When Self-Correction Fails: Diagnosing Retrieval Coverage in Hallucination Mitigation for Large Language Models

We evaluate a claim-level, evidence-grounded self-correction framework for LLM
hallucination and report a mixed result with a clear diagnosis. As a detector,
evidence-grounded verification vastly outperforms LLM self-assessment on HaluEval
(F1 0.583 vs 0.066; the baseline identifies only 3.9% of hallucinated answers).
As a generator, the same machinery significantly harms factuality, losing 9.8 pp
on TruthfulQA and 11.6 pp on FEVER against an unaugmented LLM. Diagnostic analysis
attributes both losses to retrieval coverage rather than verification logic: 93.8%
of TruthfulQA queries retrieve no relevant passage and FEVER gold evidence recall
is 10.8%, so claims are marked unsupported because evidence is absent, not false.
An ablation shows our monotonic acceptance gate itself costs 5.7 pp by locking in
hedged drafts. We argue retrieval coverage statistics should be reported as
standard in the self-correction literature, since without them a coverage failure
is indistinguishable from a method failure.