ArbGraph: Conflict-Aware Evidence Arbitration for Reliable Long-Form RAG

This work was conducted as part of RUC AI Box at Renmin University of China.

The project repository under RUC AI Box is available at RUCAIBox/ArbGraph.

This repository provides an implementation of ArbGraph, a framework for improving the reliability of long-form retrieval-augmented generation (RAG) via pre-generation evidence arbitration.

The code accompanies our paper and is released for research use and partial reproducibility.

---

Overview

ArbGraph addresses a key limitation of long-form RAG systems: handling noisy and contradictory evidence.

Instead of resolving conflicts during generation, ArbGraph performs pre-generation arbitration by:

• decomposing documents into atomic claims,
• modeling support and contradiction relations,
• estimating claim credibility via conflict-aware arbitration,
• generating outputs from a validated evidence set.

---

Pipeline

1. Retrieval
2. Atomic Claim Extraction (atomization.py)
3. Claim Alignment (claim_alignment.py)
4. Evidence Graph Construction (evidence_graph.py)
5. Conflict Arbitration (conflict_arbitration.py)
6. Generation (longform_generation.py)

---

Usage

python run_arbgraph.py

---

Requirements

Install dependencies:

pip install -r requirements.txt

---

Notes

• Default backbone: Qwen3-4B-Instruct
• Retrieval based on Wikipedia
• This is a research prototype and may require GPU for efficient execution