LongFaith: Enhancing Long-Context Reasoning in LLMs with Faithful Synthetic Data

TL;DR

LongFaith introduces a pipeline for creating faithful synthetic datasets that improve long-context reasoning in LLMs by integrating verified reasoning and citations, leading to enhanced performance and reliability.

Contribution

The paper presents a novel data synthesis pipeline that produces faithful long-context reasoning datasets, addressing verification and attribution issues in synthetic data for LLMs.

Findings

Models fine-tuned on LongFaith datasets outperform baselines in reasoning tasks.

The datasets improve model accuracy on multi-hop reasoning benchmarks.

The pipeline is scalable and adaptable across different tasks and models.

Abstract

Despite the growing development of long-context large language models (LLMs), data-centric approaches relying on synthetic data have been hindered by issues related to faithfulness, which limit their effectiveness in enhancing model performance on tasks such as long-context reasoning and question answering (QA). These challenges are often exacerbated by misinformation caused by lack of verification, reasoning without attribution, and potential knowledge conflicts. We propose LongFaith, a novel pipeline for synthesizing faithful long-context reasoning instruction datasets. By integrating ground truth and citation-based reasoning prompts, we eliminate distractions and improve the accuracy of reasoning chains, thus mitigating the need for costly verification processes. We open-source two synthesized datasets, LongFaith-SFT and LongFaith-PO, which systematically address multiple dimensions of faithfulness, including verified reasoning, attribution, and contextual grounding. Extensive experiments on multi-hop reasoning datasets and LongBench demonstrate that models fine-tuned on these datasets significantly improve performance. Our ablation studies highlight the scalability and adaptability of the LongFaith pipeline, showcasing its broad applicability in developing long-context LLMs.