Knowledge-Augmented Planning and Pruning for Multi-Hop Question Answering

While large language models (LLMs) excel at language understanding and generation, their application to multi-hop knowledge graph question answering (KGQA) faces challenges from structural and semantic mismatches. Current approaches primarily adopt an agent-based paradigm involving planning (decomposing queries), exploration (searching subgraphs), and reflection. However, previous methods suffer from two limitations: For planning, agent-based paradigms heavily rely on LLMs to decompose questions without continuous grounding in KG structures, leading to unanchored subgoal hallucinationsfactually inconsistent reasoning steps; For KG exploration, previous retrieval strategies prioritise isolated entity matching over contextualised relational triplets, resulting in semantically fragmented retrieval that propagates redundant subgraph exploration and semantic drift. To address these challenges, this work proposes a novel KAPP (Knowledge-Augmented Planning and Pruning) framework that synergizes dynamic knowledge grounding with semantic-aware subgraph optimisation. Specifically, KAPP introduces two innovations: a dynamic knowledge-guided planner that interleaves LLM-driven subgoal generation with KG-validated path correction, eliminating hallucinated steps through iterative verification, and a triple-level semantic pruner that evaluates relational relevance via joint question-triplet encoding, preserving contextual coherence while reducing redundant retrievals. Extensive experiments on CWQ and WebQSP benchmarks demonstrate KAPP’s effectiveness, achieving 5.6-7% accuracy gains over existing KG-augmented methods and reducing LLM API calls by $\mathbf{3 2. 4 \%}$.