MXBAI Rerank Large V2

mxbai-rerank-large-v2 is a reranking model that improves the relevance ordering of documents in response to a query, designed for use in retrieval-augmented generation and information retrieval pipelines. It is a specialized small language model (SLM) that achieves state-of-the-art reranking performance while maintaining computational efficiency.

Key Strengths

The model addresses fundamental limitations of SLMs in reranking: narrow representation space and difficulty understanding task prompts without fine-tuning. It uses a two-stage training approach called ProRank, which applies reinforcement learning to improve prompt understanding and fine-grained score learning to enhance representation expressiveness. This allows the model to outperform both advanced open-source and proprietary reranking models, including large LLMs with over 7B parameters.

Benchmark Results

On the BEIR benchmark, the 0.5B parameter version of this model surpasses powerful LLM reranking models, demonstrating that properly trained SLMs can achieve superior document reranking performance while maintaining computational efficiency. The model consistently outperforms the most advanced open-source and proprietary reranking models across extensive experiments.

Architecture and Training

The model uses a two-stage training approach called ProRank. The first stage applies reinforcement learning to improve understanding of task prompts. The second stage introduces fine-grained score learning to enhance representation expressiveness, addressing the narrow representation space and prompt understanding limitations typical of SLMs.