iRAT: Replanning and Controlled Retrieval for Robust LLM Reasoning

Large Language Models (LLMs) have demonstrated significant capabilities in answering questions using techniques such as Chain of Thought (CoT) and Retrieval-Augmented Generation (RAG). CoT enables step-by-step reasoning to improve accuracy, while RAG supplements LLMs with relevant external information. Retrieval-Augmented Thoughts (RAT) combines CoT and RAG to provide a more robust factual foundation and coherence in reasoning chains. However, RAT is limited in its ability to handle uncertainty and lacks replanning, often resulting in unnecessary retrievals, inefficiencies, and globally inconsistent reasoning. To address these limitations, we introduce iRAT, a novel reasoning framework that enhances RAT through retrieval control and replanning. iRAT dynamically evaluates uncertainty in initial responses, employs controlled and filtered retrievals to obtain only the most relevant context, revises thoughts to align with new content, and uses replanning to correct previous thoughts. Evaluations demonstrated that iRAT outperforms RAT in HumanEval, MBPP, and GSM8K datasets, while reducing retrievals by a considerable amount. The source code is available at github.com/prane-eth/iRAT. The fine-tuned model used for replanning is available at huggingface.co/zeeshan5k/iRATReasoningChainEvaluatorv2.