When and How Long? The Readout-Mediator Angle in Temporal Reasoning
For interpretability researchers and AI safety practitioners, the paper demonstrates a fundamental failure mode of probe-based methods: probes can report high confidence on directions the model has silently abandoned, invalidating their use as runtime safety monitors.
The paper shows that linear probes can decode task-relevant information from language model activations while being completely orthogonal to the model's actual causal computation, as demonstrated on calendar-date duration reasoning where a sin/cos probe recovers day-of-year but ablating its direction has no effect, whereas a DAS-found subspace collapses performance. This dissociation replicates across scales (1.5-9B parameters) and model families, undermining probe-based interpretability for safety monitoring.
A linear probe can decode a representation almost perfectly and yet be completely irrelevant to how the model uses it. On calendar-date duration reasoning in language models, a $\sin$/$\cos$ probe recovers day-of-year from a layer's activations, yet ablating its direction has no effect on the model's answers -- while ablating a four-dimensional subspace found by Distributed Alignment Search (DAS) at the same layer collapses performance entirely. We measure the angle between these two subspaces -- the \emph{readout-mediator angle} -- and find it indistinguishable from the angle between two random subspaces (the Haar-uniform null), meaning the probe has learned a direction orthogonal to the model's actual computation. Reverse-engineering the circuit reveals why: attention heads route month-grained context through learned QK offsets at ${\pm}30$ and ${\pm}61$ days, and MLPs then convert \emph{when} (absolute date) into \emph{how long} (duration) -- all downstream of the causal subspace the probe never touches. Sparse-autoencoder decomposition confirms the split: probe-aligned and DAS-aligned features encode semantically disjoint concepts with negligible causal overlap. The dissociation replicates across four scales ($1.5$-$9\,$B) and two model families, with preliminary evidence on two further domains (spatial displacement, symbolic arithmetic), suggesting that readout-mediator orthogonality is a general failure mode of probe-based interpretability. This directly undermines proposals to deploy probes as runtime safety monitors: the probe can report high confidence on a direction the model has silently abandoned.