acceptodds
Under review as a conference paper at ICLR 2027

Sky sphere representation in language models

Abstract

We analyze whether language models have an internal representation of the night sky that they use in answering questions about proximity on the sky. We find such a representation (a "sky sphere") in residual streams of tested LLMs (size 32B-235B parameters) on prompts like "in the night sky [object X] appears near". In 6 out of 7 models the sky sphere is linearly decodable ( score of 80–90% and median angular error down to – in leave-one-out test). The sky sphere is indeed spherical and not a flat atlas chart, and spectral embeddings of graphs derived from constellation grouping and adjacency or from corpus co-occurrence failed to fit the observed geometry as well as the straight sphere. However the neural sky sphere is bent: including deg-2 spherical harmonics in the model was crucial for the following intervention experiments that tested causal use of the sky sphere. Ablating the 8 directions utilized by the bent sky sphere raises the angular distance to a neighbor provided by LLM's output from - to - in most models (for control, ablating the complementary to the "sky sphere" in top 16 PCA raises the separation by less than ). In steering, when asked for a neighbor of constellation A, replacing "sky" directions with the model fit for a distant constellation B shifts its prediction (probability 60%-85% near A) to the target (47%-83% near B, 0%-4% still near A), which shows active use of the "sky sphere". To our knowledge, this representation is the first example of a (curved) multi-dimensional irreducible feature manifold. Code is released.

Primary area:
interpretability and explainable AI
Keywords:
AI, Mechanistic Interpretability, Feature Manifold

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_GEW9OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether language models use a geometrically structured representation of the night sky when answering proximity questions. It analyzes residual-stream activations for 188 astronomical objects and 25 prompt templates across seven models. Celestial coordinates are linearly decodable in six models, with reported leave-one-object-out R² values of 73–91% and median angular errors of approximately 11°–18°. The authors fit first- and second-order spherical harmonics, ablate the resulting subspaces, and replace fitted representations to steer responses toward another constellation. They also compare spherical geometry with flat, cylindrical, graph-based, and corpus-derived alternatives, arguing for an irreducible spherical feature manifold.

Critical strengths
  1. The representation finding is supported across multiple models and objects. Six of seven models show substantial coordinate decodability, supported by object-level validation and additional 80/20 splits with permutation tests. The paper also reports the weaker gpt-oss result and the reduced prominence under generic prompts. These observations make the finding informative within its stated proximity-task setting.

  2. The interventions provide evidence that the fitted spatial subspace is behaviorally relevant. Ablating the eight-dimensional harmonic subspace markedly degrades proximity judgments, while ablating its equal-dimensional complement has little effect. Steering toward another constellation shifts responses toward that target, and writing back the original constellation provides a useful control. Together with the usable-response analysis for ablation, this supports a functional role for the identified information, even though the precise geometric mechanism requires further controls.

Critical weaknesses
  1. The exclusion of competing geometries is stronger than the analysis supports. Appendix F.1 regresses the activation squared-distance matrix on squared differences of individual candidate coordinates. For a general linear embedding h = Aφ, the corresponding distance is ΔφᵀAᵀAΔφ, which includes cross terms unless the induced metric is diagonal. Equation (3) therefore does not cover arbitrary linear mixtures of the candidate features. Moreover, with strongly correlated predictors, nonpositive coefficients cannot establish that a representation is absent. The paper acknowledges coefficient instability but nevertheless concludes “complete absence” in F.1 and claims high-confidence exclusion of reducible representations in the abstract. These conclusions should be narrowed to the tested models and metric assumptions, or supported by held-out comparisons that allow suitably regularized candidate metrics and report uncertainty. This matters for the claimed irreducibility, although it does not invalidate the separate decoding and intervention results.

    There is also a mathematical ambiguity in the co-occurrence comparison. F.3 equates M* = ΦΛΦᵀ with (ΦΣ)(ΦΣ)ᵀ for Σ = √|Λ|. The latter is Φ|Λ*|Φᵀ and differs from M* if negative eigenvalues are present. A signed or separate input/context factorization may be intended; the paper should clarify the construction and whether the Q_text comparison is affected.

  2. The validation and intervention controls limit the strength of the generalization and geometric-specificity claims. Appendix D fits PCA using all objects before holding objects out of the coordinate regression; the accompanying repository protocol explicitly describes this as evaluation conditional on a globally fitted basis. Table 5 does not explain how its reported layers were selected. Appendix C also describes selecting proximity-focused prompts based on performance. The resulting validation is useful for describing the observed representations, but does not fully isolate generalization to objects and prompts that played no role in constructing or selecting the analysis. PCA does not use held-out coordinate labels, so this should not be interpreted as demonstrated label leakage.

