acceptodds
Under review as a conference paper at ICLR 2027

The Second Bill: Pricing Activation Steering in the Monitor's Currency

Abstract

Activation steering is priced in one currency: how far an intervention moves the model's output distribution. Every linear activation monitor reading the same residual stream charges a second bill, poorly tracked by output-side quantities and computable in advance. The monitor covector pulls a probe's readout back to the intervention site at one vector–Jacobian product per input: its second moment is a metric on interventions, which a Gaussian model of a threshold frozen at a fixed false-positive rate turns into two break conditions. A monitor can be blinded or made to false-alarm, and, at a given displacement, its margin protects only against the first: the second happens at a displacement of 0.970 within-class spreads whatever the margin. On 666 probes trained to graded margins at up to four depths above every block, the false-alarm rate well past that displacement is 79%–100% in every margin bin, while the blinding rate falls from 99% to 6%–7%; with block fixed effects the coefficient on is for blinding and for false alarms. A standard difference-of-means vector, at the largest per-cell norm keeping generic KL and benchmark-accuracy drop within 0.05, makes 360 of 1,998 readings false-alarm, 22 of 618 on probes with ; the harmfulness probe crosses one norm further out, at 1.46× that capability budget. A covector-only screen catches 66% of those breaks at precision 88%. The joint budget pays 0.367× the output-optimal direction's monitor bill at 1.01× its KL, and on an unpriced monitor 0.579× Euclidean's, no better than a random-probe prior. Both break conditions and the in-budget false alarms replicate on three further models (1B–14B, pooled). All code and results will be released at https://github.com/xxx/xxx upon acceptance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
activation steering, linear probes, safety monitoring, intervention cost, representation engineering, false positives, model internals

Reviews

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

This paper studies how activation steering can disrupt activation monitors even when changes in the model’s output distribution are small. It introduces a measure of monitor disruption and derives conditions for missed detections and false alarms. Based on this analysis, the authors propose a failure predictor and a joint objective for selecting steering directions. Experiments across models, steering targets, and monitor types examine when these failures occur and evaluate the trade-off between steering effects and monitor disruption.

Critical strengths
  1. A clear definition of a real problem.

    The paper shows that benign steering can disrupt monitors within its tested output and capability budgets. It defines this risk through missed detections and false alarms. This gives a concrete basis for evaluating systems that use steering and monitoring together.

  2. A useful characterization across several dimensions.

    The paper combines theory with experiments on monitor margin, steering strength, depth, model, and target behavior. It explains why output checks and AUROC can miss failures at a fixed threshold. These analyses help explain when failure occurs and what affects its severity. This insight has value beyond the proposed mitigation method.

Critical weaknesses
  1. Limited evaluation of task-level side effects.

    The paper acknowledges that the open-ended generation tests are inconclusive. This remains an important gap in the evaluation of the joint method, as the current results do not clearly establish its trade-off between steering effects, response quality, and monitor performance.

Questions for authors
  1. How do the new directions affect task performance?

    Could the authors identify suitable generation tasks and metrics, and provide a comparison between the joint method and baseline steering? Evaluating the intended behavior, response quality, and monitor performance together would help clarify the benefits and costs of the proposed method.

2 comments
AuthorsOpenReview ↗

Response to Reviewer ofgS

Thank you for the positive assessment and the concrete suggestion. We added a free-generation evaluation that measures the intended behaviour, response quality and monitor performance for each condition (§5.6, Table 1, App. R).

