acceptodds
Under review as a conference paper at ICLR 2027

Cone Geometry is a Routing-Controlled State

Abstract

Transformer layers are prone to over-mixing: token representations can lose distinguishability with depth or context, appearing as rank collapse, spectral compression, or narrow cone-like directional concentration. This concentration is usually described through coupled correlates–attention sinks, norm outliers, massive activations, and spectral compression–yet it remains unclear which is a controllable cause of the cone rather than a co-occurring symptom. We test one candidate: recipient-side structural sink routing. We suppress incoming attention to structural sink tokens and measure the resulting change in cone geometry, applying the same intervention to a text-to-image diffusion transformer (Stable Diffusion 3) and to a 16-model language-model suite. In language models, norm controls (output-norm replay, output-norm stress, and residual-stream norm stress) do not reproduce the effect, even though residual norm stress moves spectral-compression metrics. Thus, in the controllable sense, the cone is routing-controlled rather than norm-locked. Across both domains, this routing handle is signed: special/EOT-like SD3 text carriers open the image-token cone while padding-like carriers close it, and LM first-token routes form a carrier- and depth-dependent landscape rather than a uniformly positive sink effect. A pre-registered two-dataset breadth audit shows that these LM signs are mostly dataset-stable, and a first-order value-replacement probe acts as a cheap pre-intervention sign diagnostic: without running the full routing suppression, it predicts whether candidate edits will open or narrow the cone above chance. The same routing interventions also move next-token logit distributions far more than matched controls. Recipient-side routing is therefore a separable, predictable, and signed causal axis for cone geometry–and for immediate output behavior–in the models we study.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic Interpretability, Representation Geometry, Attention Routing, Attention Sinks, Transformers

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_t8ewOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how attention routing affects the directional concentration (cone geometry) of Transformer representations. With inputs and weights fixed, the authors suppress attention to selected recipients (e.g., the first token, EOT, or padding) before the softmax, and measure whether the cone opens or narrows using the mean pairwise cosine similarity between tokens.

The experiments have three main parts:

  • In SD3, suppressing EOT-like tokens opens the image-token cone, while suppressing padding narrows it.
  • Layer-wise first-token suppression in LMs such as Qwen3-14B shows that the sign of the effect varies with the intervention and readout layers, and can flip during propagation.
  • A fixed-depth reanalysis of 16 LMs finds that signs are mostly consistent across two datasets but differ across models.

On the theory side, the authors derive the value-replacement update induced by suppression and its first-order effect on the cone, and at four sites they decompose finite-amplitude prediction errors into readout and propagation terms. The paper also includes robustness checks on mask strength, exclusion of the suppressed token, and comparison with spectral metrics.

Critical strengths
  1. The SD3 sign split is clean and shows that sink suppression does not have a single geometric direction. At the same model, layer, and denoising step, suppressing EOT-like tokens opens the cone, while suppressing padding narrows it. This holds for nearly every prompt, with tight CIs. Padding receives more attention mass yet produces the opposite sign, so the effect is not simply driven by mass. The family-level sign split in Table 3 points the same way. This is a useful corrective to the common assumption that sinks always compress representations, and it is the main reason I find the paper's observations valuable.

  2. The experimental design is careful, and the narrow claim is credible. The checks include:

  • signs are preserved across mask strengths from 10 to 10⁴;
  • excluding the suppressed token leaves the effect essentially unchanged;
  • the breadth analysis uses a fixed depth rule rather than per-model cherry-picking;
  • cosine concentration is explicitly distinguished from spectral compression.
Critical weaknesses
  1. My main issue after reading the paper is that I am not sure what it ultimately wants to tell me. What can be concluded with confidence is that suppressing sinks changes the cone, but whether it opens or narrows depends on the token type, the layer, and the model. This reads more like a record of a phenomenon than a conclusion. The title says the cone is "routing-controlled", yet the paper never explains how it is controlled, and the sign cannot even be determined in advance. The SD3 effect is quite small to begin with and essentially vanishes after step 0. The link between cone changes and model outputs is only correlational. On top of that, the heavy hedging and process-style wording throughout the main text ("historical", "frozen criterion", etc.) make it read like an experiment log and make the main thread harder to follow. For an analysis paper, I don't think stopping at "it depends" is enough.

  2. The theory does not explain the most important part of the phenomenon. Eq. 3 and Eq. 4 are essentially softmax renormalization plus the chain rule. More importantly, the first-order prediction is reasonably accurate only at sites where the cone opens, and it largely fails where the cone narrows. For TinyLlama, for example, the linear prediction is close to zero while the observed effect is −0.146, and the response is not even monotonic in the perturbation scale. Of the four diagnostic sites, only OLMo meets the authors' own criterion. In addition, the signs in Table 3 largely split by model family. Based on the authors' own Eq. 22, this is likely related to the direction of each family's sink values, but the paper does not pursue this analysis.

Questions for authors
  1. The signs in Table 3 seem to split largely by model family: GPT-2 and Qwen3 are positive, while Pythia, OLMo, TinyLlama and others are negative. I'm curious whether this sign can be predicted directly from quantities in the baseline forward pass, e.g., the first-order term in Eq. 22, or simply the alignment between the sink value and the centroid. If this works reasonably well across the 16 models, the paper would have a much clearer conclusion, and I would reconsider my score.