    The intervention controls also leave unresolved alternatives. The complementary-subspace ablation matches dimension but is not reported to match removed activation variance; E.1 itself acknowledges dimensionality and overlap as potential explanations for the larger effect of second-order components. Steering holds out two templates while retaining the constellation identities in the fitted dataset, and directly naming target B counts as success, although Figure 7 indicates such answers are a minority. The repository specifies 88 source constellations and two templates, giving 176 prompts per arm, and documents object-cluster bootstrap intervals for ablation. The corresponding intervals and paired steering uncertainty should accompany the paper’s reported effects. These limitations weaken claims about a specific spherical computation and its generalization; they do not erase the observed behavioral effects.

Questions for authors

The following questions concern the strength and interpretation of the submitted evidence and could materially affect the assessment.

  1. Alternative geometries: Does the spherical advantage persist when each candidate receives a regularized linear map or positive-semidefinite metric, including cross terms, fitted on training objects and evaluated on held-out objects? What uncertainty and sensitivity to coordinate choices support the exclusion claim in F.1?

  2. Validation pipeline: How was the layer in Table 5 selected? What results remain when PCA, layer choice, and dimensionality are determined using training objects only, and prompts are specified independently of performance-based selection? Please include variability across the reported object splits.

  3. Intervention specificity and statistics: The repository describes 176 prompts per steering arm and object-cluster bootstrap intervals for ablation. Please include the corresponding intervals in the paper and report paired steering uncertainty and usable-response denominators. Does the effect persist with dimension- and variance-matched ablation controls and with source/target identities excluded from fitting? What is the steering success rate when direct answers of B are excluded?

  4. Co-occurrence factorization: Is M* in F.3 positive semidefinite in the experiments? If not, what signed or input/context factorization is intended, what was implemented, and does correcting the stated identity change the Q_text comparison?

4 comments

Thank you for the detailed critique, we've tried to address main points of contention as follows.

  1. Exclusion of alternative geometries. On one hand, we soften the language concerning these results, narrowing them to the geometries considered. On the other hand, we expand the competitors pool slightly (by choosing other natural axes for the cylinder and other cut meridians for the flat chart) and focus on the best competitor versions, and we switch the comparison metric from "weights of distinct dimensions in RDM under orthogonality assumption" to "unique R^2 score added by a competitor model to the base one under general ridge regression" (this no longer assumes orthogonality of models). The result stands, in the sense that competitor models add negligible R^2 to the sphere, compared to what sphere adds to them (details are in the paper).
  2. Validation issues. We removed Table 5 and instead of LOO test, score every layer with a 80/20 split validation test (abstract reports best scores across layers, per model, spreads across models). Fitting now should only use train 80 train set for choices like PCA, ridge value and the fit itself (choosing the best layer inside the training objects changes the best scores by at most 0.005). In steering, the source A and Target B are also excluded from the fitting process. We add 0.05-0.95 bands or 95% uncertainty intervals to most charts (constructed from resampling 80-20 splits, or bootstrapping the objects, depending on the situation).
  3. Intervention. Spreads are added: Fig 17 for ablation based on variability across objects; Tables 6,7,8 for steering via bootsrapping over objects, so no genuinely new info since we report the sample fraction, not a function of objects, but Fig 19 shows variability across 5 prompt templates. Paired steering uncertainty and usable-response denominators are also in these tables. A and B are now excluded from the steering fit, the effect barely changed. Regarding removing B from the success count: on one hand, we now report those figures (Tables 6,7,8); "near B but not B" is still around 50% when using l=1,2 harmonics, and "near A" is <2%. On the other hand, we would argue that excluding B may not be a refinement when testing our representation. While we don't know the details of how LLMs exactly organize their answering to our task, a reasonable picture might be "use the spherical representation to collect objects near A -> use other clues to pick the answer from them, in particular, use the A token to ban A itself". In that case, our intervention would switch the goal from "Near A, but not A", to "Near B, but not A", since "not A" part doesn't come from the proximity data. In this interpretation, B would be a successful steering output.
  4. Co-occurrence. First, we expanded the corpus until the saturation of the scores. Second, we adjusted the empirical probabilities to quantities built from counts alone (not using corpus size) in order to prevent large non-astronomy fraction from degenerating M* (details in the paper). Still, about half of its eigenvalues are negative. We show the results for versions where they are kept negative or are zeroed out. The version with absolute values used to show much worse performance on smaller datasets, and is geometrically unmotivated, so it was abandoned. Now Q_text is much closer to other metrics, but only on Qwen it beats the plain sphere (not the bent sphere), and only when restricted to constellations. So it agrees with data reasonably, but not to the point of suggesting that it is the real mechanism form which the sphere incidentally emerges.
  5. Variance-matched ablation. On one hand, we can report that "in the six models with a sky representation the control removes 40--47% of the top-16 PCA variance, comparable to the l=1,2l=1,2 ablation (53--60%) and more than the l=1l=1 ablation (29--35%)", so the variance attacked by control is on the same order (in some cases smaller), but the effects are comparatively tiny. For now we prioritized other fixes and left that untouched since ablation is costly, but we are opened to expanding controls to match the variance exactly if deemed necessary.

