Frontier models can recover up to 65% of facts they can't directly recall — just by thinking longer | VentureBeat
The Independent
Voice on AI
Orchestration
Infrastructure
Data
Security
Technology
More
Newsletters
Frontier models can recover up to 65% of facts they can't directly recall — just by thinking longer.
Ben Dickson
12:20 pm, PT, September 1, 2026
Image credit: VentureBeat with ChatGPT
When large language models (LLMs) hallucinate, developers often assume the model lacks essential facts. Teams typically diagnose the issue as a knowledge gap and attempt to address it by:
- Increasing model size
- Expanding training data
- Building complex retrieval architectures
A new study by researchers at Google Research and Technion challenges this assumption. They found that frontier models like GPT-5 and Gemini-3 encode 95-98% of tested facts. This suggests that recall, not encoding, is often the primary factor limiting factual accuracy.
By understanding how to unlock existing knowledge through inference-time computation, developers can build more reliable applications without relying on larger models or external databases.
Knowledge Profiling: Measuring What Models Actually Know
The researchers propose shifting the evaluation focus from question-level accuracy to fact-level profiling. This approach tests a single piece of information across various conditions, revealing whether it's stored parametrically ("encoded") and if the model can reliably access it ("know").
The paper illustrates this using a sample fact: Oasis played their first gig at the Boardwalk club. The study categorizes knowledge into five distinct profiles:
- Direct recall: The model readily accesses the fact to answer direct questions without extra inference compute.
- Encoding failure (empty shelves): The model neither encodes nor knows the fact. It cannot provide any information about Oasis's early days.
- Recall failure (lost keys): The model has the fact encoded but cannot access it directly.
- Recall with thinking (chain-of-thought): The fact is encoded but inaccessible to direct generation. It can only be recalled through inference-time computation, such as Chain-of-Thought, which primes the model to locate and retrieve the answer.
- Inference without encoding: The model generates an answer that appears correct, even though it doesn't have a clear "encoding" of the fact.