Path Fidelity Evaluation for Small Language Models in KGQA
September 24, 2026
This research utilizes the THESEUS framework to decouple navigation, reasoning, and generation in Knowledge Graph Question Answering (KGQA). By evaluating frozen SLMs as local action policies, the study measures Hits@1 alongside Path Edit Distance to isolate navigation failures from reasoning errors.
HOW THIS AFFECTS YOU
●
builderThis methodology helps in selecting smaller, more efficient models specifically for graph navigation tasks.
●
researcherYou can use this framework to more accurately attribute SLM failures in complex knowledge graph tasks.