Enterprise AI Analysis

GAP's effectiveness stems from a robust two-stage training strategy. Initially, a high-quality dataset of 7,000 graph-based planning traces is synthesized from the Multi-Hop Question Answering (MHQA) benchmark using GPT-4o. This dataset focuses on dependency-aware reasoning trajectories, with rigorous filtering to ensure quality, diversity, and appropriate complexity (e.g., excluding overly simplistic tasks or excessively long ones without genuine retrieval difficulty).

The first stage involves Supervised Fine-Tuning (SFT) on this curated dataset, where the Qwen2.5-3B-Instruct model learns to internalize graph-based planning strategies. This establishes a robust cold start. Subsequently, the model undergoes Reinforcement Learning (RL), fine-tuning with a correctness-based reward function. This RL stage optimizes for computational efficiency and reasoning effectiveness, allowing the model to strategically determine when, how, and how broadly to invoke child threads, balancing parallel exploration with context window constraints.

Experimental results across seven question-answering benchmarks demonstrate GAP's significant advantages. It achieves a 0.9% average performance improvement on multi-hop reasoning tasks over state-of-the-art baselines. Crucially, GAP dramatically enhances tool invocation efficiency through intelligent parallelization. For instance, it reduces interaction turns by up to 33.4% on 2WikiMultiHopQA and response length by 24.9% on HotpotQA compared to Search-R1. This translates directly to faster execution times and lower deployment costs.

The method also shows strong generalization capabilities, with learned parallel decomposition patterns transferring effectively to out-of-domain scenarios. This indicates that GAP not only improves accuracy but also makes multi-hop reasoning more practical and cost-effective for real-world applications. The performance-cost trade-off analysis on HotpotQA further confirms GAP's superior balance of accuracy and efficiency.