Sifei at SemEval-2026 Task 8: Hybrid Retrieval and Query Rewriting for Multi-Turn RAG

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, medium

Summary

Sifei Meng and Dmitry Ilvovsky presented a training-free hybrid retrieval pipeline for multi-turn Retrieval-Augmented Generation (RAG) at SemEval-2026 Task 8. This system addresses challenges like evolving user intent and context limits by integrating dense and sparse retrieval, controlled query rewriting, and cross-encoder reranking. On Task A, the pipeline achieved a normalized Discounted Cumulative Gain at 5 (nDCG@5) of 0.5453, securing 3rd place among 38 teams and surpassing the strongest baseline of 0.4795. For Task C, the system utilized documents retrieved from Task A within a lightweight generation pipeline, resulting in a harmonic mean of quality and faithfulness score of 0.5312 and a 15th place ranking out of 29 teams. The retrieval components are open-source, while query rewriting and generation use LLM APIs.

Key takeaway

For Machine Learning Engineers developing multi-turn RAG systems, this research demonstrates a robust, training-free hybrid retrieval approach. You should consider integrating dense and sparse retrieval with LLM-powered query rewriting and cross-encoder reranking to enhance performance in conversational contexts. This strategy can significantly improve document relevance (0.5453 nDCG@5) and generation quality, offering a competitive solution without extensive model training. Explore the open-source retrieval components for rapid deployment.

Key insights

The paper proposes a training-free hybrid RAG pipeline combining diverse retrieval and rewriting for multi-turn conversational contexts.

Principles

• Hybrid retrieval improves multi-turn RAG performance.
• Query rewriting is crucial for evolving user intent.
• Cross-encoder reranking refines retrieval results.

Method

The method involves a training-free pipeline: dense and sparse retrieval, controlled query rewriting, and cross-encoder reranking, followed by a lightweight generation using LLM APIs.

In practice

• Combine dense and sparse retrieval for RAG.
• Implement query rewriting for conversational AI.
• Use open-source retrieval with LLM APIs.

Topics

• Multi-turn RAG
• Hybrid Retrieval
• Query Rewriting
• Cross-encoder Reranking
• SemEval-2026 Task 8
• Large Language Models

Code references

• mengsifei/MultiturnRAG

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Benchmarking Retrieval Strategies for Biomedical Retrieval-Augmented Generation: A Controlled Empirical Study — Cross-Encoder Reranking significantly improves biomedical RAG performance by enhancing contextual precision.
• Hybrid Search and Re-Ranking in Production RAG — Combining hybrid search, cross-encoder re-ranking, and metadata filtering significantly improves RAG system accuracy for enterprise knowledge.
• Rerankers Aren’t Magic Either: When the Cross-Encoder Layer Is Worth the Cost — Cross-encoder rerankers offer inconsistent performance gains over strong embeddings, often failing on complex query types.
• RCEM: Robust Conversational Search EMbedder in Distributional Shift — RCEM distills LLM query rewriting into an embedder for robust, low-latency conversational search without explicit rewriting.
• Uncertainty-Aware Hybrid Retrieval for Long-Document RAG — UMG-RAG adaptively fuses multi-granularity, multi-expert retrieval based on query-specific uncertainty to create compact, coherent RAG contexts.
• An Efficient Approach for Answering Not Readily Attainable Questions for RAG-based Applications — This paper proposes an efficient approach for RAG systems to answer questions not readily found in retrieved data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.