Recruiting · Production

How LinkedIn built domain-adapted EON foundation models to power Hiring Assistant candidate matching

The problem

LinkedIn's earlier fine-tuned domain-specific model required significant effort to add new features or generalize to new applications, because representative high-quality training data had to be regenerated and the model retrained for each change. Proprietary GPT-4-class models also posed latency and cost concerns.

Workflow diagram · grounded in source
1
Recruiter query intake
trigger
“This product automates recruiters' repetitive tasks by breaking down queries into multiple steps through an orchestration layer”
2
Orchestration layer decomposes query
ai_action
“breaking down queries into multiple steps through an orchestration layer”
3
Evaluation agent processes candidate data
ai_action
“Nearly 90% of large language model (LLM) calls in the hiring assistant flow come from the evaluation agent. This agent uses a language model to parse extensive contexts, including the candidate's profile, resume, recruiter notes, and job…”
4
Match output with explanation
output
“providing both a categorical output indicating the candidate's alignment with job requirements and a concise explanation and summary of the findings”
5
Recruiter approves matching criteria
human_review
“This allows recruiters to easily approve and update AI-generated matching criteria, streamlining the process of sourcing the right candidates for a job”
Reported outcome

In testing, LinkedIn's EON-8B model is 75x cheaper than GPT-4 and 6x cheaper than GPT-4o, and significantly improved candidate-job-requirements matching accuracy in Hiring Assistant, outperforming GPT-4o mini by an absolute 4% and Llama-3-8B-instruct by an absolute 30%.
Prompt simplification strategies also reduced prompt size by 30%.

Reported metrics
cost efficiency vs GPT-475x
cost efficiency vs GPT-4o6x
Prompt size reduction30%
LLM calls from evaluation agentNearly 90%
Show all 7 reported metrics
cost efficiency vs GPT-475x
cost efficiency vs GPT-4o6x
prompt size reduction30%
LLM calls from evaluation agentNearly 90%
matching accuracy improvement over GPT-4o miniabsolute 4%
matching accuracy improvement over Llama-3-8B-instructabsolute 30%
training data sizearound 200M tokens
Reported stack
Hiring AssistantEONRAGRLHFDPOLlamaMixtralMistralKubernetesDeepSpeedvLLMMLFlowHDFSGPT-4LM Evaluation HarnessMeta
Source
https://www.linkedin.com/blog/engineering/generative-ai/how-we-built-domain-adapted-foundation-genai-models-to-power-our-platform
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

In testing, LinkedIn's EON-8B model is 75x cheaper than GPT-4 and 6x cheaper than GPT-4o, and significantly improved candidate-job-requirements matching accuracy in Hiring Assistant, outperforming GPT-4o mini by an ab…

What tools did this team use?

Hiring Assistant, EON, RAG, RLHF, DPO, Llama, Mixtral, Mistral, Kubernetes, DeepSpeed.

What results were reported?

cost efficiency vs GPT-4: 75x; cost efficiency vs GPT-4o: 6x; Prompt size reduction: 30%; LLM calls from evaluation agent: Nearly 90% (source-reported, not independently verified).

How is this recruiting AI workflow structured?

Recruiter query intake → Orchestration layer decomposes query → Evaluation agent processes candidate data → Match output with explanation → Recruiter approves matching criteria.

Related recruiting cases