Active Learners as Efficient PRP Rerankers
For practitioners using LLMs for reranking under limited call budgets, this work offers a more efficient and robust alternative to classical sorting-based PRP.
The paper reframes pairwise ranking prompting (PRP) for LLM reranking as active learning from noisy comparisons, showing that active rankers improve NDCG@10 per call in call-constrained settings. A randomized-direction oracle converts position bias into zero-mean noise, enabling unbiased ranking without bidirectional calls.
Pairwise Ranking Prompting (PRP) elicits pairwise preference judgments from an LLM, which are then aggregated into a ranking, usually via classical sorting algorithms. However, judgments are noisy, order-sensitive, and sometimes intransitive, so sorting assumptions do not match the setting. Because sorting aims to recover a full permutation, truncating it to meet a call budget does not produce a dependable top-K. We thus reframe PRP reranking as active learning from noisy pairwise comparisons and show that active rankers are drop-in replacements that improve NDCG@10 per call in the call-constrained regime. Our noise-robust framework also introduces a randomized-direction oracle that uses a single LLM call per pair. This approach converts systematic position bias into zero-mean noise, enabling unbiased aggregate ranking without the cost of bidirectional calls.