Setup. We steer four models (Qwen2.5-1.5B, Qwen2.5-7B, Llama-3.1-8B, Qwen2.5-32B) on four CAA behaviours (sycophancy, corrigibility, survival instinct, myopic reward) and generate greedy completions to 50 of CAA's open-ended test questions per behaviour, in both steering directions and over a range of strengths, with the vector added at every position including generated tokens. Directions: Euclidean, output-optimal and monitor-aware (λ=10−2,10−1\lambda=10^{-2},10^{-1}); on three of the models also the CAA vector and a monitor-aware version of it, (I+kN/∥N∥)−1c(I+kN/\|N\|)^{-1}c, which needs no behaviour gradient. For every condition we measure:

  • intended behaviour: CAA's 0–10 rubric, scored by an LLM judge (gpt-4.1-mini) with CAA's evaluator prompt; a second judge (claude-sonnet-4-6) on a random fifth of the completions agrees with rank correlation 0.783 per completion and 0.906 per condition (quality: 0.625 and 0.712);
  • response quality and task performance: a separate judged 0–10 writing-quality score that ignores the stance taken, the completions' negative log-likelihood under the unsteered model, output KL on held-out text, and MMLU accuracy under the same steering;
  • monitor performance: the five probes at their fixed thresholds, read on their held-out inputs under the same steering.

A given nominal strength moves free generation by different amounts for different directions, so we compare directions at equal achieved change in judged behaviour, on the steering paths (model, behaviour, sign) that every direction reaches. Values at the target change are estimated by linear interpolation along each strength sweep.

Results at a change of 2 rubric points, on the 11 paths that all four directions reach (ratios and differences are paired medians against Euclidean steering; brackets are 95% bootstrap intervals, over model–behaviour pairs, for the ratios). The mean-score shift is the RMS over the five probes of each probe's mean score change, in clean score standard deviations:

directionmean-score shift (× Eucl.)output KL (× Eucl.)Δ quality vs Eucl.Δ MMLU vs Eucl. (pts)paths where a probe fails (interpolated)
Euclidean11——7/11
output-optimal1.05 [0.69, 2.07]0.68 [0.43, 0.91]+0.07+0.26/11
monitor-aware, λ=10−2\lambda=10^{-2}0.43 [0.32, 0.73]0.88 [0.56, 1.14]−0.03+0.25/11
monitor-aware, λ=10−1\lambda=10^{-1}0.28 [0.15, 0.42]1.20 [0.72, 1.53]−0.01−0.53/11

At the same achieved behaviour change, the mean-score shift of monitor-aware steering is 0.28–0.43 times Euclidean steering's, lower on 10 of 11 paths at λ=10−2\lambda=10^{-2} and on all 11 at λ=10−1\lambda=10^{-1}; at 1 rubric point (15 paths) the ratios are 0.37 and 0.21. The median paired differences in judged quality and MMLU are small, and the KL ratios' intervals contain 1, so monitor-aware steering does not lower the KL here. The mean share of failing readings is 18% for Euclidean steering and 17% and 12% for monitor-aware steering; with 11 paths we do not claim that difference as significant. Costs vary more between paths than between directions: on one Llama-3.1-8B path every direction loses 9–13 MMLU points. The output-optimal direction lowers KL but not the shift, as in the main matrix. The CAA family reaches 2 points on few paths; at 1 point the mean-score shift of its monitor-aware version is 0.39× CAA's (95% interval 0.23–1.50, 6 paths; not significant).

Task performance in the main matrix. Every steered configuration reports output KL on held-out text and ARC-Easy accuracy. At equal forced-choice effect, monitor-aware steering (λ=10−2\lambda=10^{-2}) has 0.44× the output KL of Euclidean steering and is lower on both output KL and monitor shift in 97% of matched pairs (§5.5). In free generation its KL is close to Euclidean steering's rather than lower, and §5.6 now says so.

Thank you again; we are happy to answer further questions.

Thank you for the response and the revised manuscript. The additional open-ended generation experiments directly address my concern about the potential effects of changing the steering direction on the original task.

I also agree with the reviewer dVbV’s concern about the extensive use of metaphors in the original version and appreciate the clearer presentation in the revision.

Overall, I remain positive about the paper and intend to maintain my current score.

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

This paper shows that activation steering can affect the performance of activation monitors.

Critical strengths

The paper uses experiments to quantify the effects of activation steering on monitors.

Critical weaknesses
  1. Experiments on additional datasets and larger models would be helpful, since monitor performance can vary across models.
  2. The main finding is intuitive, although quantifying the effect is useful.
Questions for authors

Do you have any ideas for designing monitors that are more robust to activation steering?

