When artificial intelligence models produce incorrect or fabricated details, developers usually assume the underlying system simply lacks the required knowledge. The standard industry response to these hallucination errors has been building larger neural networks, incorporating broader pre-training datasets, or setting up complex external retrieval architectures. However, new findings suggest that the necessary information is often already contained within the software, even when the model fails to display it during normal prompt responses.
A study published by researchers from Technion and Google Research reveals that frontier models such as GPT-5 and Gemini-3 hold between 95% and 98% of tested facts within their parameters. The primary barrier to factual output is not an absence of stored information, but a failure of retrieval mechanisms during output generation. By allocating additional processing time at inference, these models can successfully surface up to 65% of the information they initially failed to recall directly.
To help engineers identify these retrieval bottlenecks, the research team proposes moving away from traditional evaluation techniques. Instead of scoring performance based strictly on full answer accuracy for isolated prompts, they advocate for fact-level profiling. This diagnostic approach measures what a model truly holds in its parametric memory compared to what it actively generates.
What it means
These findings indicate that spending extra computational resources while a model thinks through a problem can unlock existing knowledge without requiring external tools. For engineering teams building artificial intelligence applications, relying on extended inference-time compute offers a path toward higher factual precision without needing bigger parameters or complex external lookup databases.
