# Who Validates the Validators? Aligning LLM-Assisted Evaluation of LLM Outputs with Human Preferences

**Authors:** Shreya Shankar, J. D. Zamfirescu-Pereira, Björn Hartmann, Aditya G. Parameswaran, Ian Arawjo  
**arXiv:** [2404.12272](https://arxiv.org/abs/2404.12272)  
**Date:** 2024-04-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2404.12272v1/

## Key points

- Study focused on evaluating Large Language Models (LLMs) using EvalGen interface
- Participants evaluated NER outputs of a dataset containing 100 tweets
- Typical workflow involved grading LLM outputs and refining criteria based on EvalGen suggestions
- Participants found EvalGen helpful for generating assertions but wanted more control due to occasional mistakes
- Satisfaction expressed with using EvalGen, importance of iterating on criteria and assertions for better alignment with human requirements noted
- Challenges highlighted: subjective criteria definition, dependency on specific LLM outputs during iterative process

## Summary

The study conducted by Shreya Shankar, J.D. Zamfirescu-Pereira, Björn Hartmann, Aditya G. Parameswaran, and Ian Arawjo focused on evaluating Large Language Models (LLMs) using the EvalGen interface. Participants were tasked with evaluating NER outputs of a dataset containing 100 tweets and engaged in a typical workflow involving various activities such as grading LLM outputs and refining criteria based on suggestions from EvalGen. While participants found EvalGen to be a helpful starting point for generating assertions, they also felt the need to exert control over the process due to occasional mistakes made by the tool. Overall, participants expressed satisfaction with using EvalGen but noted the importance of being able to iterate on criteria and assertions for better alignment with human requirements. The study highlighted challenges such as subjective criteria definition and dependency on specific LLM outputs observed during the iterative process of aligning LLM-generated evaluation functions with user grades. The study conducted by Shreya Shankar et al. aimed to evaluate Large Language Models (LLMs) using the EvalGen interface. Participants were tasked with evaluating NER outputs of a dataset containing 100 tweets and engaged in a typical workflow involving activities such as grading LLM outputs and refining criteria based on suggestions from EvalGen. While participants found EvalGen to be a helpful starting point for generating assertions, they also felt the need to exert control over the process due to occasional mistakes made by the tool. Overall, participants expressed satisfaction with using EvalGen but noted the importance of being able to iterate on criteria and assertions for better alignment with human requirements. The study highlighted challenges such as subjective criteria definition and dependency on specific LLM outputs observed during the iterative process of aligning LLM-generated evaluation functions with user grades.