latentbrief
Back to news
Research3d ago

AI's Fact Recall Struggles Revealed: Key Findings from Google Research

Google AI Research1 min brief

In brief

  • Google researchers have uncovered a critical issue in large language models (LLMs): they often get facts wrong not because they lack the information, but because they can't recall it.
  • Using a new framework called knowledge profiling, the team analyzed how LLMs handle factual queries.
  • They found that while models like Gemini3 and GPT-5 encode nearly all facts during training, many errors stem from their inability to retrieve these encoded details during inference.
  • The study introduces WikiProfile, a benchmark with 2,150 Wikipedia facts paired with detailed questions to test encoding, recall, and recognition.
  • Results show that most factual mistakes are "recall failures," akin to losing keys rather than having empty shelves.
    • This distinction is crucial for improving LLM reliability-while encoding issues require bigger models or more data, recall problems may be solved with better retrieval methods during inference.
  • Looking ahead, this research highlights the need for enhanced recall techniques in AI systems.
  • Future advancements could focus on optimizing how LLMs access stored information, potentially leading to more accurate and trustworthy responses across various applications.

Terms in this brief

knowledge profiling
A method used by Google researchers to analyze how large language models handle factual queries, focusing on their ability to encode and retrieve information during training and inference.
WikiProfile
A benchmark created by Google researchers that consists of 2,150 Wikipedia facts paired with detailed questions. It tests the encoding, recall, and recognition abilities of large language models.

Read full story at Google AI Research

More briefs