acceptodds
Under review as a conference paper at ICLR 2027

TopoSteer: Topological Layer Selection for Activation Steering in LLMs

Abstract

Activation steering controls large language models at inference time by adding a vector to their internal representations. Most work asks how to build that vector, and far less attention has gone to where to add it, even though steering at a poorly chosen layer can be worse than not steering at all. Prior selectors choose a layer without ever steering the model to test it: they score how separable the target behavior is at that layer, and how consistently the individual examples point along the steering direction. We argue that what matters is not the separability of the two activation clouds, nor the alignment of individual displacements with their mean, but the extent of the cloud of per-example displacements that mean-difference steering averages. We propose a topology-based layer selection method, called TopoSteer, which scores a layer by the total persistence of its displacement cloud, computed as the length of the cloud's Euclidean minimum spanning tree. We show that the score focuses on the displacement cloud's relative geometry, while class separation affects it only through standardization. We also provide a theoretical explanation for our method by proving that, as example-level variation increases, TopoSteer and discriminability move in opposite directions, so the two criteria can rank layers in opposite ways. Across multiple language models and steering tasks, TopoSteer provides a principled alternative to discriminability- and separability-based layer selection and yields more effective steering without requiring held-out validation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Model Alignment, Model Steering, Topological Data Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_XxMXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper provides an interesting perspective on activation steering by examining the geometry of per-example displacement vectors rather than focusing only on class separability or directional consistency. The analysis of standardization effects and the answer-polarity axis is informative and helps explain why different layer-selection criteria can produce different rankings. However, the paper provides limited insight into why the proposed connectivity measure predicts successful steering, rather than simply capturing another form of displacement dispersion. A more direct empirical connection between displacement geometry and steering effectiveness would substantially strengthen the paper's informativeness.

Critical strengths
  1. This paper tackles a real problem in activation steering: how to choose which layers to intervene on without expensive held-out steering tests. LayerNavigator relies on discriminability and directional consistency. TopoSteer instead looks at the geometry of per-example displacement vectors. That gives a useful alternative view of what information layer selection can use. The setup in Section 3.2 is simple, and the theory in Section 5 helps show how their score differs from existing criteria.
  2. The experiments cover five instruction-tuned models, six persona-steering tasks, several layer budgets, and both simultaneous and sequential steering. The comparisons with LayerNavigator and W2S-Multi are relevant, especially because W2S-Multi needs extra evaluation-based selection. The big gains from the unit-sphere variant on Qwen-3B are encouraging. I also like the answer-polarity axis analysis and the bridge share idea, which helps diagnose cases where MST length is dominated by one long edge. The limitations in Appendix O also show the authors are aware of possible failure modes.
Critical weaknesses
  1. A central motivation of the paper is that the extent and connectivity of the displacement cloud provide useful information for selecting effective steering layers. However, the theoretical analysis mainly establishes how the proposed score differs from discriminability, rather than explaining why larger connectivity should indicate better steering effectiveness. In particular, Theorem 1 shows that total (H_0) persistence increases while discriminability decreases under a controlled variation of example-level residuals. This does not establish that the resulting increase in connectivity corresponds to an improvement in actual steering performance. Similarly, the geometric analyses in Section 5 characterize how different scoring criteria respond to changes in representation geometry, but do not directly establish which geometric properties predict successful steering. Although Table 1 and Figure 3 demonstrate that TopoSteer can select better-performing layers in several settings, these results do not isolate whether the improvement is driven by displacement extent, connectivity structure, or other properties of the selected layers. This matters because the proposed geometric interpretation is central to the paper's contribution. I would encourage the authors to report layerwise correlations between MST-based scores and actual single-layer steering performance across models and tasks. An analysis of layers with high connectivity scores but poor steering performance would also help clarify the conditions under which the proposed criterion is reliable.

  2. The distinctive benefit of MST-based connectivity over simpler geometric statistics remains unclear. The proposed total (H_0) persistence score is equivalent to the total length of the Euclidean MST. While the paper establishes that this measure captures information different from discriminability and directional consistency, it does not sufficiently demonstrate that MST connectivity provides useful predictive information beyond simpler measures of displacement dispersion. For example, centered RMS displacement, mean pairwise Euclidean distance, and mean nearest-neighbor distance could also characterize important aspects of the displacement cloud without relying on persistent homology. The sum-of-norms analysis in Appendix O is useful, but it does not fully address this concern, particularly because centered dispersion measures share the translation invariance emphasized in the paper. This concern is reinforced by the substantial difference between the original and unit-sphere variants in Table 1. On Qwen-3B, the original TopoSteer score underperforms LayerNavigator at K=2 and K=3, whereas the unit-sphere variant achieves substantial improvements. Since the unit-sphere projection removes individual displacement magnitudes, these results raise an important question about whether the gains arise primarily from displacement extent or from the relative geometry of normalized displacement directions. Furthermore, although the bridge-share diagnostic can identify cases where a single long MST edge dominates the score, it does not directly influence the layer-selection rule in Algorithm 1. Appendix O also identifies settings in which MST connectivity can favor an undesirable mode-split geometry, without establishing how often this occurs in real-model experiments. These issues matter because the paper's distinctive contribution depends on the value of the proposed connectivity measure, rather than on displacement dispersion in general. I encourage the authors to compare MST length with simple translation-invariant geometric baselines under identical steering protocols and to analyze cases where these criteria produce different layer rankings. A clearer explanation of the original versus unit-sphere results and the practical role of bridge share would substantially strengthen the contribution.

Questions for authors

Q1. Could the authors report the Spearman rank correlation between total (H_0) persistence and actual single-layer steering performance across models and tasks? In particular, are there layers with high connectivity scores but poor steering effectiveness, and what geometric properties characterize these cases?

