# Google Research study reveals recall limitations in frontier models like GPT-5 and Gemini-3 > Source: > Published: 2026-09-01 19:24:22+00:00 Photo: Merlin Lightpainting / Pexels # Google Research study reveals recall limitations in frontier models like GPT-5 and Gemini-3 A new framework called 'knowledge profiling' shows that most AI hallucinations stem from retrieval failures, not missing knowledge Your AI model knows the answer. It just can’t find it. That’s the central finding of a new study from Google Research and Technion, which analyzed over 4 million responses across 13 large language models and discovered that when frontier models hallucinate, the problem usually isn’t that they lack the relevant facts. The facts are sitting right there in the model’s parameters. The model simply fails to pull them up when asked. ## The lost keys problem The paper, titled “Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality,” introduces a framework the team calls “knowledge profiling.” The framework separates two distinct capabilities: encoding (whether a model has stored a fact) and recall (whether it can actually surface that fact during generation). Using a benchmark called WikiProfile, derived from Wikipedia, the researchers found that leading models have gotten remarkably good at the storage part. Gemini-3-Pro and GPT-5 both achieve 95-98% saturation in factual encoding on the benchmark. The retrieval side tells a very different story. Direct recall fails on 26-34% of those encoded facts. Even when chain-of-thought reasoning is enabled, it still misses 11-12% of what it demonstrably knows. In some of the models tested, recall failures accounted for over 70% of all factual errors. ## Scaling won’t fix this The recall failures aren’t random. They’re systematic, disproportionately affecting long-tail facts and reverse queries. Long-tail facts are the obscure, less frequently encountered pieces of information that appear rarely in training data. Reverse queries flip the typical direction of a question: instead of asking “What is the capital of France?” you ask “Paris is the capital of which country?” Models struggle more with the second formulation, even though the underlying knowledge is identical. Making models bigger primarily improves encoding, not recall, suggesting the way current transformer-based models organize and access stored information has a built-in bias toward common, forward-direction retrievals. The research team behind the study includes Nitay Calderon from Google Research and Technion, alongside Gal Yona, Eyal Ben-David, Zorik Gekhman, and Eran Ofek. A complementary paper from June 2026, titled “Thinking to Recall,” found that reasoning strategies can recover some recall failures through latent computation, essentially letting the model work through intermediate steps to access buried knowledge. But this process can introduce its own hallucinations, as the model sometimes generates false intermediate facts on its way to an answer. ## What this means for AI development and investment If recall rather than encoding is the bottleneck, then the industry’s massive spending on ever-larger training runs and datasets may be hitting diminishing returns on factual accuracy. The next frontier in reducing hallucinations likely involves better retrieval architectures, inference-time techniques, and new approaches to how models organize parametric knowledge internally. The finding also has implications for the retrieval-augmented generation (RAG) approach that has become standard practice across the industry. RAG systems feed external documents to models at inference time, essentially bypassing parametric memory altogether. This research suggests RAG may be compensating for a recall problem rather than a knowledge problem, which opens the door to potentially lighter-weight solutions if the recall mechanism itself can be improved. **Disclosure:** This article was edited by Editorial Team. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).