Calibrated Multi-Agent Consensus for Knowledge Graph Validation using Open-Source LLM

Abstract

Knowledge graphs are being used with Large Language Model applications such as question answering, recommendation, and factual grounding, but real graphs contain errors that propagate into downstream tasks. Manually finding and resolving these errors takes time and can lead to human errors. Automated validation of knowledge graph triples is, therefore, a long-standing problem. Recent works used commercial Large Language Models for the validation of Knowledge Graphs, and showed strong results, but have struggled with open-source alternatives. The confidence values that open-source models report are systematically unreliable, for example, a model that predicts a triple is true with 90% confidence is correct on those predictions far less often than 90% of the time.

This thesis investigates whether post-hoc calibration combined with dual-perspective evidence fusion can produce reliable triple validation using a single open-source language model. The proposed framework has a plug-and-play architecture containing two agents, an External Agent that retrieves factual evidence from an external source and an Internal Agent that extracts subgraphs from the knowledge graph’s own training data, both running on Meta-Llama-3-8B-Instruct. Each agent’s raw logits are passed through a calibration layer that evaluates on post-hoc methods (temperature, Platt, isotonic) and a consensus engine that combines the calibrated probabilities through fusion strategies. The framework is evaluated on three knowledge graphs spanning distinct domains.

The results overall indicated that the framework is potentially effective across all three datasets, and calibrated confidence estimates can contribute to more reliable performance and improved accuracy across various knowledge graph types.