ReNAct: Augmenting LLM Reasoning with Dynamic Notes Writing for Complex QA

ACL ARR 2025 February Submission7758 Authors

Abstract: Retrieval-augmented generation (RAG) for complex multihop QA tasks struggle with long contexts — retrieved information accumulated across steps, leading to redundancy and inefficient queries at later reasoning steps that degrade retrieval and thus the overall response quality. We introduce ReNAct (Reason, iterative Notes Writing, and Action) which enhances long context based multi-hop QA reasoning by iteratively retrieving and accumulating only the most relevant information at each reasoning step. Notes Writing dynamically maintains a concise list of the relevant information which enables more effective query writing at each reasoning step allowing focus on missing information rather than reprocessing previously retrieved long content. By writing concise notes and guiding query formulation, ReNAct significantly improves both effectiveness and efficiency in multi-hop reasoning. Our approach achieves >20% absolute F1 score gains on long-context benchmarks such as FanOutQA and FRAMES, while reducing the number of reasoning steps by 56% on average compared to the ReAct + BM25 baseline.

Paper Type: Short

Research Area: Question Answering

Research Area Keywords: MultiHop Question Answering, Retrieval-Augmented Generation, RAG, QA

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 7758