Title: Promoting Faithful Reasoning in Retrieval-Augmented Generation: A Step Beyond Accuracy

Abstract:

Following the triumph of reinforcement learning (RL) in training Large Language Models (LLMs) for specialized fields such as mathematics and software engineering, researchers are increasingly focusing on equipping LLMs to dynamically plan, query, and reason using search engines as external tools. This emerging approach is widely known as agentic search. While these methods have demonstrated performance gains on standard short-form question-answering benchmarks, they often focus heavily on the accuracy of the final answer, neglecting the integrity of intermediate reasoning steps. This oversight can result in "chain-of-thought unfaithfulness," where the reasoning path does not logically support the conclusion.

In this study, we propose a comprehensive evaluation framework for agentic search that assesses faithfulness across three specific dimensions: Think-Search faithfulness, Information-Think faithfulness, and Think-Answer faithfulness. Our analysis indicates that standard agentic search systems, such as Search-R1 and ReSearch, which are trained via Reinforcement Learning from Verifiable Reward (RLVR) using episode-level outcome-based rewards, exhibit substantial deficiencies in these faithfulness metrics.

To address this issue and encourage more faithful reasoning, we present VERITAS (Verifying Entailed Reasoning through Intermediate Traceability in Agentic Search). This novel framework incorporates fine-grained, turn-level faithfulness rewards directly into the reinforcement learning training process. Experimental results demonstrate that models trained using \ours not only achieve markedly higher reasoning faithfulness but also outperform baseline models that rely solely on episode-level outcome-based rewards in terms of overall task performance.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC