Frontier models can recover up to 65% of facts they can't directly recall — just by thinking longer

Google Research and Technion researchers have upended conventional thinking about why large language models hallucinate, revealing that frontier models like GPT-5 and Gemini-3 actually encode 95-98% of tested facts but simply fail to retrieve them during generation.

The finding, detailed in a new study published on arXiv, challenges the industry assumption that hallucinations stem from missing knowledge. Engineering teams have historically tackled the problem by scaling models larger, expanding training datasets, or building complex retrieval augmented generation (RAG) architectures. The research suggests those approaches address the wrong bottleneck.

Instead of a knowledge problem, the researchers identified a recall problem. The models have the information stored parametrically in their weights but cannot surface it when prompted. Their experiments demonstrated that frontier models can recover up to 65% of facts they initially fail to recall simply by thinking longer, through techniques that extend reasoning chains and internal computation.

The implications ripple across the AI industry. Teams building AI applications have spent enormous resources solving what they believed was an encoding problem. Prompt engineering firms, RAG vendors, and model developers pursuing scale-heavy solutions may need to recalibrate their strategies. If recall is the true bottleneck, the answer lies in better decoding techniques, longer chains of thought, and improved prompting strategies rather than more parameters or more training data.

This research carries particular weight coming from Google Research, which develops Gemini models alongside DeepMind. The findings validate internal approaches Google has explored around chain-of-thought prompting and extended reasoning. OpenAI's GPT-5 performing similarly in the study suggests the pattern holds across different model architectures and training approaches.

The 95-98% encoding rate means most frontier models contain vastly more knowledge than developers realize. The gap between encoded knowledge and recalled facts represents untapped performance. If a model encodes a fact but cannot retrieve it, developers can potentially unlock that knowledge through better inference techniques rather than expensive retraining.

Several applications follow directly. Customer service AI systems hallucinating product details might improve dramatically through extended reasoning rather than retrieval databases. Medical AI systems misremembering clinical facts could gain accuracy through longer thinking steps. Code generation models producing false library calls might surface correct knowledge they already possess.

The research does include caveats. The 65% recovery rate means roughly one-third of forgotten facts stay inaccessible even with extended thinking. Encoding completeness varies by domain. Obscure or recently-added facts face lower encoding rates than widely-distributed training data. The findings apply specifically to frontier models and may not generalize to smaller, open-source models.

Still, the work reframes the hallucination problem for an industry that has invested heavily in scale. Companies building AI products can now explore inference-side improvements before committing to larger models or more training data. This opens doors for optimization techniques, specialized decoding algorithms, and reasoning architectures that unlock existing knowledge rather than adding new knowledge.

The discovery suggests the AI capability ceiling sits higher than current inference suggests. Models know more than they show. Developers who master the gap between knowledge and recall will extract substantially more value from existing models.