TL;DR

A large language model (LLM) is an AI system trained on massive text datasets to understand and generate natural language. LLMs power most modern AI recruitment tools - resume parsing, semantic search, conversational AI, content generation, and candidate scoring. GPT-4, Claude, Gemini, and Llama are the most widely deployed LLMs underlying enterprise recruitment technology as of 2024.

How Large Language Models Work

An LLM doesn't know anything - it predicts what comes next based on patterns in everything it was trained on. This distinction is critical for understanding both the capability and the limits of LLM-powered recruitment tools.

The transformer architecture, introduced in 2017, is the foundation of every major LLM. It processes text by representing each word or token as a vector and then computing attention - how much each token should influence the interpretation of every other token in the sequence. A sentence like "He's a strong closer, but struggles with compliance" requires the model to understand that "closer" means something specific in sales, that "compliance" is a shortcoming, and that these are in tension. Attention mechanisms allow the model to connect these concepts across the sentence.

Training involves processing hundreds of billions of tokens of text: web pages, books, academic papers, code, and curated datasets. The model adjusts billions of parameters to minimize prediction error on held-out text. After training, it can generate contextually appropriate text for almost any topic - not because it has a database of answers, but because it has learned the statistical structure of language at scale.

Fine-tuning and alignment are what make base models usable in recruitment applications. Fine-tuning trains the model on domain-specific examples: recruitment emails, job descriptions, interview transcripts. Alignment techniques like RLHF (reinforcement learning from human feedback) shape the model to follow instructions, avoid harmful outputs, and produce responses that humans prefer. A customer-facing recruitment chatbot and the base model it's built on may share the same architecture but behave very differently because of post-training.

Context windows define how much information the model can hold in its working memory during a single interaction. Early GPT models had context windows of 4,000 tokens (roughly 3,000 words). Current models range from 128,000 tokens (Claude 3) to 1 million tokens (Gemini 1.5 Pro). For recruitment, larger context windows mean the model can process an entire candidate file, a complete job description, and a conversation history simultaneously - enabling more coherent and context-aware responses.

Why It Matters in Recruitment

Every significant AI capability that has entered recruitment in the last three years runs on an LLM. Resume parsing that understands context, semantic search that finds candidates without exact keyword match, conversational AI that handles candidate questions, JD generation that produces compelling copy - these all depend on LLM capability. Understanding the foundation helps recruitment leaders make better vendor decisions and set appropriate expectations.

The quality gap between LLM-powered tools and their predecessors is significant. Pre-LLM resume parsing relied on named entity recognition models trained on labeled datasets - they required structured resume formats and failed on unusual layouts. LLM-based parsing reads resumes the way a human does, inferring structure and meaning from context. Error rates on LLM-based parsing are roughly 60-70% lower than on classical NLP approaches for the same task.

For talent technology buyers, LLM provider selection now matters as much as vendor selection. The underlying model (GPT-4, Claude, Gemini) determines baseline capability. Vendors that build on top of the most capable models tend to outperform those using older architectures, though fine-tuning and integration quality matter significantly too. Contract terms around data privacy - whether candidate data is used to train the provider's models - have become a standard diligence question in vendor evaluations.

Large Language Models in Practice

A global RPO provider manages recruitment operations for 12 enterprise clients across four industries. They evaluate five AI recruitment vendors over three months. The evaluation reveals that three vendors are built on GPT-3.5, one on GPT-4, and one on a proprietary model. The GPT-4-based vendor demonstrates meaningfully better performance on two benchmarks that matter to the RPO: understanding complex, multi-requirement job descriptions with conflicting priorities, and generating candidate summaries that accurately represent nuanced profiles without over-simplification.

The RPO selects the GPT-4-based vendor and integrates it with their Bullhorn instance. Recruiters use it to generate candidate summaries, write client-facing submission notes, and run semantic searches across their database. The vendor's data processing agreement confirms that no client data trains the underlying model - a requirement the RPO's clients insisted on.

Within six months, the RPO's data shows that AI-generated submission notes (reviewed and edited by recruiters) have a 22% higher client acceptance rate than notes written manually, reflecting both quality improvement and consistency. Time spent per submission drops from 35 minutes to 12 minutes.

Key Considerations