acceptodds
Under review as a conference paper at ICLR 2027

Topology Before Steering: Localising and Gating Cyclic Concepts with Persistent Homology

Abstract

As Large Language Models become increasingly widespread, the need for effective safeguards and methods for controlling model behavior has grown rapidly. One prominent approach is activation steering, which directly intervenes in a model’s residual stream by adding a steering vector to encourage or suppress a target concept. Yet two fundamental questions arise before steering and have received comparatively little attention: where does the target concept reside within the prompt, and should we intervene on a given input at all? We study these questions for cyclic concepts, whose activations trace loop-like structures in representation space. We show that tools from Topological Data Analysis, particularly persistent homology, provide a principled way to address both problems. By analyzing activations across prompts, persistent homology identifies where a cyclic concept emerges and which tokens carry it. For new inputs, the same topological signal then determines whether the concept is present and whether steering should be activated, allowing the intervention to remain off when it is unnecessary. This yields a topology-aware approach to activation steering that jointly addresses both where and when to steer.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Topological Data Analysis, steering, gating

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ekyYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a persistent-homology-based framework for activation steering, addressing intervention localization and conditional activation. The authors analyze activation point clouds across prompts, use first homology (H1) to identify cyclic concepts, and convert recovered circular coordinates into an inference-time gating mechanism. Experiments on multiple language models, cyclic concept tasks, and visual settings evaluate the effectiveness of topology-based localization and conditional intervention.

Critical strengths

The problem addressed in this paper is meaningful. Rather than designing another steering vector, the authors investigate where and when an intervention should be applied, which is a practically relevant direction. The paper also includes Gaussian-null calibration, matched-variance distractors, and rate-matched random gating controls. The authors explicitly report several negative findings, which helps distinguish topological structure from activation magnitude and the effects of simply reducing intervention frequency.

Critical weaknesses

This is my most direct concern with the paper. Section 4.2 and Appendix I show that, across the six well-posed model-task configurations, always-on steering achieves an average target percept of 0.611, whereas gated steering achieves only 0.340. The gate therefore retains approximately 56% of the original intervention effect. This is substantially below the authors' pre-registered target of 90% (yourself say that, and i dont know why "it retains 0.56 of the always-on effect, below our pre-registered 0.90"). More importantly, the method is not consistent across cyclic concepts. While dense_clock demonstrates relatively strong selectivity, the months task exhibits inverted target/off-task firing rates in two model configurations. Consequently, I do not find the current evidence sufficient to establish that the proposed method can reliably determine when steering should be applied. Its effectiveness appears concentrated in particular cyclic tasks rather than consistently generalizable.

Questions for authors

/

6 comments
AuthorsOpenReview ↗

Rebuttal (1/3): Weakness 1 Part 1

[W1-Part 1] This is my most direct concern with the paper. Section 4.2 and Appendix I show that, across the six well-posed model-task configurations, always-on steering achieves an average target percept of 0.611, whereas gated steering achieves only 0.340. The gate therefore retains approximately 56% of the original intervention effect. This is substantially below the authors' pre-registered target of 90% (yourself say that, and i dont know why "it retains 0.56 of the always-on effect, below our pre-registered 0.90").

We thank the reviewer for reading the appendix this closely. Both numbers are correct, and the sentence quoted is ours. But 0.56 and 0.90 are not two values of the same quantity: 0.90 is a setting that we chose, and 0.56 is a measurement. In one sentence: 0.90 is how often the gate was set to fire on the prompts used to build it, and 0.56 is how often it actually fired on new prompts.

More details on what 0.56 is: It is a ratio of two measured percepts over the six well-posed cells: 0.340 (gated) divided by 0.611 (always-on). In words: of the steering successes obtained by patching every prompt, 56% remain when only the prompts selected by the gate are patched. The patch is identical in both arms. A prompt on which the gate fires therefore gives exactly the always-on answer, and a prompt on which it does not fire gives the unsteered answer. So 0.56 measures one thing only: how often the gate fired on the target prompts (its recall, or sensitivity). Table 3 shows it: the mean firing rate on target prompts over the six cells is 0.556, and 0.340/0.611 = 0.556. One cell as an example: on Qwen dense_clock, always-on steers 1.000 of the prompts, the gate fires on 0.796 of them, and the gated percept is 0.796.

0.56 does not mean that the steering is 56% as strong, or that the method is right 56% of the time. Where the gate fires, the effect is 100% of always-on. It means that the gate stayed silent on 44% of the prompts it should have accepted.

More details on what 0.90 is: It is not a result. It appears in two places, both fixed before the experiment was run:

  • A setting of the gate. Its two thresholds are chosen so that 90% of the calibration prompts are accepted (§3.3, coverage 0.90), in the way one fixes a confidence level.
  • A pass/fail bar. In our analysis script we wrote in advance that the gate should keep at least 90% of the always-on effect. We took 90% because it is the coverage: if new prompts behaved like the calibration prompts, the gate would fire on 90% of them, and 90% of the effect would be kept.