Q2. How does MST-based connectivity compare with centered RMS displacement, mean pairwise Euclidean distance, and mean nearest-neighbor distance under the same extraction data, standardization, and steering protocol? Are there cases where MST and these simpler measures select different layers, and does MST more reliably identify the better-performing layers?

Q3. Table 1 shows that the unit-sphere variant substantially outperforms the original TopoSteer score on Qwen-3B, particularly at K=2 and K=3. What explains this difference? Does the improvement primarily come from the geometry of normalized displacement directions rather than variations in displacement magnitude? How does this finding relate to the paper's motivation concerning displacement-cloud extent?

Q4. Does a high bridge-share value predict poor steering performance in the real-model experiments? Have the authors examined whether selected layers with high MST scores are dominated by long inter-cluster edges or outliers? Could incorporating bridge share into the selection procedure improve reliability without requiring additional validation data?

14 comments
AuthorsOpenReview ↗

Response 1/12

We sincerely thank the reviewer for their time and careful evaluation of our submission. We greatly appreciate the reviewer's recognition of the real problem of choosing intervention layers without expensive held-out steering tests, the useful alternative view offered by per-example displacement geometry, the simple setup and the theory distinguishing our score from existing criteria, the breadth of the experiments (five models, six tasks, several layer budgets, simultaneous and sequential steering), the relevant comparisons with LayerNavigator and W2S-Multi, and the answer-polarity and bridge-share analyses. We address each concern below. If our clarifications and additional results satisfactorily resolve the reviewer's concerns, we respectfully ask the reviewer to reconsider their assessment. We will incorporate all of these discussions into the revised version.

Weakness 1.1: Theorem 1 shows that total (H_0) persistence increases while discriminability decreases under a controlled variation of example-level residuals. This does not establish that the resulting increase in connectivity corresponds to an improvement in actual steering performance. Similarly, the geometric analyses in Section 5 characterize how different scoring criteria respond to changes in representation geometry, but do not directly establish which geometric properties predict successful steering.

Response: We agree with the author comment on Section 5. However, we would like to clarify that the Theorem 1 was not written to show that one ordering is better. It shows that our score and discriminability can rank the same configurations in opposite ways, so which one to trust must be tested on real models. Sections 5.1 and 5.3 play a similar role. They explain why the two criteria disagree (class separation affects our score through the standardization, and the answer-polarity axis affects it through a single bridge edge), not which one is right. To further address the reviewer concerns. We steered each of the 1,128 layers of the five models on its own and measured, across layers, how well each geometric property predicts the steering gain (full tables in Weakness 1.2). The result show that connectivity predicts steering. Specifically, the mean Spearman correlation is 0.59 for L\mathcal{L} and 0.63 for Ours$_{\ell_2}$ (the same score on unit-length displacements), and both are positive in 28 of the 30 model–task settings. The mean gain rises steadily from the lowest to the highest quarter of L\mathcal{L} (+0.04 to +2.56 points), and the share of layers that improve the behavior rises from 58% to 85%. Additionally, connectivity also predicts steering better than separability. Discriminability gives 0.42 (paired Wilcoxon over the 30 settings, p < 0.001), and the full LayerNavigator score gives 0.52 (p = 0.050). The size of the cloud does not explain the result: the sum of displacement norms, which measures only extent, is negatively related to the gain (−0.53).

So, across real layers, a larger L\mathcal{L} goes with better steering, and what matters is how the displacements are connected, not how large they are. This evidence is a correlation across layers, not a mechanism. It holds on all five models (the correlation of L\mathcal{L} is positive on each). It is stable between validation and test prompts, whose per-layer gains correlate at 0.92.

AuthorsOpenReview ↗

Response 2/12

Weakness 1.2: Although Table 1 and Figure 3 demonstrate that TopoSteer can select better-performing layers in several settings, these results do not isolate whether the improvement is driven by displacement extent, connectivity structure, or other properties of the selected layers. I would encourage the authors to report layerwise correlations between MST-based scores and actual single-layer steering performance across models and tasks.

Response: We thank the reviewer for the insightful suggestion. We report these correlations below. They cover the MST scores and scores that isolate each property the reviewer names: extent, connectivity, and other properties of the layers (separability and depth). We give both the rank (Spearman) and the linear (Pearson) correlation and they lead to the same conclusions.

Setup. We steered each of the 1,128 layers of the five models on its own, with the same vectors and α=1\alpha=1 as in Table 1 of the paper, and measured the gain in aligned-answer probability over the unsteered model on the 100 test prompts. For each of the 30 model–task settings, we correlate each score with the gain across that setting's layers, and we report the mean over the 30 settings.

Table R1. Correlation between each score and single-layer steering gain (mean over 30 settings; "Positive in" = number of settings with a positive correlation).

ScoreProperty it isolatesSpearmanPositive inPearsonPositive in
L\mathcal{L} (Ours)connectivity, with magnitudes0.5928/300.5127/30
L\mathcal{L} on unit displacements (Ours$_{\ell_2}$)connectivity, magnitudes removed0.6328/300.5227/30
Sum of displacement normsextent only−0.531/30−0.443/30
Centered RMS displacementglobal dispersion−0.562/30−0.513/30
Mean pairwise distanceglobal dispersion−0.512/30−0.453/30
Mean nearest-neighbour distancelocal spacing0.5928/300.5127/30
LayerNavigatorseparability + direction0.5227/300.4727/30
Discriminabilityseparability0.4227/300.3826/30
Layer indexdepth0.5328/300.4227/30

Table R2. Spearman correlation by model (mean over six tasks).

