The Geometry of Inference in Transformer Residual Streams
Six language models show residual streams become geometrically specific to their own final states during inference.
3 stories, newest first · one card per story
Six language models show residual streams become geometrically specific to their own final states during inference.
Decomposes transformer representation updates into parallel and perpendicular components, linking geometry to editing robustness and better pretraining.
Researchers decompose transformer updates into parallel and perpendicular components, linking representation geometry to editing robustness, compression diagnosis, and training interventions.
Today 1.6K readers · 2.5K views · 779 reads · all time 189K views