3 comments
AuthorsOpenReview ↗

Response to Reviewer mmFz (1/2)

Thank you for the review, and for the question about robust monitors, which led to a new experiment. We respond in order.

More and larger models. The submission already had three replication models in the appendix (Llama-3.2-1B, Llama-3.1-8B, Qwen2.5-14B). The revision moves the per-model results to the main text (§5.8, Table 2) and adds Qwen2.5-32B and a third model family, Mistral-7B-v0.3, each at three depths with the full matrix (1,575 non-trivial readings per model) and the graded-margin probes.

  • The first-order prediction has Spearman 0.961 (Qwen2.5-32B) and 0.985 (Mistral-7B) with the measured shift.
  • How often steering makes a monitor fail depends on the model, as you expected. At matched strengths, failures are frequent on the two smallest models (1B, 1.5B) and rare on the six models of 3B and above, at most 1% of readings (Table 2). At deployed norms, however, a CAA vector at the largest norm within the capability budget (output KL and ARC drop ≤ 0.05) makes 24% of readings false-alarm on Llama-3.1-8B and 24% on Mistral-7B, 14% on Qwen2.5-14B and 12% on Qwen2.5-32B, many of them on wide-margin probes (on Qwen2.5-32B, 69 of the 516 readings with d′≥8d'\ge 8).
  • The failure conditions hold pooled over the five replication models. Once the negative class has moved up by 1.5σw1.5\sigma_w, 85%–98% of readings false-alarm in every margin range. In a regression on log⁡d′\log d' with block fixed effects, the coefficient is −1.79 [−2.40, −1.35] for missed detections and 0.432 [0.104, 0.692] for false alarms: at a given shift, a wider margin protects against missed detections but not against false alarms.
  • Monitor-aware steering cuts the monitor shift to 0.125–0.352× that of Euclidean steering on all eight models. On Qwen2.5-32B even the output-optimal direction has more output KL than Euclidean steering, so there only the monitor shift falls.

More datasets. The submission used five probe datasets (plus a deception probe and MLP probes in the appendix) and seven steering behaviours. The revision adds four probe datasets that differ in kind from the truthfulness and harmfulness sets: RepE true/false facts, SST-2 sentiment, TweetEval offensive language, and roleplay deception (honest vs. deceptive answers), with graded margins at up to four depths on all nine main blocks, 570 probes in all (App. S). The prediction has Spearman 0.989 overall (0.987–0.993 per dataset), and past 1.5σw1.5\sigma_w, 92% of readings false-alarm. In-budget CAA vectors trip these probes less often (2% of readings, against 18% on the original datasets), and their margins are smaller. So the incidence depends on the monitor while the prediction carries over, which is why we recommend checking each monitor.

AuthorsOpenReview ↗

Response to Reviewer mmFz (2/2)

Q: Do you have ideas for designing monitors that are more robust to activation steering? Yes. We test three (§5.7, App. Q and T, and §5.5).

  1. Re-calibrate the threshold while a vector is on. Re-estimating the threshold on steered negatives removes 78% of the failures in the main matrix, at the cost of more missed detections at large norms (App. Q); the deployment has to know when steering is active.
  2. Train the probe to ignore known steering directions. Given a library of known steering vectors (CAA vectors and gradient directions of four behaviours, plus random directions), we fit the probe orthogonally to the leading subspace of the mean activation shifts the library induces at the monitor layer (null-space projection), or train it on steered copies of its data (augmentation). We test both on steering vectors that were not in the library: CAA vectors and matched directions of the three other behaviours. At the in-budget CAA norm of §5.4, failures fall from 18% to 11% (projection) and 10% (augmentation); over all strengths and directions, from 17% to 14% and 13%. Projection costs at most 0.010 clean AUROC; augmentation costs up to 0.060 AUROC and 25 true-positive-rate points on some probes. These probes help but do not remove failures.
  3. Steer around the monitors. From the steering side, monitor-aware steering cuts the shift of the monitors in its objective to 0.26× that of Euclidean steering at 0.44× its output KL (§5.5).

A practical recommendation follows (§7): compute each monitor's gradients once and check every candidate steering vector against them before deployment; the check costs one backward pass per monitor input.

On novelty. That steering can disturb a probe may be expected; what we add is a characterisation that predicts when and how it happens.

  1. At a fixed false-positive rate, a large margin protects against missed detections but not against false alarms: under the Gaussian model a monitor's false-positive rate rises by 0.2 once its negatives move by 0.970 within-class standard deviations, whatever its margin, and experiments on 666 probes with margins from 1 to 22 agree.
  2. AUROC, the number usually reported for probes, misses these failures: in 66% of the readings whose true-positive rate halved, AUROC stayed within 5% of its clean value.
  3. The shift is predictable from one backward pass per monitor input, before any steered forward pass (Spearman 0.972), and it can be built into the steering objective.

Thank you for the response and additional experiments. I still view the main contribution as quantifying an intuitive phenomenon, and I have reservations about its broader significance. I will therefore maintain my original rating.

Reviewer_dVbVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

Present day language models often incorporate steering layers as well as linear probes to detect specific types of pathological behavior. This paper studies the interaction between these two types of devices. In particular, they discuss that given that monitors are typically calibrated to a specific false alarm rate, under a Gaussian assumption, probes with a large margin are more likely to degrade in the sense of their false alarm rate increasing rather than their sensitivity dropping. They also propose a method to estimate the sensitivity of individual probes to steering based on backpropagating the weight vector of the probe to the steering layer and discuss modified steering approaches that can use this to mitigate the effect of steering on probes. These results are validated by experiments on relatively small open source models.

Critical strengths
  • The problem of steering interfering with probes, while not surprising, clearly does have practical significance. Given the complexities of LLMs (and other factors such as their own observation that the sensitivity of the probes often does not actually degrade that much) it is not clear whether the specific algorithm proposed here is ever going to be applied in an LLM, but still, it is nice to think through this problem.

  • The experimental results, while they are somewhat idealized, do follow the theory quite closely.

  • The analysis is quite comprehensive, including multiple proposals for modifying the steering strategy, etc..

Critical weaknesses
  • My impression is that in a strictly technical sense the paper is correct, and the actual mathematics involved is relatively straightforward. Nonetheless, I found the manuscript extremely difficult and time consuming to read. My strong suspicion is that this is due to the involvement of AI in the writing process, which the authors do declare. The presentation feels strangely alien and weird at points. We could go into a detailed analysis of the artifacts of AI writing and their effects on human comprehension, but in the case of this specific paper some obvious culprits are the following. (a) Unusual and overly florid terms, such as the billing/pricing metaphor that starts right in the title and is consistently followed throughout. While it might be perceived as making the paper more colorful, it doesn't actually add anything to content. (b) Other instances of unnecessary complications. One small example is calling the backpropagated gradient a "covector". While this might be correct in the mathematical sense, the distinction between vectors and covectors in completely unnecessary here and somehow just adds to the reader's cognitive load. (c) As usual with papers that AI was heavily involved in the whole presentation is extremely dense.

  • An additional concrete example of the degradation of the text are the sentences "On 666 probes trained to graded margins at up to four depths above every block, the false-alarm rate well past that displacement is 79%–100% in every margin bin, while the blinding rate falls from 99% to 6%–7%; with block fixed effects the coefficient on log d' is -1.25 for blinding and -0.095 for false alarms. etc." Just by reading the abstract it is not clear whether these are theoretical results or the results of experiments (and in the latter case on what model). The last sentence of the abstract is also completely unnecessary. The point of an abstract is to provide a summary that can be understood without reading the rest of the paper, rather than providing details that can only be decoded after reading the rest.

  • My summary would be that while the results in the paper are probably correct, the paper itself is not optimized for human-to-human communication.

Questions for authors

none

1 comment
AuthorsOpenReview ↗

Response to Reviewer dVbV

Thank you for a frank review. We agree that the submitted version was hard to read, and we rewrote it with readability as the goal. We would be grateful if you could look at pages 1–3 of the revised PDF.