ModelL\mathcal{L} (Ours)Ours$_{\ell_2}$LayerNavigatorDiscriminabilityCentered RMS
Llama-3B0.870.880.880.84−0.67
Llama-8B0.720.820.660.42−0.73
Qwen-3B0.420.540.270.14−0.55
Qwen-7B0.420.410.470.46−0.38
Qwen-32B0.490.510.340.24−0.49
AuthorsOpenReview ↗

Response 3/12

Table R3. Mean single-layer gain (points) and share of layers with positive gain, by quarter of each score within each setting (Q1 = lowest 25%, Q4 = highest 25%; 282 layers per quarter).

ScoreQ1Q2Q3Q4
L\mathcal{L} (Ours)+0.04 (58%)+0.51 (62%)+1.80 (82%)+2.56 (85%)
Ours$_{\ell_2}$+0.02 (57%)+0.45 (62%)+1.50 (78%)+2.94 (90%)
LayerNavigator+0.02 (57%)+0.68 (68%)+1.76 (84%)+2.44 (78%)
Discriminability+0.09 (58%)+1.01 (73%)+1.86 (84%)+1.95 (72%)
Centered RMS+2.53 (89%)+1.77 (79%)+0.58 (62%)+0.02 (57%)

First, the sum of displacement norms, which measures only the size of the cloud, is negatively related to steering gain (Spearman −0.53), and so are the two global dispersion measures. Removing the magnitudes from our score does not weaken it; it makes it slightly stronger (0.63 vs 0.59; paired Wilcoxon over the 30 settings, p = 0.006).

Secondly, the connectivity can track the gain. Both MST scores are positive in 28 of the 30 settings (27 under Pearson), and the gain rises steadily from the lowest to the highest quarter of each score. In the top quarter, 85% (Ours) and 90% (Ours$_{\ell_2}$) of layers improve the behavior. L\mathcal{L} predicts the gain better than discriminability (0.59 vs 0.42, p < 0.001) and better than the full LayerNavigator score (0.59 vs 0.52, p = 0.050). Discriminability is least reliable at the top of its ranking, which is where a selector picks layers: its top quarter steers worse than its third quarter (72% vs 84% of layers helpful). Other properties of the layers do not explain the gain either. The mean nearest-neighbour distance gives the same correlation as L\mathcal{L} (0.59), as expected, because every nearest-neighbour edge is also an MST edge. Together, these results show that the useful property is the local spacing of the displacements. Their global spread, which is dominated by the answer-polarity gap (Sec. 5.3), is negatively related to the gain.

Finally, the relation is monotone rather than linear. For every score, the Spearman correlation is larger in absolute value than the Pearson correlation, and most of the gain is in the top half of the ranking (Table R3). The values are also stable: on 200 separate validation prompts, the mean Spearman correlation of L\mathcal{L} is 0.61, and the per-layer gains on the two prompt sets correlate at 0.92.

Reviewer_yT8ROpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

Activation steering has received increased attention as a key methodology to increase alignment and safety of language models. This work asks an important question: Where should I steer the model? Without guidance, steering on a “wrong” layer can be worse than not steering at all. The authors argue that the decision should be guided by the extent of the cloud of per-example displacements that a mean-difference steering averages. They propose a method based on persistent homology scoring a layer by the total H0 persistence of its displacement cloud, computed as the length of its minimum spanning tree. They also provide a theoretical explanation of how the method differs from a discriminability, showing that the two move in opposite directions as example-level variation increases at fixed class means. They the run experiments across models and steering tasks showing that TopoSteer improves over other methods in some but not all settings.

Critical strengths
  1. The paper addresses a practically relevant and neglected question in activation steering: where to steer, rather than how to build the steering vector. The proposed score is cheap and potentially easy to adopt. It reuses the activations already computed for the steering vector and needs no validation runs.
  2. The theoretical part is transparent about what the score can and cannot capture. The assumptions are stated explicitly. The authors acknowledge that standardization reintroduces a dependence on class separation (Prop. 1). They also propose a diagnostic, the bridge share, for one of the method’s failure modes.
Critical weaknesses
  1. It seems to me that for how the score is defined, it neglects the steering vector itself. It decreases as the two classes become more separated (Prop. 1) and increases as individual examples deviate more from the average displacement (Theorem 1). So, all else equal, it favors layers where individual examples agree least with the vector that is actually injected. The paper does not explain why this should lead to better steering. Theorem 1 shows that the proposed score and LayerNavigator’s discriminability term order a family of configurations oppositely, but not that the proposed ordering is the better one. (see questions below).

  2. The main results (Corollary 1, Prop. 1, Theorem 1) are not specific to persistent homology, it is not clear to me why persistent homology is chosen over other measures of spread. The claims would probably hold equally for simpler measures of spread, such as the variance of the displacements, so it is unclear what the topological formulation contributes.

Questions for authors
  1. Which variant is the proposed method, Ours or Ours l2? If I have counted correctly, the reported count (better than LayerNavigator in 10 of 15 settings) appears to be obtained only by taking, in each setting, the better of the two variants; with Ours alone it is 7 wins. How would a user choose between the two? Also, the largest gains (Qwen-3B) come from the unit-sphere variant. This variant is defined only in a table caption and discards the length of each displacement, which is exactly the information the introduction argues is essential. Can the authors specify better the second variant, and how to choose them?

  2. Table 1 omits the unsteered baseline and reports no uncertainty, while many differences are below one percentage point on 100 evaluation examples per task. Can the authors quote uncertainties and quantify a significance to their claims?

  3. Why should a larger spread of per-example displacements around their mean indicate a better layer for mean-difference steering?

