# Knowledge Graph Question Answering

The work introduces XoG (eXplore-over-Graph), a framework for multi-hop question answering over incomplete knowledge graphs that uses type-level entity-relation statistics to identify candidate relations and KG embeddings to retrieve plausible missing entities, with the LLM acting as semantic selector and reasoner inside an iterative planning-exploration-reasoning process; on WebQSP, CWQ, and the Wikidata-based BRINK benchmark, XoG remains competitive on complete KGs and consistently outperforms comparable methods without task-specific KGQA training under incompleteness, with gains persisting across multiple LLM backbones and LLM token consumption reduced by up to 33% versus a closely related planning-based approach.

The work introduces XoG (eXplore-over-Graph), a framework for multi-hop question answering over incomplete knowledge graphs that uses type-level entity-relation statistics to identify candidate relations and KG embeddings to retrieve plausible missing entities, with the LLM acting as semantic selector and reasoner inside an iterative planning-exploration-reasoning process; on WebQSP, CWQ, and the Wikidata-based BRINK benchmark, XoG remains competitive on complete KGs and consistently outperforms comparable methods without task-specific KGQA training under incompleteness, with gains persisting across multiple LLM backbones and LLM token consumption reduced by up to 33% versus a closely related planning-based approach.

Targeting the Logic Drift that appears in LLM outputs during structured knowledge reasoning, this work proposes the Logits-to-Logic framework, whose core modules are logits strengthening and logits filtering, to directly correct the logits produced in the autoregressive generation process; experiments report significantly improved logic consistency in structured knowledge reasoning and state-of-the-art performance on multiple KGQA benchmarks.

Targeting the Logic Drift that appears in LLM outputs during structured knowledge reasoning, this work proposes the Logits-to-Logic framework, whose core modules are logits strengthening and logits filtering, to directly correct the logits produced in the autoregressive generation process; experiments report significantly improved logic consistency in structured knowledge reasoning and state-of-the-art performance on multiple KGQA benchmarks.