Update: we added a variance-matched-controls version of the ablation experiment. The trends stay the same: controls barely move the performance (they even increase parsed fraction for unknown reasons), l=1 and l=2 on their own harm, but not too much, together they bring accuracy close to chance.

Please see whether this, and the prior changes address your concerns and change your assessment. Given that your review brought up the most of the issues we address in the revision, that your rating was the most harsh, and that you indicated that addressing these issues might change it, we would really appreciate you posting your opinion on the revised paper.

Reviewer_GEW9OpenReview ↗

Official Comment by Reviewer GEW9

Thank you for the detailed responses and revisions. The authors have addressed several of my concerns, and I have decided to raise my rating to Weak Reject. I have two remaining concerns:

  1. The paper uses only 8 pages despite the 9-page limit. Could the authors clarify whether the current presentation sufficiently develops the methodological contributions and experimental analysis?

  2. The manuscript contains noticeable formatting issues, such as single words appearing on separate lines (e.g., Lines 38 and 145), which affect readability. I encourage the authors to improve the typesetting in the revision.

Thank you for reading the revision and for raising your score. We have uploaded a new version that addresses both points.

  1. Length and development of the methodology. We agree that the main text was too compressed: several of the controls you asked for in your first review appeared only in the appendix. The additions are:

Section 3.3: a formal statement of the harmonic fit and of the ablation and steering interventions (Equation 1). Section 3.3: the variance-matched ablation controls with 95% confidence intervals (Figure 6), together with the variance removed by the dimension-matched control. Section 3.3: per-model steering rates with paired bootstrap intervals and usable-response denominators, including rates with direct answers of B excluded (Table 1). Section 3.3: the variance explained by the l=1 and l=1,2 fits, which shows the bend without any intervention. Section 3.2: the check on 25 new prompt templates that were fixed before any activations were extracted. Section 3.4: the held-out unique-R² protocol for the alternative geometries, including the cylinder axes and chart cuts tested. Section 5: a new Limitations section.

All of this material was already in the appendix or in our earlier replies. It has been moved or restated in the main text, and no claims were changed.

  1. Typesetting. We fixed the single-word lines at Lines 38 and 145, removed the other single-word line endings in the main text, rescaled the corpus table in Appendix F.3 that extended into the margin, and corrected the remaining typos.

The points in your original review (alternative geometries, validation pipeline, intervention controls, co-occurrence factorization) are not among your remaining concerns. If any technical issues still keep the paper below the acceptance threshold, we would be grateful to know and will address it in the time remaining.

Reviewer_WhnCOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper conduct a series of experiments to discover the spherical feature manifold within the action of multiple language models. The author trains a linear prober from the top-4/8/16 principal components of the model activation when prompted with some questions like "the age of X is estimated to be" or "the name of X comes from". Then the author uses leave-one-out to validate the capacity of the linear probe to find that the linear prober is able to predict the position of stars. The author further uses least-square algorithms to fit three first harmonics and five second harmonics. By ablating the first harmonics and second harmonics from the 16 principle components, the author measures the contribution of each harmonics. It also includes an experiment to compare several different competing representations

