StepGap: A Hybrid NLI-LLM Checker for Step-Level Evidence-Gap Detectionin Multi-Hop Question Answering
Researchers have developed StepGap, a novel hybrid system that combines Natural Language Inference (NLI) models with Large Language Models (LLMs) to identify evidence gaps in multi-hop question answering. This system categorizes these gaps into three types: Contradicted Claim, Irrelevant Evidence, and Missing Bridge, each suggesting a specific repair action. While StepGap's overall F1 score is comparable to LLM-only baselines, its structured approach offers greater interpretability and avoids error cancellation issues seen in purely LLM-based methods. When used to guide reinforcement learning, StepGap significantly improved the Exact Match score of the Qwen2.5-7B-Instruct model. AI
IMPACT This hybrid approach offers a more interpretable and robust method for improving multi-hop QA systems, potentially leading to more reliable AI assistants.