The gate basically decides whether a prompt looks enough like the examples it was built from. Coverage (the 0.90), is how strict that decision is:

  • At 0.90 (the paper's setting), it accepts the 90% most typical examples and rejects the 10% most unusual ones.
  • At 1.00, it accepts anything that looks at least as typical as the most unusual example it was built from.

So raising the number makes the gate looser

The distance between 0.56 and 0.90 is a generalisation gap: an acceptance region tuned on the calibration prompts is too tight for prompts it has not seen.

Is 56% a bad trade? We propose new evidence. Always-on scores 100% on this measure by definition, because it patches everything, including the prompts it should leave alone. Retention must therefore be read together with the damage on off-task prompts. We extended the grid from 6 to 15 well-posed cells (five cyclic concepts, three models, 5 seeds), with evaluation templates that were not used for calibration:

armtarget perceptoff-task correctness
unsteered0.0000.561
always-on0.6960.251
gated (ours)0.3160.498
random gate, same firing rate0.1610.480

Always-on keeps the whole effect and loses 0.31 of off-task correctness. The gate keeps 45% of the effect and avoids 80% of that loss (gated is above always-on on off-task correctness in 15/15 cells). A random gate that fires equally often keeps 23% of the effect; ours is ahead of it in 14/15 cells, where the submission could report 4/6.

AuthorsOpenReview ↗

Rebuttal (2/3): Weakness 1 Part 2

[W1-Part 2] More importantly, the method is not consistent across cyclic concepts. While dense_clock demonstrates relatively strong selectivity, the months task exhibits inverted target/off-task firing rates in two model configurations. Consequently, I do not find the current evidence sufficient to establish that the proposed method can reliably determine when steering should be applied. Its effectiveness appears concentrated in particular cyclic tasks rather than consistently generalizable.

The observation is correct. In the submission the gate is selective on dense_clock and inverted on months for two models, and two tasks cannot show that it works across cyclic concepts. Since then we have widened the test and traced the months failure to its cause. The cause is one that can be detected without labels, which turns an unexplained inconsistency into a stated operating condition.

We run a wider test for the clarifications. Five well-posed cyclic concepts on three models (the grid of Answer 1), with the gate exactly as in the paper and evaluation templates not used for calibration. Firing rate on target / off-task inputs:

taskGemma-2-9BLlama-3.1-8BQwen2.5-3B
clock0.18 / 0.120.58 / 0.100.44 / 0.06
dense_clock0.40 / 0.090.58 / 0.010.02 / 0.00
heading0.22 / 0.020.68 / 0.580.75 / 0.04
hue0.54 / 0.120.71 / 0.460.90 / 0.34
months0.50 / 0.230.33 / 0.040.15 / 0.30

The gate fires more on target than on off-task inputs in 14 of 15 cells (sign test p = 0.001; mean 0.47 against 0.17), and gated steering beats the rate-matched random gate in 14 of 15. The behaviour is therefore not confined to dense_clock. We do not present this as the most accurate detector (a distance test does better, see our response to Reviewer 1bgw), and point 3 shows that it is not always the cycle that the gate reads.

Is months simply not cyclic in these models? It is. We estimated the twelve month centroids at the gated layer, averaging out the prompt template. In their principal plane the months lie around a closed loop in exact calendar order on Gemma and on Qwen, and with a single swap of neighbours (January and February) on Llama. The December to January gap is 1.2 to 1.3 times the average gap between consecutive months (1 for a perfect cycle), and the pairwise distances follow those of a regular 12-gon (r = 0.85, 0.49 and 0.84 for Gemma, Llama and Qwen; permutation p < 0.001 on each). All three models represent the months as a cycle, consistent with Engels et al. (2025).

The cause of the problem: twelve points are too few for the loop to stand out, so the gate was built on another loop.

Persistent homology detects a loop only if it is sampled densely enough to stay open longer than a noise cloud of the same spread. Once the effect of the prompt template is removed, the ratio ρ of the concept's own loop falls with the number of labels per cycle (ρ > 1 means above the null):

tasklabels per cycleGemmaLlamaQwen
dense_clock480.951.541.61
heading360.941.110.92
hue240.790.550.73
clock120.390.760.60
months120.470.130.37

We thinned the 48-point clock to 24 and to 12 labels per cycle, on the same activations. Its ρ falls from 1.54 to 0.81 and 0.89 on Llama, and from 1.61 to 0.74 and 0.83 on Qwen. Sampled like months, the clock is lost like months.

