JCT at SemEval-2026 Task 8: Resource-Efficient Multi-Turn RAG via Nano-LLM Rewriting and Hybrid Reranking

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

JCT's system submission for SemEval-2026 Task A (MTRAGEval) addresses multi-turn Retrieval-Augmented Generation (RAG) with a focus on resource efficiency. It proposes a pipeline that decouples query understanding from retrieval, utilizing a 1.5B parameter Nano-LLM (Qwen) for query rewriting. This is followed by parallel hybrid retrieval via Qdrant and Cross-Encoder reranking. The optimized system achieved an nDCG@5 score of 0.1991 on answerable queries during internal development, surpassing the official BM25 baseline. On the official blind test set, it scored 0.1744. While its absolute performance trails 20B parameter models, the work establishes a crucial baseline for extreme resource efficiency in conversational RAG. The team also provided an error analysis on domain shifts and retrieval funnels, alongside a qualitative analysis of "Underspecified" queries, highlighting generative query rewriting's vulnerabilities.

Key takeaway

For Machine Learning Engineers developing conversational RAG systems, this work demonstrates that resource efficiency is achievable without massive models. If your project has strict compute or memory constraints, consider implementing a decoupled pipeline using a Nano-LLM for query rewriting and hybrid retrieval. Be aware that generative rewriting can struggle with underspecified queries, necessitating robust error analysis and potentially alternative strategies for such cases.

Key insights

Resource-efficient multi-turn RAG can be achieved by decoupling query understanding and using a Nano-LLM for rewriting.

Principles

• Decouple query understanding from retrieval for efficiency.
• Nano-LLMs establish resource-efficient RAG baselines.
• Generative query rewriting has vulnerabilities.

Method

The pipeline involves a 1.5B Qwen Nano-LLM for query rewriting, followed by parallel hybrid retrieval (Qdrant) and Cross-Encoder reranking to address conversational query ambiguity.

In practice

• Utilize 1.5B parameter Nano-LLMs for query rewriting.
• Implement parallel hybrid retrieval with Qdrant.
• Analyze generative rewriting for "Underspecified" queries.

Topics

• Multi-turn RAG
• Nano-LLM
• Query Rewriting
• Hybrid Retrieval
• SemEval-2026
• Resource Efficiency
• Qwen

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

Related on AIssential

• Issue #115 - Reranking in your RAG pipeline — Reranking enhances RAG pipeline precision by reordering initially retrieved documents using a Cross-Encoder.
• Long Context Is Not Replacing RAG - It Is Forcing Us to Rethink Context Engineering — Long context models enhance, rather than replace, RAG, demanding advanced context engineering.
• Beyond Semantic Search: The Complete Guide to Advanced RAG with Milvus | Sagar Patil — Advanced RAG requires multi-stage retrieval, combining hybrid search with techniques like query rewriting and reranking for production reliability.
• Benchmarking Retrieval Strategies for Biomedical Retrieval-Augmented Generation: A Controlled Empirical Study — Cross-Encoder Reranking significantly improves biomedical RAG performance by enhancing contextual precision.
• Rethinking the Necessity of Adaptive Retrieval-Augmented Generation through the Lens of Adaptive Listwise Ranking — Adaptive retrieval's role shifts from noise filtering for weaker LLMs to efficiency optimization for stronger ones.
• LLM-Powered Deep Parsing for Industrial Inventory Search — LLM-powered deep parsing extracts structured data from messy industrial text for improved search and deduplication.

Open in AIssential →

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