9 comments
AuthorsOpenReview ↗

Response 1/5

We sincerely thank the reviewer for their time and careful evaluation of our submission. We are encouraged by the reviewer's recognition that the paper addresses a practically relevant and neglected question (where to steer rather than how to build the vector), that the score is cheap and easy to adopt because it reuses the activations already computed for the steering vector and needs no validation runs, and that the theory is transparent about its assumptions and limits, including Prop. 1 and the bridge-share diagnostic. We address each concern below. If our clarifications and additional results satisfactorily resolve the reviewer's concerns, we respectfully ask the reviewer to reconsider their assessment. We will incorporate all of these discussions into the revised version.

Weakness 1.1: The paper does not explain why this should lead to better steering. Theorem 1 shows that the proposed score and LayerNavigator's discriminability term order a family of configurations oppositely, but not that the proposed ordering is the better one.

Response: We thank the reviewer for the comment. However, we would like to clarify that, the Theorem 1 was not written to show that one ordering is better. It shows that our score and Discriminability can rank the same layers in opposite ways, so which ordering to trust is a question that must be tested on real models. We tested it directly. We steered each of the 1,128 layers of the five models on its own (same protocol as Table 1) and measured how much it raised the aligned-answer probability over the unsteered model on the test prompts. We then asked, for each of the 30 model–task settings, how well each score predicts this gain across layers (Spearman correlation, table below). L\mathcal{L} is the best predictor: its mean correlation is 0.59, positive in 28 of 30 settings, against 0.52 for LayerNavigator and 0.42 for Discriminability (paired Wilcoxon over the 30 settings: p = 0.050 and p < 0.001). The link is also visible directly in the gains. When we split each model's layers into four equal groups by L\mathcal{L}, the mean gain rises steadily from the lowest group to the highest (+0.04, +0.51, +1.80 and +2.56 points), and the share of layers that improve the behavior grows from 58% to 85%. Discriminability does not behave this way: its top group steers worse than its third group (72% vs 84% of layers helpful). The same pattern appears in Table 1, where Ours is significantly better than Discriminability-only selection in 8 of 15 cells and never worse (95% paired bootstrap over test prompts). So, where the two orderings disagree, the ordering of L\mathcal{L} is the one that matches actual steering gain.

Spearman correlation across layers between each score and single-layer steering gain, averaged over the 30 model–task settings. "Positive in" counts the settings with a positive correlation.

ScoreMean ρPositive in
L\mathcal{L} (Ours)0.5928/30
LayerNavigator0.5227/30
Discriminability0.4227/30
AuthorsOpenReview ↗

Response 2/5

Weakness 2: The main results (Corollary 1, Prop. 1, Theorem 1) are not specific to persistent homology, it is not clear to me why persistent homology is chosen over other measures of spread. The claims would probably hold equally for simpler measures of spread, such as the variance of the displacements, so it is unclear what the topological formulation contributes.

Response: We agree that the Corollary 1, Prop. 1 and Theorem 1 are not specific to persistent homology and also hold for the variance. These results were not meant to be the reason for choosing the MST. The reason is that every displacement cloud is split into two groups by the answer-polarity axis (Sec. 5.3), and the MST and the variance react to this split in opposite ways. To test whether this matters, we used three simpler measures in place of L\mathcal{L} under the protocol of Table 1: the variance (centered RMS distance to the mean), the mean pairwise distance, and the mean nearest-neighbour distance. We also correlated each score with the gain from steering each of the 1,128 layers on its own.

ScoreSpearman with single-layer gainPositive in (of 30 settings)Mean gain over unsteered (15 cells, pts)vs unsteered (W / L / T)
Variance (centered RMS)−0.562−0.022 / 3 / 10
Mean pairwise distance−0.512−0.013 / 3 / 9
Mean nearest-neighbour distance0.5928+4.1911 / 0 / 4
L\mathcal{L} (Ours)0.5928+3.739 / 1 / 5

W / L / T: number of the 15 model × KK cells in which the 95% paired-bootstrap interval of the difference is above zero / below zero / contains zero.

From the table above, we can see that the variance does not work as a layer selector. It is negatively related to steering gain, it always picks the first layers, and on average it gives no gain over the unsteered model (−0.02 points). Ours is significantly better than it in 9 of the 15 cells and worse in 1. The mean pairwise distance behaves the same way. The variance fails because it mostly measures the gap between the "Yes"-matching and "No"-matching groups. This gap carries 98% of the variance at the first layer and 63% at the last, but it carries no information about the behavior (Figure 4f). The MST crosses this gap with only one edge: cutting the longest edge splits the cloud exactly into the two groups (350/350) in all 150 clouds we checked, and this edge is at most 3.9% of L\mathcal{L}. As a result, the topological formulation contributes a measure of local connectivity that counts the gap between the groups only once, together with the bridge share, which reports when one long edge dominates the score. We do not claim that the MST is the only statistic that does this. The table shows that the mean nearest-neighbour distance, another local measure, behaves like L\mathcal{L}: it selects the same layers in 28 of 30 settings and ties with Ours in 12 of 15 cells. This is expected, because every nearest-neighbour edge is also an MST edge. The key distinction is local connectivity versus global spread, and the variance falls on the wrong side of it.

AuthorsOpenReview ↗

Response 3/5

Question 1: Which variant is the proposed method, Ours or Ours l2?. How would a user choose between the two?

