gpt2-circular-rep-fails-downstream

IN premise — summaries/2026/08/24/engels-2024-not-all-features-linear-s4-s-parse-autoencoders-find-multi-d-imensional-features.md

Created 2026-08-25T02:58:02+00:00

GPT-2 exhibits the circular representation for days/months in its hidden states but fails the downstream modular arithmetic task, indicating the representation alone is insufficient without appropriate downstream circuitry.

Summary

GPT-2 arranges days and months in a circular pattern inside its internal representations, yet it still fails when asked to actually perform arithmetic on those values. This means the model "knows" the shape of the data without having the functional wiring to compute with it, so checking internal geometry alone is not a reliable test of whether a model can do a task.