The calibration cloud pools twelve prompt templates, and in the months cloud the template explains as much variance as the month (46–53% against 43–51%). Template clusters are structure that a Gaussian null does not have, so the pooled cloud passes ρ > 1. But its longest loop is not the calendar: the circular coordinate is constant on 107 of the 108 calibration points, and the month explains 12–16% of it. A gate built on this coordinate does not respond to the month. Where it fires is then an accident of each model's geometry, which is why the outcome differs between models: selective on Llama, inverted on Qwen and on Gemma. For the same reason we do not count the Llama cell as a success. (On Gemma, in addition, the patch itself rarely steers months: always-on reaches 0.11.)

AuthorsOpenReview ↗

Rebuttal (3/3): Weakness 1 Part 2 (continued)

The method's own criterion flags these cells. The inverted months cells in the submitted figure come from a gap between §3.3 and the gate grid as run. §3.3 licenses a site only if ρ > 1 and the circular coordinate passes the non-degeneracy check U(θ). The grid applied the first condition only and failed to apply U(θ) for the months, so it never rejected a site. On months, U(θ) ≈ 0.02 on all three models, that is, the coordinate is constant on 107 of 108 calibration points, and a gate built on it does not respond to the month, which is why its firing was selective on one model and inverted on two. The coordinate is degenerate because twelve points per cycle are too few for persistent homology to resolve the loop: the months do form a loop in calendar order in all three models, but once template effects are removed its ratio is 0.13 to 0.47, and thinning the 48-point clock to 12 points lowers its ratio from 1.5–1.6 to 0.8–0.9 in the same way. With the check applied at a stated threshold (the 0.95 quantile of U under the same null), the pipeline declines months on Gemma and Qwen in every seed; the corrected figure (which we added to the pdf) shows these cells as abstentions. Overall it licenses 8 of the 15 cells in a majority of seeds; on these the gate fires on 0.64 of target and 0.12 of off-task inputs and is selective in 8/8.

Why we think the method remains of interest.

  • It addresses a decision that is usually skipped (is the concept read at this site, and should this input be steered?) from unlabelled prompts and against a null, so it is able to answer "no".
  • Where the loop is resolved, every step can be checked. The coordinate is the clock time. The site it selects is the one where patching works (1.00, against 0.00 at the downstream copy that carries the strongest signal, §4.1). With the check enforced, the dense_clock gate transfers to unseen templates on the two models where the loop is resolved (Qwen 0.73 / 0.00, Llama 0.58 / 0.01).
  • The failure now has a cause, a label-free test, and a prediction that can be checked: a calendar cycle sampled densely (the day of the year instead of the month) should be licensed where months is not. The limit is on how densely the concept is sampled, not on the kind of concept.
  • The diagnostic is itself a result: it shows, without a probe, that three models lay the months out in calendar order around a loop.

The revision we propose for the claim. We claim the gate for cycles that are sampled densely enough for their loop to exceed the null once template effects are removed, and the pipeline abstains otherwise. §4.2 will carry the 15-cell table, the licensing decision for every cell and the diagnosis above, with the unchecked months inversion reported next to its abstention. This was added to the pdf.

Both weaknesses concern the gate of §4.2. The review raises no concern about the localisation results of §4.1; the token scan there uses a single aligned template, so the pooling issue does not arise in it. If specific passages were hard to follow, we would be grateful for pointers and will revise them.

We hope that the clarification of the retention measure and the diagnosis of months will satisfy the reviewer and we would appreciate it if we could get your further feedback at your earliest convenience.

Reviewer_1bgwOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2(+1)
Presentation
1
Soundness
1
Read review
Paper summary

Activation steering in language models has gained widespread attention as effective safeguards and controlled outputs are needed with increasing urgency. This paper addresses the question of where to target the steering within the prompt, and whether to intervene at all, for cyclic concepts. Because of their cyclic nature, the authors exploit topological data analysis as a method for identifying persistent loops emerging from cyclic concepts, and use it to inform about activation steering before it intervenes.

Critical strengths
  1. The paper asks relevant questions about how activation steering should be approached: knowing beforehand whether the steering needs to take place at all, and where, would increase its effectiveness greatly. Table 1 shows that "always on" steering loses 20% accuracy over gating it. Proposition 3.1 gives a clear reason why the check should be done before steering and not after.

  2. The authors report their results transparently without hiding negatives, which I think should be encouraged in scientific papers.

Critical weaknesses
  1. The paper's central claim is that persistent homology determines where and whether to steer. There is no evidence in the paper that topology is needed at all: I didn't see any comparison to a non-topological alternative. The only motivation seemingly supporting the use of persistent homology is the fact that cyclic concepts are investigated (though, see below), which however is not at all a sufficient condition to use TDA against other non-topological baselines.

  2. It is unclear how a "cyclic concept" is defined. If they are concepts whose activations generate loops, the argument becomes circular, as one justification to use TDA would have been the presence of cyclic concepts. Moreover, looking at the results, persistence score does not seem to be able to cleanly separate "cyclic" from "non-cyclic" concepts (though again there is no definition of what's cyclic and what is not).

Given these two weaknesses hinder both motivation (why TDA?) and experimental setup (why cyclic concepts, and what are they), I am not reporting other weaknesses in the analysis, as they are dependent on solving these two first.

Questions for authors
  1. How does the gate compare with non-topological gates on the same inputs? For instance, CAST (Lee et al. 2025), cited in the paper, would be a method to test, but also simpler methods, a distance-to-centroid metric, or a kNN detector maybe?

  2. How is a cyclic concept defined at the task level, independently of the representation? On what basis were e.g colors or emotions classified as cyclic?

Minor question:

  1. What is the motivation for choosing PCA to 16 dimensions? And what is the motivation for PCA at all? Ripser would be able to compute persistent homology from the distance matrix at this input sizes. Did you do a sweep of dimensions to check whether PCA is stable here, or needed at all?
12 comments
AuthorsOpenReview ↗

Rebuttal (1/5): Weakness 1

Thank you for your thoughtful review and constructive feedback. We have revised our manuscript in response to your suggestions, with all changes highlighted in blue. Below, we address each of your concerns in detail.

[W1] The paper's central claim is that persistent homology determines where and whether to steer. There is no evidence in the paper that topology is needed at all: I didn't see any comparison to a non-topological alternative. The only motivation seemingly supporting the use of persistent homology is the fact that cyclic concepts are investigated (though, see below), which however is not at all a sufficient condition to use TDA against other non-topological baselines.

Answer:

We thank the reviewer for the careful reading, and for recognising that we report negative results openly; we are going to answer in the same spirit. We agree with the core of this weakness: the submission did not place the topological gate next to non-topological gates, and "the concepts are cyclic, hence TDA" is not a sufficient argument. We have now run the comparison the reviewer asks for (Answer 3). First, what we claim.

We do not claim that topology is necessary, or that it gives the most accurate gate. The paper studies how much of where and whether to steer can be read from the shape of the unsteered activations, given only an unlabelled set of prompts that vary the concept: no labels, no negative examples, no trained probe. Persistent H1 suits that study for three reasons. It needs no supervision. It is scored against a matched null, so "is there structure at this site at all?" has an answer and the method can abstain. And it does not assume that the loop lies in a plane or along a direction.

Non-topological alternatives already in the paper (for "where").

  • Linear probe. Accuracy is at least 0.998 on all 11 tasks, so it cannot rank sites, while donor-patch success on the same tasks ranges from 0.49 to 1.00 (§4.1, Fig. 5 right).
  • Activation variance. On dense_clock, a random-name token with matched variance (1885 vs. 1837) has ratio 0.96 and does not steer; the concept token has 1.70 and steers at 1.00 (Fig. 5 left).
  • Circle fit. Over 160 sites the H1 ratio and PCA roundness are uncorrelated (−0.018), and 38 sites carry a strong loop that is not a flat circle (App. H). A new sweep agrees: in the top-2 PCA plane only 20% of cyclic clouds exceed the null, against 93–97% in 12 to 32 dimensions (Answer 5).
  • Sparse autoencoder. Similar token ranking (r_s = 0.68), but saturated inside the mention span, and it requires a dictionary trained on about 4B tokens (§4.1).

These comparisons are spread over §4.1 and App. H and are easy to miss. The revision gathers them in one table, next to the new gating baselines.

What the new comparison shows (for "whether"). For the membership decision alone, topology is not needed: a distance-to-centroid or kNN test at the same site separates target from off-task inputs better than our reader-plane test (Answer 3). What the topological route contributes is the site analysis above and a pipeline that, gate included, runs without labels or negatives.

AuthorsOpenReview ↗

Rebuttal (2/5): Weakness 2

[W2] It is unclear how a "cyclic concept" is defined. If they are concepts whose activations generate loops, the argument becomes circular, as one justification to use TDA would have been the presence of cyclic concepts. Moreover, looking at the results, persistence score does not seem to be able to cleanly separate "cyclic" from "non-cyclic" concepts (though again there is no definition of what's cyclic and what is not). Given these two weaknesses hinder both motivation (why TDA?) and experimental setup (why cyclic concepts, and what are they), I am not reporting other weaknesses in the analysis, as they are dependent on solving these two first.

Answer:

We agree that the paper used "cyclic concept" without defining it, and that this made the argument look circular. The definition we use is at the level of the task and never refers to activations:

A concept with labels y0,...,yn−1y_0, ..., y_{n−1} is cyclic if the task's ground-truth successor is a single n-cycle, s(yi)=y(i+1)modns(y_i) = y_{(i+1) mod n}: the last label is followed by the first (after December comes January; after 23:30 comes 00:00). It is linear if the order has two ends (nothing follows "true" on a truthfulness scale).