Response: TopoSteer is the score L\mathcal{L} of Eq. 6 and Algorithm 1 (Ours): the total MST length of the standardized displacement cloud. Ours${\ell_2}$ is an ablation of it. It first normalizes each displacement individually, δ^li=δli/∥δli∥\hat\delta_l^i=\delta_l^i/\lVert\delta_l^i\rVert, and then computes the MST length of the resulting point cloud. We recommend Ours as the default, since it is the score defining TopoSteer and our experiments provide no evidence that the scale-free variant is generally superior. The two variants capture complementary geometric properties: L\mathcal{L} measures the extent and connectivity of the displacement cloud, retaining magnitude information, whereas Lℓ2\mathcal{L}^{\ell_2} removes individual magnitudes and measures directional dispersion. The two scores rank layers very similarly (Spearman 0.97 across layers), and in a direct paired-bootstrap comparison over the 15 simultaneous cells neither dominates: Ours${\ell_2}$ is significantly better in 4 cells, worse in 3, and tied in 8. Most of the difference between them comes from one model. On Qwen-3B, the size of the displacements keeps shrinking in the last layers, so L\mathcal{L} favors middle layers while the scale-free Ours${\ell_2}$ selects the last layers, which steer much better there. Without Qwen-3B, the two scores give almost the same mean gain over the unsteered model (+4.99 for Ours vs +4.73 for Ours${\ell_2}$ over the remaining 12 cells). This distinction is also supported by our theoretical analysis: Remark 2 establishes that Lℓ2\mathcal{L}^{\ell_2} is bounded in terms of directional consistency (ClC_l) and tends to zero as Cl→1C_l\to1. Thus, the two variants capture different aspects of the displacement cloud rather than being interchangeable scores. We have now clarified the distinction in Section 3.2 (Remark 1) and included the comparison with LayerNavigator's directional consistency in Section 5.2 (Remark 2).

Reviewer_RJj1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose TopoSteer as a method for selecting layers for mean difference vector steering without held-out activations. Instead, they standardize activations by measuring the topological spread of paired activation differences through their Euclidean minimum spanning tree length. They propose to steer the top-scoring layers with raw mean difference vectors. The paper shows that this score captures residual structure rather tahn mean displacement or class separation. Experiments across multiple persona tasks and five instruction-tuned models compare TopoSteer with other baselines (that use discriminability, random-layer, or LayerNavigator).

Critical strengths
  1. answer-polarity analysis: the paper identified and analyzes a yes/no polarity axis in contrastive persona data. this axis splits the displacement cloud into two parts that carry little class information, which offers a useful geometric finding (beyond the proposed layer selector).
  2. layer-selection signal beyond separability: the proposed method outperforms discriminability-only and random mid-stack selection in the simultaneous steering results, without requiring additional forward passes. this supports the narrower claim that geometric structure contains information missed by separability scores.
Critical weaknesses
  1. The main reported gains rely on an undefined variant? the reported gains over LayerNavigator come from "Oursℓ2\text{Ours}_{\ell_2}" which is undefined in the paper. TopoSteer performs roughly at parity with LayerNavigator. The author also switch between variants in its claims, and the reported results lack confidence intervals, making its central efficacy claim difficult to assess.
  2. No demonstrated link between MST length and steering success: the paper does not show that layers with longer minimum spanning trees produce better steering, report the proposed bridge-share diagnostic at selected layers, or compare the score with simpler dispersion measures. As a result, the theory explains how the selectors differ but does not establish why the MST score should be preferred. It would be useful to report the Spearman correlation between L (and D, C) and single-layer steering gain across all layers, tasks and models
Questions for authors
  1. What is "Oursℓ2\text{Ours}_{\ell_2}"? Which variant do the abstract, Sec. 4, Table 2 and the conclusion call "TopoSteer", and how should a practitioner choose between them without validation?
  2. Can you give paired bootstrap CIs (or sign tests) over test items for each Table 1 cell, and average the Random baseline over several seeds?
  3. What is the per-layer rank correlation between L and single-layer steering gain across all layers, tasks and models? What are the bridge shares at the selected layers?
8 comments
AuthorsOpenReview ↗

Response 1/8

We sincerely thank the reviewer for their careful evaluation and constructive feedback. We appreciate the reviewer's recognition of the answer-polarity analysis as a useful geometric finding beyond the layer selector, and the evidence that geometric structure carries layer-selection signal missed by separability scores, without additional forward passes. We address each concern below. If our clarifications and additional results satisfactorily resolve the reviewer's concerns, we respectfully ask the reviewer to reconsider their assessment. We would be happy to further discuss any remaining questions. We will incorporate all of these discussions into the revised version.

Weakness 1.1: The main reported gains rely on an undefined variant? the reported gains over LayerNavigator come from Our_l2 which is undefined in the paper.

Response: We thank the reviewer for this point. We agree that the submission defined this variant only in the caption of Table 1. Specifically, TopoSteer is the score L\mathcal{L} of Eq. 6 and Algorithm 1 (Ours): the total MST length of the standardized displacement cloud. Ours${\ell_2}$ is an ablation of it where we first normalizes each displacement individually, δ^li=δli/∥δli∥\hat\delta_l^i=\delta_l^i/\lVert\delta_l^i\rVert, and then computes the MST length of the resulting point cloud. Thus, Ours${\ell_2}$ removes the information about magnitudes and focuses on their directional arrangement. We have now clarified the distinction in Section 3.2 (Remark 1) and included the comparison with LayerNavigator's directional consistency in Section 5.2 (Remark 2).

AuthorsOpenReview ↗

Response 2/8

Weakness 1.2: TopoSteer performs roughly at parity with LayerNavigator. The author also switch between variants in its claims, and the reported results lack confidence intervals, making its central efficacy claim difficult to assess.

