Technology Category
Knowledge Representation and Reasoning: Qualitative ReasoningNatural Language Processing: Question AnsweringSearch and Optimization: Metareasoning and Metaheuristics
Self-Critique Guided Iterative Reasoning for Multi-hop Question Answering
🤖 AI Summary
To address retrieval bias and reasoning errors in knowledge-intensive multi-hop question answering—caused by the absence of intermediate supervision—this paper proposes the Self-Critical Iterative Reasoning (SCIR) framework. SCIR enables end-to-end training across three synergistic stages: automatic question decomposition, dynamic self-assessment of intermediate reasoning steps, and multi-branch trajectory exploration. Its core innovation is a novel large language model–based self-critical feedback mechanism that generates fine-grained intermediate evaluation signals without human annotation, thereby guiding iterative refinement of reasoning paths. Evaluated on HotpotQA, 2WikiMultiHopQA, and MuSiQue, SCIR outperforms state-of-the-art methods by up to 8.6%, and—critically—provides the first empirical validation of a positive coupling between controllability of intermediate reasoning and final answer accuracy.
Technology Category
Application Category
📝 Abstract
Although large language models (LLMs) have demonstrated remarkable reasoning capabilities, they still face challenges in knowledge-intensive multi-hop reasoning. Recent work explores iterative retrieval to address complex problems. However, the lack of intermediate guidance often results in inaccurate retrieval and flawed intermediate reasoning, leading to incorrect reasoning. To address these, we propose Self-Critique Guided Iterative Reasoning (SiGIR), which uses self-critique feedback to guide the iterative reasoning process. Specifically, through end-to-end training, we enable the model to iteratively address complex problems via question decomposition. Additionally, the model is able to self-evaluate its intermediate reasoning steps. During iterative reasoning, the model engages in branching exploration and employs self-evaluation to guide the selection of promising reasoning trajectories. Extensive experiments on three multi-hop reasoning datasets demonstrate the effectiveness of our proposed method, surpassing the previous SOTA by $8.6%$. Furthermore, our thorough analysis offers insights for future research. Our code, data, and models are available at Github: https://github.com/zchuz/SiGIR-MHQA.
Problem
Research questions and friction points this paper is trying to address.
Addressing knowledge-intensive multi-hop reasoning challenges in LLMs
Improving accuracy of iterative retrieval and intermediate reasoning steps
Guiding reasoning process via self-critique feedback and branching exploration
Innovation
Methods, ideas, or system contributions that make the work stand out.
Self-critique feedback guides iterative reasoning
End-to-end training enables question decomposition
Branching exploration with self-evaluation selection
🔎 Similar Papers
No similar papers found.
