@inproceedings{farhan-liebeskind-2026-jct,
    title = "{JCT} at {S}em{E}val-2026 Task 8: Resource-Efficient Multi-Turn {RAG} via Nano-{LLM} Rewriting and Hybrid Reranking",
    author = "Farhan, Tal  and
      Liebeskind, Chaya",
    editor = "Kochmar, Ekaterina  and
      Ghosh, Debanjan  and
      North, Kai  and
      Komachi, Mamoru",
    booktitle = "Proceedings of the 20th {I}nternational {W}orkshop on {S}emantic {E}valuation (2026)",
    month = jul,
    year = "2026",
    address = "San Diego, California, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.semeval-1.173/",
    pages = "1326--1331",
    ISBN = "979-8-89176-414-9",
    abstract = "This paper describes our system submission for SemEval-2026 Task A (MTRAGEval), focusing on multi-turn Retrieval-Augmented Generation (RAG). Conversational queries often suffer from contextual ambiguity, rendering standard retrieval methods ineffective. We propose a highly resource-efficient pipeline that decouples query understanding from retrieval using a 1.5B parameter Nano-LLM (Qwen) for query rewriting, followed by parallel hybrid retrieval (Qdrant) and Cross-Encoder reranking. During internal development, our optimized system achieved an nDCG@5 score of 0.1991 on answerable queries, outperforming the official BM25 baseline. On the official blind test set, the system achieved a score of 0.1744. While our absolute performance trails behind baselines utilizing massive 20B parameter models, our work establishes a crucial baseline for extreme resource efficiency in conversational RAG. We provide a comprehensive error analysis detailing the impact of domain shifts, retrieval funnels, and we conduct a qualitative analysis on the organizers' surprise ``Underspecified'' class to highlight the vulnerabilities of generative query rewriting."
}