Frontier AI Models Can Recover 65% of Forgotten Facts Through Extended Inference Time, Study Finds

When artificial intelligence models generate inaccurate statements or hallucinations, developers typically assume the underlying system lacks the required training data. However, new research indicates that large language models often store the correct information deep within their parameters, failing only to surface it during initial generation.

A joint study conducted by researchers at Google Research and Technion revealed that top-tier artificial intelligence systems hold vastly more information than their immediate responses suggest, as first reported by VentureBeat. The study demonstrated that state-of-the-art models like GPT-5 and Gemini-3 have actually encoded between 95% and 98% of the factual data points tested, pointing to recall mechanisms rather than data storage as the primary bottleneck in model accuracy.

By expanding computation during the inference phase—effectively allowing the system additional processing time to reason through a prompt—the researchers found that models could retrieve up to 65% of the facts they initially failed to produce. This finding suggests that artificial intelligence engineering teams may be able to significantly diminish hallucinations without continually expanding dataset sizes, scaling up parameter counts, or deploying complex external retrieval architectures.

To better evaluate the mechanics behind storage and retrieval, the research team introduced an evaluation framework centered on fact-level profiling rather than traditional question-level metrics. Instead of grading a model based on whether it provides a correct answer to an isolated query, fact-level profiling tracks how a single piece of information is processed across diverse conditions, phrasings, and querying angles.

Under this new testing methodology, the researchers drew a clear boundary between whether a piece of data is parametrically encoded versus whether it is fully known by the model. A fact is classified as encoded if the language model can reproduce it when provided with its original training context. Conversely, a fact is considered known only if the model can consistently recall and apply it across varied sentence structures and prompt directions.

The distinction highlights why standard benchmark metrics often fail to guide appropriate engineering interventions. As the study's authors noted, encoding failures and recall failures are visually indistinguishable under traditional accuracy benchmarks, despite demanding entirely different technical solutions and reflecting separate system limitations.

For enterprise technology leaders and developers building commercial applications, the research provides a pragmatic path toward improving output reliability. Rather than focusing exclusively on scaling training data or expanding vector databases, software teams can achieve substantial accuracy gains by implementing advanced inference-time computing strategies that give existing architectures the necessary processing time to retrieve latent parametric knowledge.