Improving Context Fidelity via Native Retrieval-Augmented Reasoning
Suyuchen Wang, Jinlin Wang, Xinyu Wang, Shiqi Li, Xiangru Tang, Sirui Hong, Xiao-Wen Chang, Chenglin Wu, Bang Liu
2025-09-18
Summary
This paper addresses a common issue with large language models: they sometimes give answers that don't quite match the information they were given, leading to inconsistencies.
What's the problem?
Large language models, while powerful, often struggle to consistently use the specific information provided to them when answering questions. Current solutions either require a lot of extra work to get the model to show its sources *after* answering, or they focus on having the model search the internet, which doesn't necessarily help it use the information you *already* gave it.
What's the solution?
The researchers developed a new system called CARE that helps language models learn to actively use the provided information *while* they are thinking through the answer. It's like teaching the model to highlight the relevant parts of the text as it reasons. This method doesn't need a huge amount of example data and improves both how well the model finds the right information and how accurately it answers the question, by strategically including key pieces of the provided text in its thought process.
Why it matters?
This work is important because it makes large language models more trustworthy and useful for tasks that require accurate knowledge. By improving their ability to stick to the facts and use provided context, it moves us closer to having AI systems that can reliably answer questions and solve problems based on specific information.
Abstract
Large language models (LLMs) often struggle with context fidelity, producing inconsistent answers when responding to questions based on provided information. Existing approaches either rely on expensive supervised fine-tuning to generate evidence post-answer or train models to perform web searches without necessarily improving utilization of the given context. We propose CARE, a novel native retrieval-augmented reasoning framework that teaches LLMs to explicitly integrate in-context evidence within their reasoning process with the model's own retrieval capabilities. Our method requires limited labeled evidence data while significantly enhancing both retrieval accuracy and answer generation performance through strategically retrieved in-context tokens in the reasoning chain. Extensive experiments on multiple real-world and counterfactual QA benchmarks demonstrate that our approach substantially outperforms supervised fine-tuning, traditional retrieval-augmented generation methods, and external retrieval solutions. This work represents a fundamental advancement in making LLMs more accurate, reliable, and efficient for knowledge-intensive tasks.