Docs + Environment pattern: How to Incorporate LLM as a Judge in Verification

[cwing-nvidia]

Tutorial: How to Incorporate LLM as a Judge in Verification Logic

Background

Users have asked how to use LLM-as-a-judge for verification in their resource servers. This is particularly important for tasks where ground truth is difficult to verify (e.g., creative writing, instruction following).

Problem

Users need guidance on:

• When to use LLM-as-a-judge vs. other verification methods
• How to deploy the judge model
• How to configure the judge model endpoint

Acceptance Criteria

• [ ] Verification design - when to use LLM as a judge
• [ ] Configuration - how to set up the judge model in NeMo Gym
• [ ] Architecture - where the judge model runs
• [ ] Deployment options - co-located on the same cluster, separate endpoint/cluster, using external APIs
• [ ] point to existing resource servers that utilize LLM as a judge

Priority

High - common pattern for verification

Related

• This pattern is used internally but not documented for external users
• Consider showing integration with multiple LLM providers (OpenAI, Gemini, Claude, local models)