IRCoT: Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions
- Paper: Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions
- Authors / Affiliations: Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, Ashish Sabharwal
- Published: 2022-12-20
- Source: https://arxiv.org/abs/2212.10509
- Type: Academic paper
- Category: FreshLLM and similar architectures (LLM and large-scale search)
- Retrieval quality improves by up to 21 points.
- Downstream QA improves by up to 15 points.
- Evaluated on four multi-step QA datasets: HotpotQA, 2WikiMultihopQA, MuSiQue, and IIRC.
- Similar substantial gains hold in out-of-distribution (OOD) settings and with much smaller models such as Flan-T5-large, without additional training.
- IRCoT reduces hallucination, producing more factually accurate CoT reasoning.
- Enhancing Noise Robustness of Retrieval-Augmented Language Models
- FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation
- Gorilla: Large Language Model Connected with Massive APIs
- Long-form factuality in large language models
- ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning
- Search-R1: Training LLMs to Reason and Leverage Search Engines with RL
- Paper: https://arxiv.org/abs/2212.10509
- Code, data, and prompts: https://github.com/stonybrooknlp/ircot
One-line Summary
IRCoT interleaves retrieval with chain-of-thought reasoning steps, so that retrieval guides the reasoning and the reasoning guides what to retrieve next.
Background and Motivation
Prompting-based large language models (LLMs) are surprisingly powerful at generating natural language reasoning steps — Chains-of-Thought (CoT) — for multi-step question answering (QA). However, they struggle when the necessary knowledge is either unavailable to the LLM or not up-to-date within its parameters. Retrieving relevant text using the question helps, but this one-step retrieve-and-read approach is insufficient for multi-step QA: what to retrieve depends on what has already been derived, which in turn may depend on what was previously retrieved.
The IRCoT Approach
IRCoT interleaves retrieval with steps (sentences) in a CoT:
1. CoT-guided retrieval: Reasoning sentences generated so far are used to form retrieval queries for an external knowledge source. 2. Retrieval-augmented CoT: Retrieved results are fed back into the model to improve subsequent reasoning steps. 3. Iterative loop: This alternation continues across the reasoning chain, aligning each retrieval with the current reasoning state.
Conceptually, this replaces the classic "retrieve once, then read" pipeline with an iterative process in which the retrieval query evolves as the derivation progresses — an early step toward the later agentic search / ReAct-style paradigms.
Experimental Results
Using IRCoT with GPT-3:
Key Takeaways for Search / RAG Practitioners
1. Architecture: Interleaved retrieval + generation remains a strong baseline; later work (ReSearch, Search-R1) makes the number and policy of retrieval steps itself learnable via reinforcement learning. 2. Engineering: The iterative loop adds latency and token cost per query; cascades with early stopping and query caching are common mitigations. 3. Evaluation: Beyond answer accuracy, check the factual accuracy of intermediate reasoning steps, since IRCoT's main qualitative benefit is reduced hallucination in the chain.
Limitations and Future Directions
Potential limitations include benchmark scale, mismatch between benchmarks and real user query distributions, and the safety risks of open retrieval. Follow-up directions include more efficient test-time compute allocation, deeper integration with knowledge graphs and structured databases, and retrieval policies learned via RL.