Critical strengths
  • It verifies and reveals the feasibility of predicting the position of stars from the action of open-source language models given specific prompt;
  • It conducts extensive experiment with fine-grained control, ablation and steering experiments to provide the casual evidence.
  • It proposes to model and testify the spherical feature manifolds with not only the first harmonics but also the second harmonics.
Critical weaknesses
  • The preprocessing to build PCA includes the stars used for validation and testing.
  • In table 5, it provides the accuracy from different models and different layers but there is no explanation on how these layers are selected or any validation set is involved.
  • In the more strict second harmonic ablation experiments where only the components that are orthogonal to the first harmonics are ablated, the change seems only to be substantial for GLM-4.5-Air and Qwen3-235B as in Figure 16. Does the author have any explanation for this?
  • In competing representation section, the author uses corpus-derived embeddings from Wiki dump 2023 but we don't know if it is a good representative of the open-sourced LLM's t training corpus.
Questions for authors
  • How is the R-square when fitting the 16-d PCA activation with the three first harmonics and five second harmonics?
2 comments

Thank you for pointing out crucial issues and providing your concerns. Here is our response.

  1. Validation issues. We removed Table 5 and instead of LOO test, score every layer with a 80/20 split validation test (abstract reports best scores across layers, per model, spreads across models). Fitting now should only use train 80 train set for choices like PCA, ridge value and the fit itself (choosing the best layer inside the training objects changes the best scores by at most 0.005). We sample 100 splits, report median and 5-95% spread.

  2. Anomalous (l=2)⊥(l=2)^\perp ablation. Two models in that chart were mislabeled (fixed now), we assume the question is about Qwen3-32B and gpt-oss, as per current chart. In short, we don't have a robust explanation for the outliers. we note that gpt-oss (aside from generally having a weak and potentially more fragile spherical representation) is also an outlier in having l=1 and l=2 spans being less orthogonal than in other models, so it is more affected by l=1 no longer being cleanly extractable after l=2 is projected onto it along (l=2)⊥(l=2)^\perp. However all other models are roughly similar in that regard, Qwen3-32B being not worse than Qwen3-235B, yet suffering in ablation, unlike the larger Qwen. Maybe, being the smallest model, it is generally more fragile to interventions, but whatever the reasons for these peculiarities, we would argue that our findings do not rest on them, and while that is an interesting question, deeming it a critical weakness seems excessive to us.

  3. Co-occurrence corpus. The exact training data is not publicly available for analyzed models, so we cannot be sure about how representative our corpus was. However, it was far too small. We expanded the corpus until the saturation of the scores. In order to prevent non-astronomy portions from affecting the counts and degenerating the learned co-occurrence matrix M*, we adjusted the empirical probabilities to quantities built from counts alone (not using corpus size), details in the paper. Now co-occurrence derived fit is much closer to other metrics, but only on Qwen it beats the plain sphere (but not the bent sphere), and only when restricted to constellations. So it agrees with data reasonably, but not to the point of suggesting that it is the real mechanism form which the sphere incidentally emerges.

  4. Fitting with higher harmonics. Since higher harmonics are components of the embedding of the sphere (encoding), but we mainly measure R^2 for the decoder (linear projection out of the neural space), the comparison to our figures wouldn't be apples to apples. Nevertheless, here is a small test on layer 80 of Mistral (although the decoder r^2 has a large plateau, the representation grows in size across depth, so encoder r^2 is better in later layers). The top-16 PCA variance explained by l=1 encoder is 0.61 [0.57, 0.64] (5-95% range over splits), but with l=1,2 it is 0.78 [0.74, 0.82], 17 percentage points higher. Here we allowed individual intercepts for different types (constellation vs star vs other), without it we got 0.37 -> 0.53, lower values (the type component of the variance is missing), but the same .17 gain. For a "bent decoder", we trained a projection to R8\mathbb{R}^8, a quadratic distortion on R8\mathbb{R}^8: x↦x+Q(x)x\mapsto x+Q(x) where Q is degree 2, and a final linear projection to the target R3\mathbb{R}^3. The score on top-16 PCA barely moved from 0.88 [0.83, 0.92] by the increment of +0.017 [+0.005, +0.036], improving angular error by less than a degree. One possible (although, unlikely in our view) interpretation of this apparent ceiling (that many models seem to come close to) is that the neural space might encode locations not of the named reference object, but already of its neighbor, so some angular error of ∼10∘\sim 10^\circ might be intrinsic to the data. Adding l=2l=2 only slightly worsened decoder performance on the top-8 PCA or if the decoder was trained as a simple quadratic ridge regression. We haven't added such investigations to the paper.

Dear reviewer, please let us know whether our edits addressed your concerns and affect your assessment, thank you.

