In short. Parametric knowledge is information encoded in model weights, tested through factual recall without external retrieval and through controlled weight edits.

Definition: Parametric knowledge is information encoded in a language model's learned weights, the internal parameters used to produce predictions. Researchers test factual recall by asking questions without external search or retrieval.

Key facts

The March 2026 revision of SimpleQA Verified describes a benchmark of 1,000 prompts evaluated without tools.
The retrieval-augmented generation paper combines a pretrained model with an external index of Wikipedia passages.
The ROME study tests factual storage by changing model weights and evaluating the resulting predictions.

How does SimpleQA Verified test factual recall?

The SimpleQA Verified paper focuses on short factual answers. Its construction removes similar questions, balances topics and answer types, and reconciles conflicting reference sources. These choices affect which facts the evaluation samples and whether its reference answers are reliable.

An automated grader assigns each response to a category: correct, incorrect, or not attempted. The paper reports an F1 score, the harmonic mean of overall accuracy and accuracy among attempted answers. This combines factual accuracy with the model's willingness to answer.

The grading instructions focus on the direct answer and ignore extra information. A response listing several candidates must commit to a preferred answer to count as attempted. The paper's tool-free setting isolates recall from assistance supplied by search.

How do researchers probe individual facts?

Petroni and colleagues' Language Models as Knowledge Bases? introduces the LAMA probe, a collection of tests for factual and commonsense relations. A probe is an input designed to elicit a particular piece of knowledge. Here, the model fills a missing word in a sentence, called a cloze statement.

The paper illustrates the method with this input:

Dante was born in [Mask].

The expected completion is Florence. The researchers operationally treat successful prediction of the missing object as evidence that the model knows the corresponding relation. They test pretrained models without further training for the task and find that some types of factual knowledge are learned more readily than others.

How is stored knowledge different from retrieved information?

Retrieval-augmented generation, or RAG, supplies retrieved documents as additional context for generating an answer. In Lewis and colleagues' RAG architecture, the parametric component is a pretrained sequence-to-sequence model, which maps an input sequence to an output sequence. The external memory is a searchable index of Wikipedia passages.

The paper demonstrates updating the system's knowledge by replacing this external index. That update concerns the documents available during answering. The distinction matters when assessing a website's role: an answer can draw on retrieved text as well as information encoded during training. A tool-free factuality test examines the latter mechanism.

Can researchers change knowledge inside the weights?

The Locating and Editing Factual Associations in GPT study uses causal interventions: researchers corrupt internal activations, then restore selected values to examine their effect on a factual prediction.

Feed-forward modules are network components that transform internal representations. The study finds an important role for these modules in middle layers of the tested models. Its Rank-One Model Editing method, abbreviated ROME, changes feed-forward weights to alter a factual association. Evaluation checks whether the change carries across different phrasings and whether unrelated associations remain intact. This complements answer-based tests with evidence from interventions inside the model.

FAQ

Can I use web search during a parametric knowledge test?

SimpleQA Verified evaluates models without tools. Adding search changes what the test measures by supplying external information.

Can I test factual recall without changing the model?

Yes. Fine-tuning means further training an existing model for a task. The LAMA study tests pretrained models through missing-word predictions without this additional training.

Sources

Check how a page reads to answer engines. Open the GEO score analyzer