S-Path-RAG: Semantic-Aware Shortest-Path Retrieval Augmented Generation for Multi-Hop Knowledge Graph Question Answering

A research team led by Rong Fu has introduced S-Path-RAG, a novel framework designed to tackle the complex challenge of multi-hop question answering over large knowledge graphs. Unlike traditional one-shot retrieval methods that rely heavily on text, S-Path-RAG employs a sophisticated hybrid strategy combining weighted k-shortest paths, beam search, and constrained random walks to enumerate semantically relevant candidate paths of bounded length. This approach is both token-efficient and topology-aware, learning a differentiable path scorer alongside a contrastive path encoder and a lightweight verifier to filter results.

The system operates within an iterative Neural-Socratic Graph Dialogue loop, where concise diagnostic messages from the language model trigger targeted graph edits or seed expansions, enabling adaptive retrieval when the model expresses uncertainty. This creates a feedback mechanism that refines searches based on the LLM's confidence. The selected path information is then injected into the language model via cross-attention as a compact soft mixture of path latents, preserving interpretable, path-level traces for diagnostics and human intervention.

Validation on standard multi-hop KGQA benchmarks demonstrates that S-Path-RAG delivers consistent improvements in answer accuracy and evidence coverage compared to existing graph- and LLM-based baselines, while also enhancing end-to-end efficiency. The paper includes detailed ablation studies analyzing trade-offs between semantic weighting, verifier filtering, and iterative updates, providing practical deployment recommendations for scenarios with constrained compute and token budgets. This makes the framework particularly relevant for enterprise applications where both accuracy and operational cost are critical.