Reviewer_iF7bOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper shows that large language models represent the positions of stars and constellations on an internal sphere, which can be read out from their activations in six of seven open models. Removing or overwriting this representation changes the models’ answers about which objects are close on the sky, so the models appear to use it. The authors also compare the sphere with alternative geometries (flat chart, cylinder, graph- and text-based embeddings) and argue that the sphere fits best.

Critical strengths
  1. The main strength is being able to show causally that the models actually use the representations for the task using the removal and overwrite experiments. From what I can tell of previous literature, previous papers on internal representations stop at showing that some information can be read out. They also show that a slightly bent sphere explains the behaviour better than a perfect one. This makes the work more than an exploratory study.

  2. A second strength is the wide range of experiments run. The results are consistent across seven models from different families and come with reasonable checks: a permutation test, and a comparison between proximity-focused and generic prompts showing that the signal depends on the task.

Critical weaknesses
  1. It seems to me that the claim that alternative representations built from one-dimensional pieces are ruled out with high confidence is overstated, because the alternatives checked, the flat chart and cylinder, are not exhaustive of possible ways of assembling one-dimensional pieces. Also, I didn't see any uncertainty estimation in the experiments results.
Questions for authors
  1. Is the PCA fitted on individual prompts or on per-object averages?

  2. The 90° “random chance” level holds for points spread uniformly over the sphere, which neither the objects nor the models answers are. Could you report an empirical baseline, for example by shuffling the models answers across query objects?

  3. It might be interesting to test whether distance is really absent. For instance, testing directly whether star distance can be decoded anywhere in the activations, ideally with prompts that ask about distance.

2 comments

Thank you for your help in improving our paper.

  1. Exclusion of alternative geometries. On one hand, we soften the language concerning these results, narrowing them to the geometries considered. On the other hand, we expand the competitors pool slightly (by choosing other natural axes for the cylinder and other cut meridians for the flat chart) and focus on the best competitor versions, and we switch the comparison metric from "weights of distinct dimensions in RDM under orthogonality assumption" to "unique R^2 score added by a competitor model to the base one under general ridge regression" (this no longer assumes orthogonality of models). The result stands, in the sense that competitor models add negligible R^2 to the sphere, compared to what sphere adds to them (details are in the paper).
  2. Uncertainty estimations have been added to most charts, either from variation across 80-20 splits that are now the basis of the most of R^2 measurement, or from variation across objects (for example in ablation experiments).
  3. PCA is fitted on individual prompts, and now only on the 80 train portion (that is also exclusively used for fitting).
  4. We have added null angular separation baselines to Figure 2 (linear decoder accuracy) and Figure 5 (LLM accuracy under ablation). While the second one is close to 90∘90^\circ, the first one is notably smaller, about 83∘83^\circ, aligned with the models in early layers (with little to no signal, the regression sends everything near the center of mass, then projects onto the sphere, into the more densely populated hemisphere, where the median angle to other points is less than from a random point).
  5. We haven't tried yet to decode the distance, focusing instead on other fixes. There are a few issues with it, which already appeared when we tested the Euclidean representation. The distance is not defined (and doesn't have a natural proxy) for the cleanest signal carriers in our dataset, the constellations; and the stars have a large portion that is not firmly known by the tested LLMs. Second, the distances span orders of magnitude, so we wouldn't expect linear encoding. Maybe some tricks akin to RoPE are utilized, maybe the Log in encoded instead, but without a clear target this problem becomes actually harder that the sky position task, despite lower dimensionality; so we don't have an answer here at this time.

I thank the authors for addressing my comments, I think their modifications have improved the manuscript. I acknowledge their justification for not answering the distance decoding question at this time, I believe their arguments make sense. I also appreciate the conciseness of the replies.

I keep my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

First revision main changes summary.

  1. Methodology of scoring the signal changed from LOO tests on globally fit PCA to 80/20 splits validations where fits see only the train 80%. 2) Claims about excluding alternatives built from one-dimensional representations significantly softened to reflect the actual results better.
  2. The prior comparison of alternative geometries switched to measuring unique R^2 added by individual models; no longer relies on the assumption that representations coexist orthogonally; variations of cylinder (other natural axes) and plain chart (other meridian cuts) are added.
  3. For most plots, statistics is assembled to plot 0.05-0.95 spread bands or 95% confidence intervals.
  4. For co-occurrence model, the corpus is expanded significantly; saturation of performance is observed; the model now keeps the negative eigenvalues of the co-occurrence matrix or zeroes them out, rather than taking their absolute value (that has led to much worse scores).
  5. Other minor changes.
    • Later revision: added a variance-matched version of ablation experiment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.