Definition
Plain language
Earlier research showing that whether a statement is true or false shows up as a direction inside a model's internal state.
As stated in the literature
A line of interpretability work (Marks and Tegmark) establishing that true and false statements are linearly separable in a transformer's residual stream, motivating later probes for truthfulness, staleness, and judgment directions.
Why it matters: It suggests a model internally tracks truth in a readable way, opening the door to probes that detect when a model 'knows' something is false.
For example, this work found you can draw a line inside a model's internal state that separates the statements it treats as true from those it treats as false.
Heard on the show
“And there's a well-known result from a couple years back — the "Geometry of Truth" line of work — showing that true-versus-false sits along one such direction.”Episode 037 — Why Hallucination Detectors Miss Stale Facts: A Geometric Story About What Models Know But Don't Say