Editorial illustration for AI models can retrieve 65% of "inaccessible" facts by thinking longer
AI Models Unlock 65% of Hidden Facts With Extended Thinking
AI models can retrieve 65% of "inaccessible" facts by thinking longer
A model that gets a question wrong doesn't necessarily lack the answer. That's the finding from a new study by researchers at Google Research and Technion, who tested GPT-5 and Gemini-3 on facts the models had already failed to produce in normal use. Rather than treating a wrong answer as a gap in training data, the team checked whether the fact was still sitting somewhere in the model's parameters, just unreachable through the usual generation path.
The distinction matters because it changes what engineers should build. If a model is missing facts, the fix is more data, bigger architectures, or bolted-on retrieval systems. If a model has the facts but can't surface them, the fix is entirely different: better prompting, more inference-time computation, or techniques that help the model search its own memory more thoroughly.
The researchers found that 95-98% of tested facts were encoded somewhere in these frontier models, even when the models couldn't answer questions about them directly. That gap between "encoded" and "retrievable" is what the study sets out to measure, using a method the researchers call fact-level profiling.
Providing models with extra computational effort successfully retrieves 40-65% of the encoded facts that models initially fail to directly recall. The researchers compare this to the human tip-of-the-tongue state, where deliberate effort, such as mentally retracing context, eventually helps remember the information.
Why this matters
If the Google Research and Technion team is right, a chunk of what we've been calling "hallucination" is really a retrieval failure, not a knowledge gap. That reframes years of engineering instinct. Teams that chase hallucinations by scaling parameters or bolting on retrieval-augmented generation may be treating the wrong problem.
The fact sits in the weights; the model just can't reach it under the prompting conditions used. A 65% recovery rate from extra inference-time compute is a big number, and it suggests cheaper fixes than another training run: better prompting, more test-time reasoning, or decoding strategies tuned to how facts were originally learned. We'd push back a little on the neatness of "just think longer" as a fix, since it trades latency and compute cost for accuracy, and that tradeoff won't work for every product.
Still, for researchers, this is a reason to instrument models for recall failures specifically, not just accuracy. For founders building on top of frontier models, it's a reminder that the missing feature might already exist, buried and unreachable.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Anthropic's Claude 5.1 models cut Fable cache read costs by 75%
From Archives Google Pics Brings AI Image Generation to Google Workspace
Common Questions Answered
What percentage of inaccessible facts can AI models retrieve by thinking longer?
According to the study by Google Research and Technion, AI models can retrieve 40-65% of encoded facts that they initially fail to directly recall by providing extra computational effort. This recovery rate demonstrates that many facts are stored in the model's parameters but are simply unreachable through standard generation paths.
How does the tip-of-the-tongue phenomenon relate to AI model retrieval failures?
Researchers compare AI retrieval failures to the human tip-of-the-tongue state, where deliberate effort such as mentally retracing context eventually helps remember information. This analogy suggests that when models fail to produce facts, the information may still exist in their parameters but requires additional computational effort to access.
Why does distinguishing between retrieval failure and knowledge gaps matter for AI development?
If retrieval failure rather than knowledge gaps explains many AI errors, it reframes years of engineering instinct and suggests that teams chasing hallucinations through scaling parameters or retrieval-augmented generation may be treating the wrong problem. This distinction implies that the solution lies in improving inference-time computation rather than simply adding more training data or model parameters.
Which AI models were tested in the Google Research and Technion study?
The researchers tested GPT-5 and Gemini-3 on facts that the models had already failed to produce in normal use. These frontier models were specifically chosen to evaluate whether providing extra computational effort could retrieve previously inaccessible facts from their parameters.
Further Reading
- Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality in Large Language Models - hyper.ai
- Google Research Finds WikiProfile Reveals Recall, Not Knowledge, Breaks Frontier LLMs - AlphaSignal
- Google: Subject/Object Entity Order Affects AI Answers - Search Engine Journal
- Thinking to Recall: Why Reasoning Helps LLMs Remember Facts - toknow.ai
- Recall is the Bottleneck of Factuality in LLMs - Automaziot