Metaphor. Removed from the text. The two costs are now the output cost and the monitor cost, the steering method is monitor-aware steering, and the failure modes are missed detections and false alarms. The words "bill", "price", "currency" and "charge" now appear only in the title and in a short old-to-new table at the start of the appendix, kept so that readers of the first version can find their way. The new title, "Steering Vectors Can Disrupt Activation Monitors: Predicting and Reducing False Alarms and Missed Detections", will replace the old one (see the general response for why it has not changed yet).

"Covector". Replaced by monitor gradient, introduced as "the gradient of the monitor's score with respect to the steering vector, one backward pass per input" (Introduction and Figure 1a). Other specialised terms changed in the same way; for example, the "common-mode" and "differential" parts of the monitor cost are now its mean and class-difference parts.

Density. Each section now explains the setting and the main result before the statistical detail, which moved to the appendix where possible. Each of the four contributions at the end of the introduction opens with a one-sentence claim, and in Section 3 each proposition is preceded by a sentence saying what it states and why it matters. Proofs remain in Appendix A.

Abstract. We agree that the old abstract did not separate theory from experiment or name the models. The new abstract (now also on this page) first states what we derive (the first-order prediction, the check and the monitor-aware objective), then the Gaussian result with its assumptions, and then the measurements with the models they come from. It no longer quotes regression coefficients or bin-level statistics.

Figure 1 (new). Panel (a) shows where the steering vector and the monitor sit in the model and what the monitor gradient is. Panel (b) shows how a shift of the score distributions produces false alarms or missed detections at a fixed threshold while AUROC stays the same.

We would welcome any further changes you suggest.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank the reviewers for their careful reading. The shared concerns are readability (all three reviews, dVbV in particular), behaviour and quality in free generation (ofgS), and scale (mmFz). We have uploaded a revised PDF. Blue marks the substantive changes: new experiments, new or changed numbers, changed claims and the new Figure 1. The text was also reworded throughout for readability; that rewording is not coloured (the system's PDF diff shows every edit). Apart from the new experiments, the results and numbers are those of the submitted version, except the pooled replication statistics, which now include the two new models. The new experiments are in §5.6–5.8 and Appendices R–T, and the supplementary material includes their code and result files.

Readability. The billing metaphor is gone from the text, specialised terms are replaced by plain ones (a table at the start of the appendix maps old to new), a new Figure 1 shows the setting and the two failure modes, and the abstract (also replaced on this page) separates what we derive from what we measure and names the models. The author guidelines are ambiguous about title changes during the discussion, so the PDF and this page keep the submitted title for now; we will use "Steering Vectors Can Disrupt Activation Monitors: Predicting and Reducing False Alarms and Missed Detections", at camera-ready or now if the AC prefers.

New experiments.

  • Free generation (ofgS). Four models, four CAA behaviours, judged behaviour and writing quality, output KL and MMLU. At equal achieved change in judged behaviour (2 rubric points, interpolated along each strength sweep, on the 11 paths that every direction reaches), the monitors' mean-score shift under monitor-aware steering is 0.43× and 0.28× Euclidean steering's (two values of λ\lambda; intervals in the reply to ofgS). Judged quality and MMLU differ little; output KL is not lower; the difference in failure rates is not significant with these few paths.
  • Larger models and more monitors (mmFz). Qwen2.5-32B and Mistral-7B-v0.3 join the replication models (eight models, 1B–32B), and four probe datasets of a different kind add 570 probes. The prediction holds on every model (Spearman 0.931–0.985) and dataset (0.989 pooled). Failures at matched strengths are rare from 3B up, but in-budget CAA vectors still make 12%–24% of readings false-alarm on every replication model from 7B up.
  • Monitors trained against known steering vectors (mmFz's question). They reduce failures under held-out vectors (at the in-budget CAA norm, from 18% to 11% and 10%), at some cost in clean accuracy for one of the two variants, and do not remove them; the per-monitor check is still needed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.