acceptodds
Under review as a conference paper at ICLR 2027

Activation Steering Transfer to Agents: One Gain Ratio Does Not Identify Potency and Efficacy

Abstract

Activation Steering Transfer to Agents: One Gain Ratio Does Not Identify Potency and Efficacy Additive activation steering is calibrated in single-turn chat and deployed inside agent scaffolds. The quantity usually reported for that move is a gain: a ratio of steered effects, T=Delta_agent/Delta_chat. It is a summary of the same kind as a ratio of Braun's steerability slopes (Braun, 2026), though not an identity with one: the two coincide only under the proportional linear response this paper rejects. It is structurally a member of the relative-steerability family of Tan et al. (2024). We show on eight family x arm dose-response cells over six models, swept in both deployment contexts, that this ratio does not identify potency and efficacy: reconstructing the published estimator in both of its forms on our own grids, its realized range contains 1 in five of five scorable cells, it moves with dose in four of five, and two cells with opposite potency shifts, both CI-clean on the primary grid, return gain intervals that overlap at a width under 0.08. Every scorable cell in the frozen five-cell gain audit is an amplifier at one dose and an attenuator at another, so an amplify/attenuate taxonomy reports the dose it was read at. We replace the gain with a location: dEC50 = EC_50^agent - EC_50^chat, the cross-context difference in a curve location whose context-dependent horizontal reparametrization Taimeskhanov et al. (2026) derive analytically, in a simplified model, for one context at a time. It is a curve location rather than a ratio of effects, signed in both directions on this roster's primary grid across four distinct model families (+1.013[+0.777,+1.273] against -12.368, -10.855, -5.497 and -886.066 elsewhere), and beats a vertical rescaling at equal complexity in all five cells of the frozen horizontality audit (four under the registered alpha^*-trim). Because separating a mechanism from an artifact is the question at issue, we report the discipline at the same volume as the result: one cell is quarantined loudly, because its agent response inverts past a coherence collapse and the fit's orientation flips, manufacturing the roster's largest positive, which we delete. Three deflationary accounts (residual-norm rescaling, baseline alignment, dose transduction) are measured and rejected on sign pattern and magnitude, including wrong-sign predictions on the roster's positive anchor. Our pre-registered forecaster of dEC50 was refuted out of sample and published as refuted, our registered salvage claim (a location estimable where a gain ratio dies against a ceiling) produced no such cell and is reported unanswered, and a census of our own pre-registration register reports the registered branches no code in this repository could have emitted. The practical consequence is a measurement instruction rather than a theorem: a transfer conclusion read at one strength does not identify what changed, because a displacement and a gain are not distinguishable from a single operating point.

Primary area:
interpretability and explainable AI
Keywords:
activation steering, representation engineering, LLM agents, dose-response, identification, interpretability, measurement, steering vectors

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_aN1nOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

In this paper, the experiments support a key conclusion: steering strength calibrated in a chat setting must be recalibrated when applied to an agent setting. Measuring effectiveness at a single steering strength cannot distinguish between a shift in the strength required to achieve an effect and a change in the magnitude of the achievable effect.

Critical strengths

None, I think no substantive strength can be identified.

Critical weaknesses

Weakness 1: Unclear Background and Motivation

The introduction does not clearly establish the background or motivation for this work. In particular, the role of Additive Activation Steering is insufficiently explained, making it difficult to understand the problem being addressed and the rationale behind the proposed approach.

Weakness 2: Incorrect Format

The author seems to have included citations in the abstract, which is quite uncommon.

Tips: As this topic falls outside my primary area of expertise, and the paper’s lack of clarity further hinders my assessment, I find it difficult to identify substantive issues at this stage. I will revisit my assessment in light of the points raised by other reviewers.

Questions for authors

Please refer to Weakness.

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

This paper investigates how activation steering changes the effect between chat and agent. The existing work evaluates the ratio of the effects at a steering strength. However, the effect changes depending on steering strength. This paper proposes to use dEC50 considering the dose-response curve. In the experiments, the dEC50 is calculated with several LLMs.

Critical strengths

The analysis of the difference between chat and agent is important.

Critical weaknesses

The presentation should be improved. Many variables (alpha, u, …, etc) are used without definition. In abstract, equations and citations should be minimum. Delta_agent/Delta_chat is not defined.

The technical contribution is limited. The use of EC50 might be new for analyzing activation steering transfer, but it has been used in other fields.

The importance of the problem of activation steering transfer is not well described. There might be other points that need to be considered when transferring between chat and agent.

The problem of the gain is described. However, it is unclear how dEC50 can be useful. It might be helpful to show the usefulness by controlling the transfer of activation steering using dEC50 for example.

Questions for authors

Please see the weakness above.

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

The paper studies how activation steering calibrated in single-turn chat transfers to ReAct-style agent settings, questioning transfer assessments based on an agent-to-chat gain ratio at a single steering strength.

It proposes fitting dose-response curves in both contexts and using their midpoint difference, dEC50, to measure changes in steering potency. Experiments cover six open-weight models and eight refusal-induction/bypass configurations, starting with 200 items per configuration; two configurations are excluded by the validity screen.

The authors report shift signs varying across models on the primary grids and amplifier/attenuator labels reversing with dose. Their preregistered shift predictor fails out of sample, and the planned causal scaffold-editing experiment was not run.

The main finding is that a single-dose gain ratio cannot separately identify changes in steering potency and coherent efficacy when transferring activation steering from chat to agent contexts.

Critical strengths

The dose-response sweeps may provide useful evidence that amplifier/attenuator classifications can reverse with steering strength, exposing a limitation of single-dose transfer evaluation.

The authors also report failed predictions and exclude a spurious positive result, helping clarify the limits of their findings.

Critical weaknesses

Readability is poor. The paper is difficult to follow, beginning with the fifth sentence of the abstract, which introduces numerous terms and compressed expressions without sufficient explanation. Examples include, but are not limited to “family×arm dose–response cells,” “CI-clean,” “frozen five-cell gain audit,” and “registered salvage claim,” alongside “amplifier,” “attenuator,” and “quarantined.” Although some underlying concepts are familiar, their paper-specific meanings and relationships should not be assumed. This terminology obscures a relatively straightforward research question and makes the abstract unnecessarily difficult to understand.

Structure and logical flow is also very loose. Several headings and passages read more like internal research or revision notes than a reader-oriented scientific argument. For example, many paragraphs start with "what xxx", "why xxx" with no connective narrative around them. Appendix C explicitly discusses how to position the work because “under-conceding is the failure mode a reviewer punishes.” The manuscript repeatedly mixes scientific explanation with audit history and commentary on its own presentation, leaving the reader to reconstruct the connections between the research question, experimental design, and conclusions.

Figure is not readable. The only figure in the main text has text that's too small to read.

Limited contribution and the conceptual advance incremental. The proposed measurement compares fitted dose–response midpoints across contexts, and the main finding is a limitation of single-dose transfer evaluation. The study also does not isolate the effect of the ReAct scaffold. The chat-to-agent comparison simultaneously removes a compliance instruction, and the registered causal scaffold-editing experiment was not run. Furthermore, the only statistically clear positive shift loses significance under the registered dose cutoff, weakening the highlighted bidirectional result.

Taken together, these limitations lead me to recommend clear rejection.

Questions for authors

My main concerns are the clarity of presentation and the significance of the contribution, as detailed above (in the critical weaknesses). Addressing these would, in my assessment, require substantial revision rather than clarification alone.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.