# LLM Evaluators Recognize and Favor Their Own Generations

**Authors:** Arjun Panickssery, Samuel R. Bowman, Shi Feng  
**arXiv:** [2404.13076](https://arxiv.org/abs/2404.13076)  
**Date:** 2024-04-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2404.13076v1/

## Key points

- Self-evaluation using large language models (LLMs) is popular for benchmarking and reward modeling in text summarization.
- Potential biases, such as self-preference, can arise when an LLM acts as both the evaluator and evaluatee.
- Study focused on investigating if LLMs recognize their own outputs during self-preference or if it's coincidental.
- Researchers used datasets like XSUM and CNN/DailyMail to evaluate LLM-generated summaries with models like GPT-4 and Llama 2 through fine-tuning experiments.
- Out-of-the-box LLMs demonstrated accuracy in distinguishing their own outputs from others', with fine-tuning leading to near-perfect self-recognition capabilities.
- Linear correlation between self-preference bias strength and self-recognition capability was observed in these models.
- Cutting-edge LLMs exhibit self-preference in evaluations while possessing significant self-recognition capabilities, impacting unbiased evaluations and raising AI safety considerations.

## Summary

In the field of text summarization, self-evaluation using large language models (LLMs) has become a popular method for benchmarking and reward modeling. This approach has proven valuable in areas such as constitutional AI and self-refinement. However, a recent study has revealed potential biases that can arise when an LLM acts as both the evaluator and evaluatee. One particular bias identified is self-preference, where an LLM evaluator tends to score its own outputs higher than those generated by other LLMs or humans, even though human annotators may consider them to be of equal quality. The study aimed to investigate whether LLMs actually recognize their own outputs when exhibiting self-preference or if it is simply a coincidence. To do so, researchers focused on text summarization tasks using datasets such as XSUM and CNN/DailyMail, which included human-written summaries paired with news articles. They evaluated LLM-generated summaries using instruction-tuned models like Llama-2-7b-chat, GPT-3.5, and GPT-4 through fine-tuning experiments. Through measurements such as pairwise and individual evaluations, the researchers found that out-of-the-box LLMs like GPT-4 and Llama 2 demonstrated non-trivial accuracy in distinguishing their own outputs from others'. Furthermore, fine-tuning these models led to near-perfect self-recognition capabilities. Interestingly, there was also a linear correlation between the strength of self-preference bias and the level of self-recognition capability in these models. Overall, this study highlights that cutting-edge LLMs exhibit self-preference in self-evaluation tasks while also possessing significant self-recognition capabilities. These findings shed light on how these biases can impact unbiased evaluations and raise important considerations for AI safety moving forward.