The Year in AI Papers
The year/Independent research

Paper 2602.12192

Query-focused and Memory-aware Reranker for Long Context Processing

Yuqing Li, Jiangnan Li, Mo Yu, Guoxuan Ding, Yanyu Chen, Zheng Lin, Wei Zhang, Jie Zhou

Published
Feb 2026
Research lab
Independent
Citations
5
GitHub
Not linked

01 In brief

Summary

The paper introduces QRRanker, a listwise reranking framework that trains LLMs to score passage-query relevance using attention scores from selected Query-focused Retrieval (QR) heads, avoiding generation-based scoring and Likert-scale supervision.

It uses a contrastive loss on continuous attention-derived scores, enabling training on arbitrary retrieval datasets.

QRRanker, built on a 3B-parameter model, outperforms existing pointwise and listwise rerankers (e.g., Qwen-Reranker, ReasonRank) and memory methods (e.g., HippoRAG-v2) across Wikipedia multi-hop QA (MuSiQue, HotpotQA), long-context story QA (NarrativeQA, DetectiveQA), and dialogue memory (LoCoMo), setting new state-of-the-art results.

It also shows competitive performance against retrieval agents and on the BRIGHT benchmark.

The framework supports flexible extensions: prepending summary prefixes improves ranking on long narratives and dialogues, and training heads from middle layers maintains performance while reducing inference latency.

QRRanker is robust to input shuffling, indicating content-based scoring, and generalizes across backbones (Qwen3-4B, Llama-3.2-3B).

02 From the paper

Abstract

Built upon the existing analysis of retrieval heads in large language models, we propose an alternative reranking framework that trains models to estimate passage-query relevance using the attention scores of selected heads. This approach provides a listwise solution that leverages the holistic information within the entire candidate shortlist during ranking. At the same time, it naturally produces continuous relevance scores, enabling training on arbitrary retrieval datasets without requiring Likert-scale supervision. Our framework is lightweight and effective, requiring only small-scale models, such as 3B parameters, to achieve strong performance. Extensive experiments demonstrate that our method outperforms existing state-of-the-art pointwise and listwise rerankers across multiple domains, including Wikipedia and long narrative datasets. It further establishes a new state-of-the-art on the LoCoMo benchmark, which assesses dialogue understanding and memory usage. We further demonstrate that our framework supports flexible extensions. For example, augmenting candidate passages with contextual information further improves ranking accuracy, while training attention heads from middle layers enhances efficiency without sacrificing performance.