Source: LLM Evaluators Recognize and Favor Their Own Generations (Panickssery et al., 2024)
Paper linking an LLM's ability to recognise its own text to its tendency to rate that text higher.
Contents
- What: research paper, arXiv:2404.13076 (NeurIPS 2024)
- Authors: Arjun Panickssery, Samuel R. Bowman, Shi Feng
- Date: 15 April 2024
- Original: arxiv.org/abs/2404.13076
Summary
Studies self-preference bias: judges such as GPT-4 and Llama 2 rate their own outputs higher than others rate them. The paper tests whether this comes from the judge recognising its own text.
Key claims
- Out of the box, models distinguish their own outputs from other LLMs’ and humans’ with non-trivial accuracy.
- Fine-tuning that changes self-recognition changes self-preference in a linear relationship.
- Controlled experiments suggest the link is causal rather than explained by simple confounders.
- Self-recognition threatens unbiased LLM-as-evaluator setups (and reward modelling).