Response: We thank the reviewer for this comment. For the confidence intervals experiments, we re-ran every cell of Table 1 with per-prompt logging and computed 95% paired-bootstrap intervals (600 test prompts per model, resampled within task, 10,000 draws). A cell is a win or loss only if the interval excludes zero; otherwise it is a tie. Per-cell intervals are in Table R2 and the counts are in Table R2c.

Table R2. Simultaneous steering: aligned-answer probability (%), mean over six tasks, and paired differences in points with 95% CI.

ModelKBaseDiscriminabilityRandomLNOursOurs$_{\ell_2}$Ours − LNOurs$_{\ell_2}$ − LN
Llama-3B166.7372.7969.5273.7972.7972.79−1.01 [−1.42, −0.59]*−1.01 [−1.43, −0.59]*
Llama-3B266.7379.7271.3981.3881.3881.380 (same layers)0 (same layers)
Llama-3B366.7378.2172.9575.1280.4977.40+5.37 [+3.83, +7.00]*+2.28 [+1.71, +2.88]*
Llama-8B177.6279.7379.5081.2683.4482.70+2.18 [+1.59, +2.78]*+1.44 [+0.60, +2.27]*
Llama-8B277.6281.7081.4384.8784.2687.09−0.61 [−1.24, +0.02]+2.22 [+0.35, +4.17]*
Llama-8B377.6284.1183.1386.4385.9184.73−0.52 [−1.00, −0.07]*−1.69 [−3.35, −0.05]*
Qwen-3B164.2660.9965.2464.2364.8171.51+0.58 [−0.43, +1.63]+7.28 [+5.42, +9.21]*
Qwen-3B264.2659.5067.5664.1163.0376.95−1.07 [−2.64, +0.42]+12.84 [+9.63, +16.05]*
Qwen-3B364.2656.3469.2763.6361.0079.43−2.62 [−4.35, −0.97]*+15.81 [+11.36, +20.22]*
Qwen-7B185.1385.9185.5085.5785.7785.74+0.20 [−0.50, +0.90]+0.17 [−0.63, +0.95]
Qwen-7B285.1385.8585.6886.1485.7485.53−0.40 [−1.56, +0.78]−0.61 [−2.09, +0.85]
Qwen-7B385.1385.5986.3486.4785.7885.32−0.69 [−2.68, +1.25]−1.15 [−3.22, +0.89]
Qwen-32B182.6882.7883.2083.4283.2483.23−0.18 [−0.71, +0.36]−0.19 [−0.86, +0.47]
Qwen-32B282.6882.4783.2183.0883.9183.69+0.83 [−0.13, +1.79]+0.61 [−0.64, +1.84]
Qwen-32B382.6881.9283.4981.7583.6683.64+1.91 [+0.48, +3.37]*+1.90 [−0.18, +4.07]

Table R2c. Counts of cells (wins / losses / ties). Simultaneous: 15 cells. Sequential: 10 cells (KK = 2, 3; greedy, re-extraction, no layer repeats; Random over 5 seeds).

ComparisonSimultaneousSequential
Ours vs Disc8 / 0 / 76 / 0 / 4
Ours vs unsteered9 / 1 / 55 / 1 / 4
LayerNavigator vs unsteered6 / 1 / 85 / 0 / 5
Ours vs Random7 / 3 / 53 / 3 / 4
Ours vs LayerNavigator3 / 3 / 91 / 2 / 7
Ours$_{\ell_2}$ vs LayerNavigator6 / 2 / 73 / 2 / 5
Ours$_{\ell_2}$ vs unsteered11 / 0 / 45 / 0 / 5

From the table R2, mean gain over the unsteered model across the 15 simultaneous cells: Disc +1.89, Random +2.54, LayerNavigator +3.46, Ours +3.73, Ours$_{\ell_2}$ +6.12.

We agree that L\mathcal{L} is on par with LayerNavigator in final steering accuracy (3 wins, 3 losses, 9 ties; sequential 1 / 2 / 7). Parity does not weaken the main contribution. TopoSteer reaches this level with a single geometric score and no separability term. It is significantly better than separability alone (Discriminability) in 8 of 15 cells and never worse. It improves on the unsteered model in more cells than LayerNavigator (9 vs 6), with a higher mean gain (+3.73 vs +3.46 points). At the level of single layers, L\mathcal{L} also predicts steering gain better than the LayerNavigator score (Spearman 0.59 vs 0.52 over 1,128 layers; Table A in Weakness 2.1). The scale-free ablation Ours$_{\ell_2}$ is ahead of LayerNavigator (6 wins, 2 losses), but most of this lead comes from Qwen-3B (3 wins, 2 losses without it), so we do not treat it as the main result.

AuthorsOpenReview ↗

Response 3/8

Weakness 2.1: No demonstrated link between MST length and steering success: the paper does not show that layers with longer minimum spanning trees produce better steering, report the proposed bridge-share diagnostic at selected layers, or compare the score with simpler dispersion measures. As a result, the theory explains how the selectors differ but does not establish why the MST score should be preferred. Question 3: What is the per-layer rank correlation between LayerNavigator and single-layer steering gain across all layers, tasks and models? What are the bridge shares at the selected layers?

Response: We thank the reviewer for asking for this evidence. We answe this by providing three analyses: (i) the simpler dispersion measures as layer selectors under the protocol of Table 1, (ii) the relationship between each layer score and single-layer steering gain over all 1,128 layers, and (iii) the bridge-share diagnostic at the selected layers.

Setup. We steered each of the 1,128 layers on its own (5 models × 6 tasks, same vectors and α=1\alpha=1 as Table 1). Gain is the change in aligned-answer probability over the unsteered model on the test prompts. All scores are computed on the same standardized displacement clouds (700 training pairs per task).

