LLM Recall, Recognition Split Draws Interest

The useful research trend is splitting factuality into recall, precision, coverage, and self-awareness instead of collapsing everything into one truthfulness score. FACT-BENCH finds instruction tuning can hurt factual recall, while scaling helps and counterfactual exemplars can sharply degrade known facts. FactBench and VERIFY treat verification as supported, unsupported, or undecidable, which is closer to real fact-checking than a simple yes/no. VeriFact and FactRBench show precision and recall can diverge in long-form answers, so a model can reject bad claims without covering every required fact. Self-Alignment for Factuality and Factual Self-Awareness suggest models can sometimes judge their own correctness, but that signal is imperfect and not the same as verbatim recall. If exact wording matters, retrieval-grounded verification is still safer than trusting parametric memory alone.