GPT-5 and Gemini-3 Recovered Up to 65% of Missed Facts by Thinking Longer

A frontier model’s wrong answer may increasingly be a problem of access, not storage. A Google Research and Technion study found that extra inference-time reasoning recovered 40–65% of encoded Wikipedia facts that models initially failed to produce directly, challenging the reflex to treat every factual miss as evidence that a model needs more data or retrieval.

Stored facts and usable answers are different tests

The researchers tested 13 language models across more than 4 million responses with WikiProfile, a set of 2,150 facts drawn from Wikipedia. Their central distinction is between encoding a fact—successfully completing its original training-style context—and knowing it well enough to answer varied phrasings and directions reliably.

That framing separates two superficially identical failures. An encoding failure means the model does not reproduce the underlying information. A recall failure means it can reproduce the fact in the right context but cannot surface it for a direct question. The study also distinguishes answers recovered through thinking from answers inferred by combining other knowledge without the target fact having been explicitly encoded.

Frontier models store more than they can summon

For GPT-5 and Gemini-3, the study’s reported encoding rate was 95–98% of tested facts. That does not translate into dependable direct recall: the same frontier-model group missed 26–34% of encoded facts before being given added reasoning effort. High factual coverage and reliable generation therefore cannot be treated as the same capability.

The scaling result sharpens the contrast. Growing Gemma3 from 1 billion to 27 billion parameters reduced encoding failures from 85% to 23%, but recall failures without thinking rose as high as 40%. The finding suggests larger models can solve much of the storage problem while expanding the pool of facts that are present yet hard to retrieve on demand.

Prompt shape changes the result

The failures were not evenly distributed. Rare, long-tail facts were encoded at rates similar to popular facts, yet frontier models showed a recall gap of more than 25% between the two groups. Models also struggled with reverse questions—asking for a subject rather than an object—even when they could recognize the same answer in a multiple-choice format.

That pattern supports a narrower operational response than simply attaching retrieval-augmented generation, or RAG, to every factual task. RAG remains necessary for current information or facts absent from the model, but the study points to reformulating a query or escalating a low-confidence response to a reasoning pass when the information may already be encoded. The trade-off is compute: reasoning takes more of it than a fast direct answer.

A useful diagnostic, with a narrow evidence base

WikiProfile’s evidence is encyclopedic, not a direct test of proprietary or highly specialized company data, where storage and recall may behave differently. The article estimates a full frontier-model profile on the suite at about $500, and says teams can reduce that cost by using fewer samples or omitting multiple-choice variants. The benchmark is available through Hugging Face for teams that want to test whether their own failures are empty shelves or lost keys.