Works · Framework Demo

Ca2KG · When to Trust KG-RAG

Knowledge-graph RAG answers look confident even when the retrieved path is weak or the reasoning is fragile. Ca2KG calibrates trust by re-answering the question under counterfactual interventions — assuming the retrieved path is weak, then simulating a reasoning failure — and combines the results with a panel-based re-scoring step into a single Calibrated Confidence Index (CCI).

WWW 2026 20.1% acceptance rate MetaQA · WebQSP · ECE / Brier

How it works

2 + 1 interventions, one trust score

01

Baseline answer

The question is answered normally over the retrieved knowledge-graph path.

02

Path-quality intervention

Re-answers while assuming the retrieved path is weak — tests reliance on retrieval quality.

03

Reasoning-reliability intervention

Simulates a reasoning failure to see whether the answer holds up regardless.

04

Panel re-scoring → CCI

All three outputs are aggregated into one Calibrated Confidence Index that decides trust.

Worked example

Three questions, three trust outcomes

Illustrative scores in the shape of the paper's calibration signals — not a live model call.

Question
Baseline confidence
Path-quality intervention
Reasoning-reliability intervention

Calibrated Confidence Index (CCI)
Threshold marker at 70 — panel-aggregated trust score