Watch: Context-Dependent Recall
The finding that a model's ability to recall a stored fact depends on how closely the query matches the context in which the fact was originally encoded during training.
Transcript
Large language models often know far more than they can remember. Recent research shows that while frontier models encode up to ninety-eight percent of facts in their internal weights, they fail to recall nearly a third of them. This phenomenon is called context-dependent recall. The bottleneck is not storage, but access.
This explains the famous reversal curse. A model might know that person A created product B, but fail to answer who created product B when asked from the opposite direction. The knowledge is physically there, but the pathway to reach it is too weak. We see a similar gap with rare facts. Models actually store niche information quite well, but they struggle to retrieve it if a prompt deviates even slightly from how they first learned it.
Fortunately, there is a way to bridge this gap. When models are allowed to use intermediate reasoning steps, often called thinking modes, they recover more than half of these hidden facts. The thinking process acts as an associative bridge, helping the model navigate its own memory.
This research reframes how we think about AI hallucinations. When a model confidently states a falsehood, it might not be a gap in its training data. The correct fact may be fully encoded, but simply unreachable from the direction the user approached it. When the path is blocked, the model simply fills the silence with what sounds most likely.