ERank: Effective and Efficient Text Reranking

ERank is a series of pointwise rerankers developed by Alibaba-NLP, including 4B, 14B, and 32B parameter versions. These models are built from reasoning LLMs and are designed to provide highly effective and efficient text reranking with low latency. A key innovation is their ability to outperform some listwise rerankers on challenging reasoning-intensive tasks.

Key Capabilities & Features

• Novel Two-Stage Training: ERank is trained using a unique pipeline involving Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). SFT encourages the LLM to generate fine-grained integer scores, moving beyond traditional binary relevance classification. RL incorporates a novel listwise derived reward, instilling global ranking awareness into the pointwise architecture.
• Instruction Awareness: All ERank models support customizing input instructions, allowing adaptation to various tasks.
• Strong Performance: ERank models demonstrate competitive performance across both reasoning-intensive benchmarks (BRIGHT, FollowIR) and traditional semantic relevance benchmarks (BEIR, TREC DL). The 32B version achieves an average score of 38.1 across these benchmarks, with ERank-4B scoring 36.8.
• Low Latency: As a pointwise reranker, ERank offers significantly lower latency compared to listwise models, making it suitable for applications requiring fast response times.

When to Use ERank

• Text Reranking: Ideal for improving the relevance of retrieved documents in search or recommendation systems.
• Reasoning-Intensive Tasks: Particularly strong in scenarios requiring complex reasoning for document ranking.
• Low-Latency Applications: Suitable for environments where quick reranking is critical due to its efficient pointwise architecture.
• Customizable Ranking: When there's a need to adapt the reranking behavior through specific instructions.