Abstract
State-of-the-art Information Retrieval (IR) pipelines often rely on dense neural models to index and retrieve large document collections. After an initial retrieval stage, precision-oriented re-ranking approaches can further refine the ranked list. However, re-ranking quality strongly depends on the availability of reliable relevance signals. Recent work has explored the use of Large Language Models (LLMs) to automatically generate relevance judgments that can serve as such signals. However, achieving strong alignment with human gold labels generally requires either very large and complex LLMs or extensive human supervision. In this work, we investigate whether relatively small LLMs, despite producing limited and often noisy relevance assessments, can still be leveraged to improve the ranking of documents retrieved by dense bi-encoder models. Specifically, we introduce new approaches based on the recently proposed Dimension IMportance Estimators (DIMEs), incorporating LLM-generated relevance judgments into the re-ranking process. Our study assumes a practical setting: we consider access only to relatively small LLMs and to very small sets of relevance judgments generated for the top-retrieved documents of each query by dense retrieval models, and we investigate how to leverage these sparse, often noisy signals to refine the ranking of the entire retrieved set. Our method leads to substantial improvements in document ranking, boosting nDCG@10 and AP by up to 9% and 15%, respectively, across different datasets, LLMs, and bi-encoder backbone retrieval models. These results show that even a few noisy relevance judgments from small LLMs can provide a powerful signal to improve the effectiveness of IR systems.

This publication has 32 references indexed in Scilit: