- 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
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:
| task | Gemma-2-9B | Llama-3.1-8B | Qwen2.5-3B |
|---|---|---|---|
clock | 0.18 / 0.12 | 0.58 / 0.10 | 0.44 / 0.06 |
dense_clock | 0.40 / 0.09 | 0.58 / 0.01 | 0.02 / 0.00 |
heading | 0.22 / 0.02 | 0.68 / 0.58 | 0.75 / 0.04 |
hue | 0.54 / 0.12 | 0.71 / 0.46 | 0.90 / 0.34 |
months | 0.50 / 0.23 | 0.33 / 0.04 | 0.15 / 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):
| task | labels per cycle | Gemma | Llama | Qwen |
|---|---|---|---|---|
dense_clock | 48 | 0.95 | 1.54 | 1.61 |
heading | 36 | 0.94 | 1.11 | 0.92 |
hue | 24 | 0.79 | 0.55 | 0.73 |
clock | 12 | 0.39 | 0.76 | 0.60 |
months | 12 | 0.47 | 0.13 | 0.37 |
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.)
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_clockgate 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
monthsis 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.
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:
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:
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:
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.