AI Models Can Recover Forgotten Facts by Thinking Longer
Research shows large language models can recover up to 65% of seemingly forgotten information when given additional computation time.
1New Perspective on Model Errors
When large language models (LLMs) produce incorrect information, developers typically assume the model lacks the necessary knowledge. New research from Google and Technion reveals that most of this information is actually encoded in the model's parameters but not directly accessible. Advanced models like GPT-5 and Gemini-3 can encode 95-98% of tested facts. This finding indicates that the primary bottleneck for model accuracy is often recall rather than encoding capability.
2Knowledge Profiling Method
Researchers propose a fundamental shift in model evaluation approach. They advocate for fact-level profiling instead of traditional question-level accuracy. This method tests how a single piece of information becomes accessible under different conditions. It measures whether a fact is encoded, its queryability through various phrasings, and the computational effort required for retrieval. This approach reveals the critical distinction between encoding and knowing, enabling more targeted improvements.
3Five Distinct Knowledge Profiles
The study identifies five categories of how models process information. Direct recall: The model accesses information easily. Encoding failure: The model never encoded the fact. Recall failure: Information is encoded but inaccessible. Recall with thinking: Knowledge becomes accessible only through additional inference computation. Inference without encoding: The model deduces correct answers through reasoning without direct encoding. This classification enables error-specific solutions.
4Comprehensive Experimental Findings
Researchers tested 13 different LLMs using the WikiProfile dataset containing 2,150 facts extracted from Wikipedia. In advanced models, encoding rates reached 95-98%, yet 26-34% of these facts were initially inaccessible. Interestingly, when given additional thinking time, models recovered 40-65% of previously inaccessible information. This mirrors the human 'tip-of-the-tongue' phenomenon, suggesting similar cognitive mechanisms in artificial systems.
5Limitations of Model Scaling
The study demonstrates that increasing model size isn't always the solution. When scaling Gemma3 from 1 billion to 27 billion parameters, encoding errors decreased from 85% to 23%. However, recall errors simultaneously increased to 40%. These findings show that scaling primarily addresses storage problems rather than access issues. Larger models tend to encode more knowledge in an inaccessible state.
6Cost-Effective Solutions
The research suggests smarter alternatives to expensive scaling for improving accuracy. Inference-time computation techniques (like Chain-of-Thought) can help models access knowledge they already possess. For example, a model failing to answer directly might generate intermediate thoughts to uncover the correct information. This approach offers potential for enhancing application reliability without larger models or external databases.
7Industrial Implications
These findings could significantly impact AI application development strategies. Companies often attempt to solve model errors by training larger models. However, this research shows that in many cases, the problem lies not in model capacity but in how it utilizes knowledge. Google's Nitay Calderon states, 'When facts come out wrong, the knee-jerk reaction is to scale up. But if the knowledge is already encoded, this expensive solution doesn't help.'
8Human-Like Cognitive Processes
The study reveals striking parallels between LLM knowledge access mechanisms and human cognition. Similar to humans experiencing 'tip-of-the-tongue' states, models can recover seemingly forgotten information by mentally retracing context. This parallel highlights the value of drawing inspiration from human cognition in AI system design. Understanding model behavior could lead to more natural interactions.
9Future-Oriented Implications
This work signals a potential paradigm shift in AI research. Rather than focusing solely on increasing model capacity, developing techniques to better access existing knowledge may prove more efficient. Inference-time optimizations could enhance performance while reducing training costs. The findings may enable revolutionary applications in high-stakes domains like medical diagnosis or legal analysis where accuracy is critical.