(i) Simpler dispersion measures as layer selectors. We used each measure in place of L\mathcal{L} under the protocol of Table 1 (simultaneous steering). Centered RMS displacement and mean pairwise distance are measures of global spread (distance of each displacement to the cloud mean, and average distance between all pairs of displacements); mean nearest-neighbour distance is a local measure (average distance from each displacement to its closest neighbour). Each selector ranks the layers by its own score and steers the top KK; Table 1 reports the resulting aligned-answer probability, and Table 2 the paired-bootstrap comparison between selectors over the 15 models × KK cells.

Table 1. Aligned-answer probability (%), mean over six tasks.

ModelKUnsteeredLayerNavigatorOurs (L\mathcal{L})Mean NN dist.Centered RMSMean pairwise
Llama-3B166.7373.7972.7972.7966.7366.74
Llama-3B266.7381.3881.3881.3866.7466.74
Llama-3B366.7375.1280.4980.4966.7566.75
Llama-8B177.6281.2683.4483.4477.6277.59
Llama-8B277.6284.8784.2684.2677.5977.59
Llama-8B377.6286.4385.9185.9177.5677.56
Qwen-3B164.2664.2364.8165.5364.2764.27
Qwen-3B264.2664.1163.0366.1364.1564.17
Qwen-3B364.2663.6361.0064.0364.1964.28
Qwen-7B185.1385.5785.7785.7785.1185.07
Qwen-7B285.1386.1485.7485.7485.1085.14
Qwen-7B385.1386.4785.7885.7885.1285.15
Qwen-32B182.6883.4283.2483.2282.6882.68
Qwen-32B282.6883.0883.9183.9182.7382.73
Qwen-32B382.6881.7583.6683.6882.6982.69
Mean gain over unsteered—+3.46+3.73+4.19−0.02−0.01

Table 2. Paired-bootstrap comparison over the 15 cells (wins / losses / ties; a win or loss means the 95% interval excludes zero).

ComparisonWins / losses / ties
Ours vs Centered RMS9 / 1 / 5
Ours vs Mean NN distance0 / 3 / 12
Ours vs unsteered9 / 1 / 5
Mean NN distance vs unsteered11 / 0 / 4
Centered RMS vs unsteered2 / 3 / 10
Mean pairwise vs unsteered3 / 3 / 9

From the Table 1, we can see that centered RMS and mean pairwise distance pick the earliest layers (0–2 for RMS; within 0–7 for pairwise) and give no gain: every cell stays within 0.11 points of the unsteered model. These global measures are dominated by the polarity gap (98% of the variance at the first layer, Table E), so they rank the layers where steering helps least at the top. The MST counts this gap only once, as one edge, so L\mathcal{L} measures how pairs spread within each answer group. Mean nearest-neighbour distance behaves like L\mathcal{L}, as expected, since each point's nearest-neighbour edge is an MST edge: the two correlate at 0.999 across layers, select the same top-KK layers in 28, 29 and 28 of 30 settings, and tie in 12 of 15 cells (the nearest-neighbour selector is ahead in the other 3, all on Qwen-3B).

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response 1/2

Dear Reviewers, AC, SAC and PC,

We thank you for your time and for the constructive reviews. The reviewers found the problem relevant and under-studied, namely choosing where to steer and not only how to build the steering vector (yT8R, XxMX); the score cheap and free of validation runs (RJj1, yT8R); the theory clear about its assumptions and about how the score differs from existing criteria (yT8R, XxMX); and the answer-polarity and bridge-share analyses useful beyond the selector (RJj1, XxMX). We ran every analysis the reviewers requested and uploaded a revised PDF, with all changes in blue. Each reviewer thread gives the full tables. Here we summarize the shared concerns, what we found, and what changed in the paper.

In brief

  • TopoSteer is the score L\mathcal{L}. The unit-sphere version, Ours$_{\ell_2}$, is an ablation, now defined in the paper.
  • With per-prompt confidence intervals, TopoSteer matches LayerNavigator in steering accuracy and is never significantly below discriminability-only selection. We narrowed our claims to match this.
  • Over 1,128 single-layer runs, L\mathcal{L} tracks steering gain more closely than LayerNavigator and discriminability, while simpler measures of global spread fail. The useful property is the local spacing of the per-example displacements.

1. Which score is the method (Reviewer RJj1- Weakness 1.1, Question 1; Reviewer yT8R - Question 1; Reviewer XxMX - Weakness 2.2, Question 3)

TopoSteer is L\mathcal{L} of Eq. 6 and Algorithm 1: the total length of the minimum spanning tree (MST) of the standardized displacement cloud ("Ours" in Table 1). Ours${\ell_2}$ computes the same length after scaling each displacement to unit length. The submission defined it only in a table caption, and Figure 3 labelled it "Ours". In the revision, Remark 1 (Sec. 3.2) defines it as an ablation, and Remark 2 (Sec. 5.2) with Proposition 5 (App. D.2) bounds it by how widely the displacement directions spread. All claims about TopoSteer now refer to L\mathcal{L} alone, and Ours${\ell_2}$ is reported separately. The two give similar mean gains on four of the five models (+4.99 points for TopoSteer and +4.73 for the ablation, over the 12 cells without Qwen-3B). On Qwen-3B the ablation is better, by +6.70, +13.91 and +18.43 points at KK = 1, 2 and 3, because the displacements shrink in the last layers and L\mathcal{L} counts their size (App. J.4).

