SkepticalSam

FactSelfCheck: Can Knowledge Graphs Catch Hallucinated Facts?

FactSelfCheck: Fact-Level Black-Box Hallucination Detection for LLMs

Read the original source

What the paper says

FactSelfCheck converts long-form outputs into fact triples or knowledge graphs, samples multiple responses, and checks consistency across them. It aims to detect hallucinated facts without external databases or model internals.

The Critique

Fact-level detection is the right target, but the method depends on extraction quality. If the system extracts the wrong triples, it can score the wrong facts. It also assumes hallucinations vary across samples. Some false claims are stable because they are memorised, common, or reinforced by the prompt. A consistent hallucination can pass a consistency check.

Why It Matters

Fine-grained hallucination detection would be much more useful than sentence-level warning labels. Users need to know which claim is suspect, not that a paragraph feels risky.

What They Missed

Robustness against stable falsehoods, long-document scaling, human review of extracted triples, and comparisons with retrieval-based verification.

The Big Question

Can a model catch hallucinations by comparing itself to itself, or do consistent falsehoods slip through the net?