This has been added on the pdf.

This is fixed when the task is written. Each task in our benchmark carries a wrap-around flag, and the evaluation question asks for the successor ("The next month is", "Ten degrees clockwise is"), whose correct answer wraps for the cyclic tasks and stops at the end of the scale for the linear controls (honesty, sycophancy, sentiment, linear_order).

With this definition the logic is not circular. "Cyclic" is a hypothesis about the label space, set before any forward pass. Persistent homology then asks an empirical question that the definition leaves open: does this model represent the concept as a loop, and at which layer and token? The two can disagree, and the paper reports where they do. hue is cyclic by definition, yet on Qwen2.5-3B its ratio is 0.68 (below the null) and donor patching succeeds on 0.49 of inputs, while on Llama-3.1-8B the ratio is 1.32 and success is 0.99 (§4.1). If "cyclic" meant "its activations form a loop", such a case could not count as a failure.

Your observation that the score does not perfectly separate cyclic and non-cyclic tasks is correct and is explicitly examined in Appendices J and M. A structured finite cloud can exceed the Gaussian null without certifying semantic cyclicity. The method therefore also checks the circular coordinate for degeneracy before constructing the gate. In the reported example, the coordinate statistic is 0.85 for dense_clock, 0.10 for weekdays, and 0.05 for honesty. This is an additional empirical screen, not a general theorem that the representation is a circle. The cyclic task definition, structural licensing, and held-out gate performance are three distinct objects.

AuthorsOpenReview ↗

Rebuttal (3/5 - A): Question 1

[Q1] How does the gate compare with non-topological gates on the same inputs? For instance, CAST (Lee et al. 2025), cited in the paper, would be a method to test, but also simpler methods, a distance-to-centroid metric, or a kNN detector maybe?

Answer: We ran this comparison on the paper's models. Every gate is fitted on the same calibration prompts and scored on the same inputs (the target task, two other cyclic tasks and honesty), and the same donor patch is applied wherever a gate fires. The grid is Qwen2.5-3B, Llama-3.1-8B and Gemma-2-9B on five well-posed targets (dense_clock, clock, months, hue, heading), 15 cells, with evaluation templates that were not used for calibration.

  • Distance to centroid and kNN (k = 3): one-class tests on the same calibration cloud at the same read site, 0.90 coverage.
  • CAST (Lee et al., 2025), re-implemented from the paper and the released code: pair-centred PCA condition vector, score cos(h, tanh(proj_c h)) on the mean-pooled prompt activation, and its grid search over layer, threshold and direction that maximises F1. CAST needs negative prompts. We give it either prompts of the very tasks that are off-task at test time (its best case) or prompts of six other tasks. We also report CAST with its condition read at the concept token, the site our gate reads.
GateNeeds negativesAUROC, target vs. off-taskFires on target / off-taskOurs − gate at matched firing rate
Topological gate (ours)no0.7640.47 / 0.17n/a
Distance to centroidno0.9780.79 / 0.08−0.12
kNNno0.9950.70 / 0.01−0.14
CAST as published, negatives from the off-task tasksyes0.8900.67 / 0.23−0.04
CAST as published, negatives from other tasksyes0.8280.71 / 0.45−0.00
CAST at the concept token, negatives from the off-task tasksyes0.9630.89 / 0.03−0.11
CAST at the concept token, negatives from other tasksyes0.9440.99 / 0.19−0.10
Random gatenochanceat our rate+0.16

AUROC stands for "area under the ROC curve". It measures how well a score separates two groups, here target inputs (where the gate should fire) and off-task inputs (where it should not). The simplest reading: pick one target input and one off-task input at random. AUROC is the probability that the gate gives the target the higher score. Firing rates are at each gate's own threshold. The last column is the target percept (fraction of answers containing the steered label) of our gate minus that of the other gate when both fire on exactly the same number of inputs. All numbers are means over the 15 cells.

What the table says:

  1. Our gate is not the strongest detector. Distance and kNN at the same site are better on every measure, and CAST read at the concept token has a higher AUROC and a higher percept at a matched rate. We added it to the pdf.
  2. Against CAST as published, the label-free gate is close at a matched firing rate (−0.04 and −0.00 on average; CAST is ahead in 11 and 8 of the 15 cells) and fires less on off-task inputs on average (0.17 against 0.23 and 0.45). CAST finds more of the targets (0.67–0.71 against 0.47) and has the higher AUROC.
  3. CAST's selectivity depends on its negatives. Its off-task firing about doubles on average (0.23 → 0.45; at the concept token 0.03 → 0.19) when the negatives come from tasks other than those met at test time. Its grid search also returns a different layer (1 to 15), threshold (0.05 to 0.31) and direction for each model and concept, in line with the sensitivity Lee et al. report. Gates that use no negatives have none to mis-specify and no search to run.
  4. Where the condition is read matters for CAST too. Moving its check from the prompt mean to the concept token raises AUROC from 0.83–0.89 to 0.94–0.96. This is the "where" question of our paper, and it suggests the two approaches are complementary.

