Foresight-over-Graph: Reasoning Beyond Local Horizons for Knowledge Base Question Answering

2026-10-06

AI Summary

This study addresses the issue of premature loss of critical evidence in knowledge graph reasoning with large language models (LLMs) caused by local greedy strategies. To this end, it proposes the FoG framework, which introduces a lookahead-aware retrieval paradigm to overcome local field-of-view limitations. Specifically, FoG iteratively constructs dynamic evidence subgraphs, employs a far-to-near feedback mechanism to guide path exploration, and incorporates memory compression techniques to effectively preserve key branches within deep contexts for multi-hop reasoning. Experimental results demonstrate that FoG achieves state-of-the-art performance on KBQA benchmarks, improving the Hits@1 metric on ComplexWebQuestions (CWQ) by 16.58% while significantly reducing both LLM invocation frequency and token consumption.

Abstract

Large language models (LLMs) have demonstrated strong capabilities in question answering, yet they still frequently suffer from hallucinations on knowledge-intensive tasks. Knowledge graphs (KGs) provide LLMs with structured, interpretable, and updatable factual grounding, making them a promising external knowledge source for reliable reasoning. However, existing LLM-guided graph reasoning methods typically rely on hop-wise greedy or beam-style pruning during evidence retrieval. Such local decision processes are inherently myopic: evidence that appears weak near the source may become crucial only after deeper graph context is explored, causing answer-critical branches to be discarded prematurely and making the reasoning chain difficult to recover. To address this limitation, we propose Foresight-over-Graph (FoG), a foresight-aware evidence retrieval framework for knowledge base question answering (KBQA). FoG iteratively constructs a question-relevant evidence subgraph and uses far-to-near feedback to guide path exploration, and maintains a compact memory subgraph to support continued exploration. Extensive experiments on widely used KBQA benchmarks demonstrate that FoG achieves state-of-the-art performance, with a particularly large improvement of 16.58% in Hit on CWQ, while also reducing LLM calls and token usage. Our code is available at https://github.com/yhong7/FoG .

Problem

• Knowledge Base Question Answering
• Graph Reasoning
• Evidence Retrieval
• Large Language Models
• Myopic Pruning

Innovation

• Foresight-over-Graph
• Knowledge Base Question Answering
• Far-to-Near Feedback
• Evidence Subgraph
• Large Language Models