REALM: Smarter, Faster, and More Reliable Document Re-Ranking

​

Workflow of recursive relevance modeling framework for LLM-based document re-ranking

Invention Summary:

Most Large Language Model LLM-based re-rankers produce single-point relevance scores without modeling uncertainty, even though LLM outputs are inherently stochastic. Treating these predictions as deterministic leads to unstable rankings, sensitivity to prompt phrasing, and disordered top-k results—especially for borderline documents.

Rutgers researchers have developed REALM: Recursive Bayesian Uncertainty-Aware LLM Re-ranking, a probabilistic LLM-based re-ranking framework that models document relevance as Gaussian distributions and refines them through recursive Bayesian updates. REALM transforms LLM-based document ranking by modeling relevance as probabilistic Gaussian distributions and refining them through recursive Bayesian updates rather than relying on fixed point estimates. This uncertainty-aware approach explicitly measures confidence, stabilizes top-k results, and improves ranking accuracy while significantly reducing LLM inference cost and latency. By selectively querying only uncertain documents and avoiding redundant comparisons, REALM minimizes token usage and enables scalable, cost-efficient deployment. Its combination of robustness, efficiency, and probabilistic reasoning makes it highly significant for production search engines, enterprise retrieval systems, recommendation pipelines, and large-scale AI-powered information platforms. This framework has already been integrated into https://ipapers.ai as the final stage reranker for academic literature search, where it is actively used to enhance the relevance and stability of retrieval results.

Market Applications:

• Enterprise search platforms
• E-commerce product search
• AI chatbots and Retrieval-Augmented Generation (RAG) systems

Advantages:

• Improves ranking stability by explicitly modeling uncertainty.
• Reduces LLM inference cost through adaptive querying.
• Enhances top-k accuracy with recursive Bayesian refinement.

Publications:

•    Wang, Pinhuan, et al. "REALM: Recursive Relevance Modeling for LLM-based Document Re-Ranking." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. https://arxiv.org/abs/2508.18379.

Intellectual Property & Development Status: Provisional application filed. Patent pending. Available for licensing and/or research collaboration. For any business development and other collaborative partnerships, contact:  marketingbd@research.rutgers.edu