Core Concepts
Pairwise preference search improves alignment between LLM evaluators and human judgements.
Abstract
Large Language Models (LLMs) struggle to align with human judgement.
Calibration methods are insufficient for effective alignment.
Pairwise-preference Search (PAIRS) introduces a new evaluation paradigm.
PAIRS outperforms direct scoring and calibration techniques.
Transitivity is crucial for evaluating LLMs effectively.
Stats
"PAIRS achieves state-of-the-art performance on representative evaluation tasks."
"Misalignment in evaluation is not primarily due to biased priors over evaluation scores."
"The likelihood term reflects expected output candidates for a given score."
Quotes
"PAIRS achieves unique scalability in aligning LLM evaluations."
"Calibration consistently improves performance for Mistral 7B and Llama-2 7B."