The table uses the gate exactly as evaluated in Table 3., the pipeline licenses 8 of the 15 cells; on those, at a matched firing rate our gate is level with distance (−0.01) and with CAST at the concept token (−0.01 / +0.06) and not behind CAST as published (+0.08 / +0.12, within noise for 8 cells), with AUROC 0.86 against 0.97 for distance and 0.99 for kNN. The paper's original evaluation prompts give the same ordering (AUROC: ours 0.81, distance 0.97, kNN 0.99, CAST 0.83–0.86, CAST at the concept token 0.93–0.95).

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

The paper analyzes activations across prompts, and identifies where a cyclic concept emerges and which tokens carry it, leading to a topology-aware approach to activation steering that jointly addresses both where and when to steer.

Critical strengths

N/A

Critical weaknesses

ideas are kind obvious

Questions for authors

N/A

3 comments
AuthorsOpenReview ↗

Rebuttal (1/1): Weakness 1

[W1] ideas are kind obvious

We thank the reviewer, and are glad the paper raised no soundness concerns. However, we respectfully disagree with the reviewer that our ideas are obvious. While the research questions in our submission, where to steer and whether to steer, are simple to state, they are not easy to achieve. Please allow us to explain why our work is not "kind obvious" and novel.

1. The contribution goes beyond observing circles or proposing conditional steering.

Engels et al. (ICLR 2025) established circular language-model features and their computational role. CAST (Lee et al., ICLR 2025) established conditional activation steering. We credit both in Appendix A. Our contribution is a different construction: infer a null-calibrated site score and circular coordinate from a concept-specific calibration cloud, then convert that structure into an O(d_h) membership gate for an existing intervention. Site selection uses neither successful-intervention labels nor the supplied cyclic order; fitting the gate does not require negative condition examples. The calibration prompts are deliberately concept-specific, so this is unsupervised inference within a chosen concept family, not unrestricted concept discovery.

These predecessors establish that circular representations and conditional control are possible. They do not supply the null-calibrated localisation-to-gating procedure evaluated here. The paper's novelty lies in that construction and its controlled empirical findings, not in claiming a new persistent-homology algorithm or a new isometry theorem.

2. The localisation result is not a consequence of merely detecting a loop.

In Figure 4, the onset span selected from unsteered activations reaches target percept 1.000, matching the oracle mention span. Patching the strongest-signal downstream span, an equal-size low-score span, or the best-gated individual token each yields 0.000. A downstream copy can carry the structural signal without being an equally effective intervention site. The result therefore concerns where an intervention works, not just where information is detectable. Pre-mention zeros are automatic under the aligned causal templates and are explicitly excluded from this argument.

Figure 5 tests a concrete alternative explanation. Independently varying name and concept slots have nearly matched activation variance, 1885 versus 1837, but persistence ratios 0.96 versus 1.70. Patching the name preserves the source answer on 96% of examples; patching the concept produces the target answer on 100%. Both slots vary across prompts, so neither total activation variance nor simply choosing the first varying token explains the observed distinction. This is a controlled empirical finding, not something established by the familiarity of the ingredients.

3. The gate is evaluated beyond simply reducing intervention frequency.

On dense_clock, target/off-task firing is 0.854/0.000, 0.958/0.028, and 0.796/0.000 across Gemma, Llama, and Qwen. Table 3 holds the donor intervention fixed and compares the gate against random gating at the same overall firing rate. Across the six well-posed cells, target percept is 0.340 versus 0.164. We report this descriptive advantage together with the months failures and variability; we do not turn it into a claim of universal reliability. The VLM experiment additionally achieves 0.800 target firing and zero control firing when the varying phase is conveyed by an image rather than named in the text.

The usefulness of this construction does not follow merely from knowing that circular representations exist. A downstream loop can be an ineffective intervention site, a strong structural signal can yield an unreliable gate, and the same criterion can behave differently across models. The contribution is to specify and test this connection, identifying both effective regimes and failures. If the novelty concern rests on a particular prior construction or result, please identify the overlap so that we can address it directly.

We hope this clarifies our contribution and that the reviewer will consider revisiting the rejection rating.

AuthorsOpenReview ↗

Any Questions from Reviewer ainA on Our Rebuttal?

We would like to thank the reviewer again for your thoughtful review and valuable feedback.

We would appreciate it if you could let us know if our responses have addressed your concerns and whether you still have any other questions about our rebuttal.

We would be happy to do any follow-up discussion or address any additional comments.

