# Application of LLM Agents in Recruitment: A Novel Framework for Resume Screening

## 1 Introduction

- use their novel framework -> leveraging their own finetuned LLM on their curated resume dataset
- 11 times faster than manual screening
- always run into risks of private data leakage
- use local models to solve data privacy issue


## 2 Related Work

## 3 Resume Screening Framework based on LLM Agents


1. resumes (in diff file formats) -> all converted to JSON via rule-based algorithm
2. segment resume into sentences based on format (like line breaks); necessary as the LLM classifies per sentence and some aspects of resumes might be more important than others in certain scenarios
3. use an open source LLM to classify parts of the segmented sentences (e.g. experience, personal info, education, etc)
4. remove private information from the classified sentences
5. grade and summarize resume with GPT-3.5
6. Finish with an HR agent or a human HR person to decide whether or not this person should move onto interview

### 3.1 Sentence Classification


- LLaMA2 does not guarantee a classification from an input string as evidenced by the image above
- use LLaMA2 finetuned on a specially curated resume sentence instruction dataset
- use the prompt structured below during inference


### 3.2 Grade & Summarization


### 3.3 Decision Making


## 4 Experiment Setup

## 5 Results