ResRank: End-to-End Joint Training for Unified Retrieval and Listwise Reranking
A new framework called ResRank has been introduced by researchers to tackle two significant challenges in listwise reranking using LLMs: the 'lost in the middle' issue and super-linear inference latency. Drawing inspiration from multimodal LLMs, ResRank employs an Encoder-LLM to condense each candidate passage into a singular embedding. This embedding, along with the query, is processed by a Reranker-LLM for listwise ranking. This method not only shortens input length but also lessens ranking degradation. Designed for industrial use, the framework enhances efficiency while maintaining effectiveness. The paper can be found on arXiv with the identifier 2604.22180.
Key facts
- ResRank unifies retrieval and listwise reranking via end-to-end joint training.
- It uses an Encoder-LLM to compress each passage into a single embedding.
- The Reranker-LLM performs listwise ranking using query and compressed embeddings.
- Addresses 'lost in the middle' phenomenon in long input sequences.
- Reduces inference latency that scales super-linearly with sequence length.
- Inspired by multimodal LLMs that project visual inputs into compact tokens.
- Aims to make LLM-based reranking practical for industrial deployment.
- Paper published on arXiv with ID 2604.22180.
Entities
Institutions
- arXiv