Reproduce “ArbGraph: Conflict-Aware Evidence Arbitration for Reliable Long-Form Retrieval-Augmented Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qingying Niu, Yuhao Wang, Ruiyang Ren, Bohui Fang, Wayne Xin Zhao
ArbGraph introduces a pre-generation evidence arbitration framework for Retrieval-Augmented Generation (RAG) that enhances factual consistency in long-form text generation by explicitly identifying and resolving conflicts within retrieved evidence. The system achieved superior factual recall (up to 84.9%), information density, and faithfulness while reducing hallucination rates (as low as 33.3%) across two benchmarks and multiple LLM backbones, demonstrating increased robustness to noisy retrieval conditions.
Key insights
- Explicitly resolving factual conflicts and arbitrating evidence credibility prior to text generation is critical for improving factual consistency and reducing hallucinations in long-form RAG.
- Modeling inter-claim dependencies through a conflict-aware evidence graph is crucial for effective evidence validation, enabling structured reasoning over agreement and contradiction.
- An iterative, confidence-gated arbitration process with dynamic scheduling of high-impact conflicts leads to more stable and effective conflict resolution compared to single-step or static approaches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.