2. Uncertainty and the size of the claim (Reviewer RJj1 - Weakness 1.2, Question 2; Reviewer yT8R - Question 2)

We re-ran every cell of Table 1 with per-prompt logging, added the unsteered model (Base), drew Random from all layers over 10 seeds (5 for sequential steering), and computed 95% paired-bootstrap intervals over the 600 test prompts of each model (App. J.3). A difference counts as a win or a loss only if its interval excludes zero. Over the 15 simultaneous cells:

  • TopoSteer vs LayerNavigator: 3 wins, 3 losses and 9 ties (1 / 2 / 7 over the 10 sequential cells).
  • TopoSteer vs discriminability-only selection: 8 wins, no losses and 7 ties.
  • Significant gain over the unsteered model: 9 cells for TopoSteer, 6 for LayerNavigator and 12 for Random (mean gains +3.73, +3.46 and +2.54 points).
  • TopoSteer vs Random: 7 wins and 3 losses (Qwen-3B at KK = 2 and 3, Qwen-7B at KK = 3).

TopoSteer is therefore on par with LayerNavigator in final accuracy, not better. We removed the claim of more effective steering from the abstract, the introduction and Sec. 4.

3. Does the score predict steering? (Reviewer RJj1 Weakness 2.1, 2.2, Question 3; Reviewer yT8R - Weakness 1.1, Question 3; Reviewer XxMX W1.1–W1.3, Q1)

Theorem 1 shows that L\mathcal{L} and discriminability can order layers in opposite ways; it does not show which order is better. We tested this by steering every layer of the five models on its own (1,128 runs over 30 model–task settings; Sec. 5 and App. J.4).

  • The mean Spearman correlation between score and single-layer gain is 0.59 for L\mathcal{L} (positive in 28 of 30 settings), 0.52 for LayerNavigator and 0.42 for discriminability (paired Wilcoxon tests: p = 0.050 and p < 0.001).
  • The mean gain rises from +0.04 points in the lowest quarter of layers ranked by L\mathcal{L} to +2.56 points in the highest quarter, where 85% of the layers improve the behavior.
  • The result holds on 200 separate validation prompts per task (0.61), and the size of the displacements does not explain it: the sum of displacement norms correlates negatively with the gain (−0.53).

These are correlations across layers, not a mechanism. A high L\mathcal{L} also does not guarantee a gain: 41 of the 282 top-quarter layers (15%) lower the behavior, mostly by small amounts (median −0.63 points), with one substantive failure on Qwen-3B (layers 21–23).

AuthorsOpenReview ↗

General response 2/2

4. Why the MST rather than a simpler measure of spread (Reviewer yT8R - Weakness 2; ReviewerXxMX - Weakness 2.1, 2.2, Question 2, Question 4.1–4.3; Reviewer RJj1 Weakness 2.1, Question 3)

We agree that Corollary 1, Proposition 1 and Theorem 1 also hold for simpler measures of spread, such as the variance. The difference shows up in the data. Used as selectors under the same protocol, centered RMS distance and mean pairwise distance pick the first layers, give no gain over the unsteered model (−0.02 and −0.01 points), and correlate negatively with single-layer gain (−0.56 and −0.51). Both are dominated by the answer-polarity gap of Sec. 5.3, which carries 98% of the variance of the cloud at the first layer. Mean nearest-neighbour distance behaves like L\mathcal{L} (correlation 0.59; tied in 12 of the 15 cells), because each nearest-neighbour edge is an MST edge. So the useful property is local spacing rather than spread, and the paper now describes the score this way instead of as the "extent" of the cloud.

The full tree adds the bridge share blb_l, the share of L\mathcal{L} carried by its longest edge. At the selected layers, blb_l is at most 1.03%, so no single edge drives the score. Used as a tie-break that needs no validation data (the lowest blb_l among the five highest layers by L\mathcal{L}), it raises the mean single-layer gain at K=1K = 1 from +2.73 to +4.19 points (+2.70 to +4.16 on the validation prompts). We tested this rule only for single layers (App. O.6).

5. A clearer example in Figure 3 (RJj1 W1.1; XxMX W1.2)

Figure 3 now shows a case where TopoSteer itself selects different layers. On Qwen-3B (Religion Following, K=3K = 3), discriminability and LayerNavigator select layers 21–23 (50.13%), and TopoSteer selects layers 33–35 (68.61%), a gain of +18.49 points (95% interval [+11.47, +25.81]). Such differences are common but do not always favor TopoSteer: over all 90 model–task–KK combinations, it selects layers different from both baselines in 63, and it is significantly above both in 15 and below both in 7.

Summary of changes in the revised PDF (all in blue)

  • Abstract, Sec. 1 and Sec. 4: claims narrowed to parity in steering accuracy and a score that tracks single-layer gain; the score is described as a measure of local spacing.
  • Sec. 3.2 (Remark 1), Sec. 5.2 (Remark 2) and App. D.2 (Proposition 5): the unit-sphere ablation defined and analyzed.
  • Table 1: re-run with per-prompt logging, a Base column and Random over seeds. App. J.3 (Tables 9–11): paired-bootstrap intervals, win/loss counts and Random per seed.
  • Sec. 4 and Figure 3: the new example and counts above.
  • Sec. 5 and App. J.4 (Tables 12–18): single-layer sweep, alternative selectors, failure analysis, and bridge share at the selected layers.
  • Sec. 5.3: global spread versus local spacing.
  • App. O.6 (Tables 23–25): MST edge statistics and the bridge-share tie-break.
  • App. I: the per-prompt evaluation metric. App. L: selected layers from the re-run.
  • Limitations: a high L\mathcal{L} does not guarantee a gain.
  • The score, the algorithm and the existing theoretical results (Corollary 1, Proposition 1, Theorem 1) are unchanged.

We hope the new evidence and the revision resolve the shared concerns, and we would be glad to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.