- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper tests the circular arrangement of weekday representations in language models and explores what the spaces between the seven days mean. Moving along a curve fitted through these representations produces small changes near each weekday and larger changes between weekdays. The authors also examine whether time of day, uncertainty between days, or natural text produces representations in these intermediate regions. These findings seem to favour seven distinct weekday regions over a ring with continuous meaning.
Critical strengths
-
The paper asks a practical question about whether a clear geometric pattern also has a meaningful interpretation. It makes a distinction between observing a circular arrangement and showing that the model uses the space between the observed points.
-
The authors go beyond visualising representations and test how interventions affect the model.
-
Several experiments check alternative explanations for the findings, including time of day, uncertainty, and natural text.
Critical weaknesses
-
The contribution seems too narrow for the broader claims the paper aims to make. The study focuses on weekdays and mainly revises the interpretation of an existing finding. Even if the results hold, it remains unclear what general insight they provide beyond this particular case.
-
Weekday names provide only a coarse description of an underlying continuous temporal variable. Learning their order and cyclic relationships does not necessarily mean that interpolating between their representations will produce meaningful intermediate states. The paper does not sufficiently separate the effects of how temporal information is expressed in the training data from the properties of the learned representation itself. Moreover, even if the observed weekday representations lie on a continuous manifold, a spline fitted through these sparse points need not follow that manifold between them. The results therefore leave open whether the intermediate responses reflect a discrete representation or a mismatch between the fitted path and the underlying geometry.
-
The analysis remains mostly descriptive and does not explain why the observed structure emerges. The paper shows stable regions around weekdays, but does not identify the computations that produce these effects or the conditions under which they arise.
Questions for authors
-
Could the author test concepts other than weekdays?
-
If we sample time more irregularly and densely using continuous daytime and fit the curve through those representations, do the same sensitivity and occupancy patterns remain? Furthermore, what would be the case for other types of continuous variables that are not time-based?
-
How sensitive are the conclusions to the choice of interpolation method used to connect the observed representations?
2 comments
Thank you for the response. I maintain my score.
we thank the reviewer for their constructive review and for engaging with the specifics of the method.
on using a single manifold: we agree, we chose the weekday manifold as it has been used as the canonical example of continuous manifold steering/probing in past work (Wurgaft et al, Engels et al) so it was the natural place to test that interpretation. our secondary contribution consists of a pair of diagnostic measures (sensitivity + occupancy) that can be applied to other manifolds, we are actively using these to examine color which has natural continuous structure, to see if our methods generalise.
on dense time sampling: figure 2 does use continuous clock time (22 hours x 7 days) and we found that clock-time variation lies almost entirely outside the weekday plane (99.8% of hour-centroid variance, the leading clock time axis is 88 degrees from the plane) so fitting a curve through time sampled points would trace a direction orthogonal to the ring rather than filling its arcs. the small in-plane component is comparable to that induced by non-temporal controls (fig 2b).
on sensitivity to interpolation method: we address this concern by evaluating the Jacobian field over a polar grid covering the whole weekday plane rather than along the spline alone, and the seven minima extend radially from the centroids (fig 1c, appendix b). in Llama and Mistral the sevenfold structure also appears along radial and random orthogonal perturbation directions rather than just along the tangent and the result is also reproduced on arithmetic prompts from Wurgaft et al (appendix C). hence we don't think a different spline would change the conclusion.
on separating training data from learned representation, and mechanism: fair point, the paper is descriptive by design and does not explain why the spoke structure emerges. that said, the diagnostics should generalise to more complex manifolds where such mechanisms might be hard to identify in the first place.