AuthorsOpenReview ↗

Another Call for Your Response

Dear Reviewer ainA,

We sincerely appreciate the time you have taken to provide feedback on our work, which has helped us greatly improve its clarity, among other attributes. This is a gentle reminder that the Paper Discussion Phase will end in less than 2 days from this comment, i.e., 11:59 pm AoE on November 18. We are happy to answer any remaining questions you may have or provide further clarification before then, but we will no longer be able to respond once the discussion phase has concluded.

If you agree that our responses to your reviews have addressed the concerns you listed, we kindly ask that you consider whether raising your score would more accurately reflect your updated evaluation of our paper. Thank you again for your time and thoughtful comments!

Sincerely,

Authors

Comments on the paper

5 comments
AuthorsOpenReview ↗

Global Response

Dear AC and reviewers,

Thanks for your thoughtful reviews and valuable comments, which have helped us improve the paper significantly. We are encouraged by the endorsements that: 1) the questions we study, where and whether to steer instead of which vector to add, are relevant and practically meaningful (reviewers 1bgw, ekyY), 2) Proposition 3.1 gives a clear reason why the check should be made before steering and not after (reviewer 1bgw), 3) our controls (Gaussian-null calibration, matched-variance distractors, rate-matched random gating) help distinguish topological structure from activation magnitude and from the effect of simply reducing intervention frequency (reviewer ekyY), 4) our results are reported transparently, negative findings included (reviewers 1bgw, ekyY), and 5) the paper raises no substantive soundness concerns (reviewer ainA).

One main question emerging from the reviews, raised from different angles by reviewers 1bgw and ainA, was what the topological route contributes: whether persistent homology is needed at all next to non-topological alternatives, and whether the ideas go beyond what is already known. Another, raised by reviewer ekyY, was whether the gate decides reliably when to steer across cyclic concepts, given its failure on months. We address these two questions here; the individual responses give the details.

What the topological route contributes. On where to steer, the topological score carries information that standard signals miss. A linear probe is saturated: its accuracy is at least 0.998 on all 11 tasks, while the success of a donor patch ranges from 0.49 to 1.00. On dense_clock, a random-name token with matched activation variance has a persistence ratio of 0.96 and does not steer, while the concept token has 1.70 and steers on every input. The position with the strongest signal is a downstream copy where patching has no effect, while patching the onset span that the scan selects matches the oracle. These controlled findings about where an intervention works do not follow from knowing that circular features exist (Engels et al., 2025) or that steering can be made conditional (CAST; Lee et al., 2025).

On whether to steer, we ran the comparison that reviewer 1bgw asked for, on identical inputs, and report the outcome as it is. For the membership decision alone, topology is not needed: distance-to-centroid and kNN tests at the same site are better detectors than our gate (AUROC 0.98 and 0.995 against 0.76), and so is CAST once its condition is read at the concept token. Against CAST as published, our gate is within 0.04 in target percept at a matched firing rate and fires less on off-task inputs. We therefore do not present the gate as a higher-performing detector. Our aim is a different angle of study: site scan, licensing decision and gate all derive from one unlabelled set of concept prompts, with no negative examples, no trained probe and no supervised search, together with a null-calibrated statement about the structure at the site, which lets the method abstain (new Tables 5 and 6).

Reliability of the gate across cyclic concepts. Reviewer ekyY is right that selectivity on dense_clock and an inversion on months cannot show that the gate works across cyclic concepts. We widened the test and traced the months failure to its cause. On five cyclic concepts and three models (15 cells, held-out templates), the gate fires more on target than on off-task inputs in 14 of 15 cells (mean 0.47 against 0.17). On months, all three models do lay the months out around a closed loop in calendar order (one neighbour swap on Llama), but twelve labels per cycle are too few for persistent homology to resolve that loop against the null; thinned to 12 labels on the same activations, the 48-point clock is lost in the same way. The months gate had been built on another loop, with a degenerate circular coordinate. The non-degeneracy check U(θ) of our method detects this without labels, but the submission's experiment had a bug for U(θ) and the gate grid had not applied it. With the check applied at a stated threshold, the pipeline declines months on Gemma and Qwen in every seed and licenses 8 of the 15 cells in a majority of seeds, on which the gate fires on 0.64 of target and 0.12 of off-task inputs. The revision restricts the claim accordingly, to cycles that are sampled densely enough for their loop to exceed the null once template effects are removed, and reports every abstention (new Table 1, updated Figure 7, new Appendix I.1). On retention: 0.90 is the in-sample coverage at which the gate is calibrated, and 0.56 is its measured recall on new target prompts; where it fires, the effect is that of always-on steering.

We hope that our rebuttal has helped to clear concerns about our work. We are glad to answer any further questions you have on our submission and we would appreciate it if we could get your further feedback at your earliest convenience.

