Aligning LLM-as-a-Judge with Human Preferences

When measuring the performance of LLM applications, accurate evaluation is difficult with existing hard-coded rules or unit tests due to the nature of natural language outputs. To address this, the LLM-as-a-Judge approach, which uses a separate LLM as an evaluator, is widely used.

However, LLM-as-a-Judge has the drawback of requiring a separate prompt engineering process to design the evaluation prompt in order to work properly. To solve this, LangSmith has introduced a new 'self-improvement' solution.

The core of this feature is as follows:

• When a user modifies the output of LLM-as-a-Judge, that human feedback is stored as a Few-shot example.
• The stored examples are automatically included in the prompt during the next evaluation, improving the performance of the evaluation model.
• As a result, it is possible to build a customized evaluator that reflects user preferences without complex prompt engineering.

LLM-as-a-Judge is used in various use cases such as detecting RAG hallucinations, checking correctness, and toxicity checks, and by providing Ground Truth information at evaluation time, it can determine answers with higher accuracy than at the generation stage.