What's this about?
Probing the epistemic myopia of LLMs under long-tail divergent knowledge — what happens when the sources a model learned from disagree, and each covers only part of the truth.
What is "epistemic myopia"?
Confidently answering from the fragment of evidence you happen to hold, without registering that other fragments exist. The parable is exact: each blind man is not lying, and each is describing something real, and each is wrong about the elephant. A model trained on divergent long-tail sources has exactly this problem — the failure is not hallucination, because everything it says traces to something real, but a failure to represent that its view is partial.
Why does this matter to you?
- It is a different and harder failure than hallucination, and standard factuality evaluations do not detect it.
- Long-tail divergent knowledge is precisely the regime for niche technical questions — including security advisories, where sources routinely disagree about severity and applicability.
- The right response is a model that reports the disagreement rather than resolving it silently, which almost none currently do.