Low | Difficult |
| Natural language design | Structured text | Medium | Relatively difficult |
| **ALS** | **Atomic logic sheet** | **High** | **Easy** |

## Unique Value of ALS

The study found that ALS has unique advantages in capturing **process control violations**. Especially in scenarios 5-7 (removal of safety stock checks):

- Group A detection rate: 0/20 (0%)
- Group B detection rate: 0/20 (0%)
- Group C detection rate: 20/20 (100%)

This indicates that certain business rule constraints **cannot be effectively expressed through natural language specifications**, while the atomic logical representation of ALS can accurately capture these constraints.

---