  2. Mean cosine in LMs is often dominated by a few outlier dimensions (Timkey & Van Schijndel, 2021), and suppressing sinks is likely to affect massive activations. If you remove the top-k outlier dimensions, or standardize each dimension before computing cosine, do the sign and magnitude of ΔC still hold? If the effect mostly comes from a few coordinates, the "cone geometry" framing may need to be revised.

  3. (Clarification) For SD3 at layer 12, step 0, what is the baseline C of the image tokens, and how much does it vary across prompts and seeds? With only ΔC ≈ 0.002 reported, it is hard for me to tell whether this is a large effect. Also, does this intervention have any visible effect on the final generated images?

Reviewer_AxDNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In these works, the authors try to understand what mechanism leads to cone-like anisotropic structure in transformers models. Their main claim is that attention routing, how attention is spread on the keys builds up this narrow conic space. To prove that, they show that by suppressing attention to special text recipients, they can succesfully open up the cone. They conduce multiple experiments on transformers models (vision & language), and observe on multiple architecture and scale that routing actually influence the conic structure.

Critical strengths
  • The broad study conducted with different architectural kind of models and multiple metrics evaluation gives more credit to there hypothetical claims.
  • The mechanism is interesting and their intuition is thoroughly discussed, the research question is currently a promising fields. I see links with recent litterature trying to explain anisotropy (eg [1]), the overall goal is to understand the mechanism rather than brand it, those explanatory might actually be important to understand models capability.

[1]: Revisiting Anisotropy in Language Transformers, ICML 2026 (Bernas et al)

Critical weaknesses
  • The framing of the paper is somehow lacking, this is clearly solvable, but I wonder why they did not use common anisotropy framing, as much as the tools (eg [2]). Also it would be mandatory for the authors to discuss the fact that there mechanism cannot fully explain conic geometry since other models than transformers have been shown to produce such a conic space (their mechanism assume a transformer model) [3,4].
  • The paper is not smooth to read honestly and I did not find clearly relatives metrics that could precisely states how much concrete impact did these modifications had on the geometry. Opening the cone depends on many factors that might have been overlooked, this is why I recommand using anisotropy based methods that can give relative contribution and interesting impacts. You might as well try to produce a statistical test against the random baseline that appeared to have much lower impact (such observations fit with the theory and would give it stronger validation). The issue here is more about the scale, how important actually is this mechanism ? The paper does not really answer this.

[2]: IsoScore: Measuring the Uniformity of Embedding Space Utilization, ACL 2022 (Rudman el al)

[3]: How Contextual are Contextualized Word Representations?, EMNLP 2019 (Ethayarajh)

[4]: Anisotropy is Not Inherent to Transformers, NAACL 2024 (Machina et al)

Questions for authors
  • It might help to uniformise the notation for attention ("P" then "a" at some point, + in the litterature, people generally use "A").
  • Citing and updating the related works with some anisotropy litterature seems necessary. I recommend that authors carefully read the state-of-the-art papers (particularly the four I cited in my review, which were not included in the paper) to better follow the conventions of the field.
Reviewer_ZYyhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how attention routing shapes the directional concentration of transformer representations. With model weights and inputs fixed, it suppresses attention to selected token positions and measures changes in cosine similarity. Experiments in SD3 and language models show that suppression can either increase or decrease concentration, depending on the recipient, layer, and model. The paper derives the corresponding attention-output update and investigates how downstream propagation affects its geometric consequences.

Critical strengths
  • The experiments show that suppressing attention to structural tokens does not uniformly reduce directional concentration. The opposite responses of special and padding recipients in SD3, together with the depth-dependent sign reversals in language models, show that the effect depends on which route is suppressed and where it is measured
  • The main observations are supported by checks across models, datasets, mask strengths, and token inclusion. The spectral analysis also shows that reduced cosine concentration does not necessarily imply greater dispersion in centered spectral measures.
  • The analysis connects routing suppression to an explicit value-replacement update and examines where its first-order approximation fails. The contrasting component interventions in GPT-2 and Qwen3 further show that shared updates alone do not explain all observed responses.
Critical weaknesses
  • Structural-recipient and control interventions differ in removed attention mass, selected-head count, and update magnitude. Their effect sizes therefore cannot be attributed to recipient identity alone. Appendix A shows that attention mass alone does not explain the SD3 sign split, but the unmatched perturbations still limit what can be concluded from the smaller control effects.
  • The connection between the local response analysis and the depth-dependent sign reversals is not fully developed. The four-site diagnostics show that nonlinear propagation can substantially change the response, but do not examine how it contributes to the reversals in the layer sweep. The analysis is informative at the selected sites, while its explanation of the broader depth pattern remains limited.
  • The SD3 special-versus-padding comparison is reported at a layer and denoising step selected for the clearest separation. The CLIP-specific comparisons corroborate the result at that site, and the temporal audit tracks its persistence after the initial edit. These results leave the prevalence of the sign split across other layers and intervention times unclear.
Questions for authors
  • How much of the difference between structural-recipient and control effects remains at comparable removed attention mass or update magnitude? A limited comparison would help assess whether perturbation size accounts for the gap.
  • For a representative intervention whose response reverses downstream, please clarify whether the first-order prediction also reverses or whether nonlinear propagation drives the change. The existing decomposition could help connect the local analysis to the depth-sweep results.
  • The additional SD3 results would be useful for interpreting the selected site. If available, summarize whether the special-versus-padding sign split also appears at other layers or in independently edited denoising steps.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.