AuthorsOpenReview ↗

Summary of Revisions

Revision summary

Incorporating the comments and suggestions of the reviewers, as well as some further analyses, we summarize here the main changes in the revised paper. New and modified text is highlighted in blue. Numbering refers to the revised PDF: a new Table 1 was inserted, so Tables 1–3 of the submission are now Tables 2–4 (our individual responses cite the submission's numbering).

  1. We have conducted additional experiments comparing our gate with non-topological gates on identical inputs (new Appendix H.1, Table 5; reviewer 1bgw). Distance to centroid, kNN and CAST (Lee et al., 2025; re-implemented from the paper and released code, as published and read at the concept token) are fitted on the same calibration prompts and scored on the same inputs, on 15 well-posed cells with held-out templates. Distance and kNN are stronger detectors than our gate (AUROC 0.978 and 0.995 against 0.764). Against CAST as published, for two choices of its negatives, our gate is close at a matched firing rate (−0.04 / −0.00 in target percept) and fires less on off-task inputs (0.17 against 0.23 / 0.45).

  2. We have gathered every comparison with a non-topological alternative in one table (new Appendix H.2, Table 6; reviewer 1bgw). For where to steer (linear probe, activation variance, planar circle fit, sparse autoencoder) and for whether to steer (random gate, always-on steering, Mahalanobis distance, distance to centroid, kNN, CAST), the table states whether our method is better, similar or worse, and what each alternative needs. Its conclusions: the topological route contributes to the site analysis, a null-calibrated statement about whether a loop is present at a site, and a pipeline that runs without labels or negatives compared to other prior methods.

  3. We have widened the gate grid from 6 to 15 well-posed cells (Section 4.2, new Table 1; reviewer ekyY). The grid covers five cyclic targets on three models (5 seeds, held-out templates). For every cell, Table 1 gives the labels per cycle, the firing rates of the gate as released, and the licensing decision and firing rates under the full rule. The gate as released fires more on target than on off-task inputs in 14 of 15 cells (mean 0.47 against 0.17).

  4. We have stated the threshold of the non-degeneracy check and applied it (Sections 3.2 and 4.2, Table 1, Figure 7, Appendix I.2; reviewer ekyY). Section 3.3 licenses a site only if ρ > 1 and the circular coordinate passes the check U(θ) of Eq. 1. The threshold is now the 0.95 quantile of U under the same null draws, fixed before any behavioural evaluation. With both conditions applied, the pipeline declines months on Gemma and Qwen in every seed and licenses 8 of the 15 cells in a majority of seeds; on these, the gate fires on 0.64 of target and 0.12 of off-task inputs. Figure 7 and Section 4.2 now show the two months cells as abstentions; the firing rates of the gate as released remain reported (Tables 1 and 4). On Llama, months is licensed in only 2 of the 5 seeds. Table 1 requires a majority of seeds, so it reports "abstains". Figure 7 plots the average over the five seeds, with the three unlicensed seeds counted as zero, which gives the low rates shown (0.34 on target, 0.27 off-task).

  5. We have added a diagnosis of the months failure (new Appendix I.1, Table 7; reviewer ekyY). The months are a cycle in all three models: their centroids lie around a closed loop in calendar order (with one swap of neighbours on Llama). But twelve points are too few for the loop to stand out: once template effects are removed, the persistence ratio of the concept's own loop falls with the number of labels per cycle, and thinning the 48-point dense_clock to 12 labels lowers its ratio from 1.54 to 0.89 on Llama and from 1.61 to 0.83 on Qwen. The pooled calibration cloud still passed ρ > 1 through its template clusters, with a circular coordinate that is constant on 107 of the 108 calibration points, which is what the check of item 4 detects. Section 4.2 accordingly restricts the claim to cycles that are sampled densely enough for their loop to exceed the null once template effects are removed.

  6. We have added a task-level definition of a cyclic concept (Section 2.3; reviewer 1bgw). A concept is cyclic if the task's ground-truth successor is a single n-cycle (after December comes January), and linear if the order has two ends.

  7. We have added the motivation for the PCA step and a sweep over its dimension (Section 3.1; Appendix L, new Table 8; reviewer 1bgw). The projection is not there for computation: it puts the data and the Gaussian null at the same effective dimension. Without PCA every one of the 39 clouds scores between 2.6 and 17, and the ratio stops discriminating. From 12 to 32 dimensions the decision ρ > 1 is the same as at k = 16 for 97–100% of clouds. In the top-2 plane only 20% of cyclic clouds exceed the null.

Area_Chair_Qg7ROpenReview ↗

Official Comment by Area Chair

Dear Reviewers,

As the discussion period draws to a close, please read and acknowledge the authors’ rebuttal within the next 24 hours. I also encourage you to engage in the discussion and share any remaining questions or concerns.

Thank you for your time and thoughtful reviews.

Best,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.