Large language models absorb enormous amounts of knowledge during training, but that knowledge does not reliably appear in their outputs. Because the models are probabilistic, there is no guarantee that what they have learned will be expressed when prompted. A new arXiv paper, 'MAPLE: A Multi-Axis Parametric Knowledge Extraction Lens for LLM Evaluation,' takes this problem as its starting point.
The abstract points to Knowledge Extraction Enhancement (KEE) techniques as a way to bridge the gap between stored and surfaced knowledge. MAPLE is presented as a lens for evaluating LLMs along multiple axes, suggesting that assessment should look at how well a model can extract and surface what it knows.
The paper's framing implies that evaluation methods need to account for the probabilistic nature of LLMs. If knowledge can remain latent, then standard benchmarks may not capture the full extent of what a model truly knows. MAPLE appears to be an attempt to make that latent knowledge visible for more meaningful evaluation.