acceptodds
Under review as a conference paper at ICLR 2027

One-Step Steering is Not Sequential Control

Abstract

Sparse autoencoder (SAE) features give steering methods activation directions, but a direction specifies only one step: once the first intervention moves the hidden state, the second must be evaluated at a state where it was never specified. We call the rule that resolves this the execution law and study it for encoder–decoder controllers. We derive an exact identity showing that re-encoding after each step adds an interaction defect driven by the encoder–decoder round trip, and prove that good reconstruction does not bound this defect. Across eight public GPT-2 SAE widths, wider dictionaries reconstruct better while their round-trip error grows monotonically. On 493 held-out feature pairs in GPT-2 and Gemma-2-2B, the identity predicts which application order produces less interaction (Spearman 0.98–0.99), although accuracy falls when a controlled feature changes support. Choosing the predicted order cuts the interaction residual by 24–38%, yet direct error to the joint target changes by about 1% or less, because interaction accounts for only 4–7% of endpoint error and can cancel singleton errors. Reconstruction, interaction fidelity, and target delivery are therefore separate evaluation axes, and sequential steering must specify its execution law and measure outcomes directly.

Primary area:
interpretability and explainable AI
Keywords:
activation steering, sparse autoencoders, representation engineering, mechanistic interpretability

Reviews

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

A steering direction defines only one step. Later steps need a rule. The paper calls this rule the execution law. Re-encoding between steps adds interaction error. An exact formula links this error to decode-encode error. The paper proves that three properties are separate: reconstruction, interaction error and reaching targets. The paper tests eight public GPT-2 SAEs. In these SAEs, width improves reconstruction but raises decode-encode error. The paper also tests in-house GPT-2 and Gemma-2-2B SAEs. There, per-order terms usually pick the order with less interaction error. Final error barely changes.

Critical strengths

Figure 5 confirms the exact formula numerically.

In eight public GPT-2 SAEs, decode-encode error grows with width. This finding is a useful warning for state-dependent steering that re-encodes. Decoder conditioning explains only part of this growth (Appendix N).

Critical weaknesses

Interaction error has a narrow scope and a small effect. With a linear SAE decoder, fixed-vector steering has zero interaction error. Fixed-vector steering is the most common kind. The authors found no released method that needs their state-dependent feedback rule (Appendix A). With this rule, the chosen order cuts interaction error by 24% to 38%. Final error changes by about 1% or less. The paper itself stresses this point. Section 6 presents the formula mainly as a code check. The check tests that code follows its stated execution law. The narrow scope and small effect explain my informativeness score and rating.

The abstract supports its order claim with a rank correlation of 0.98 to 0.99. This correlation measures the size of interaction error, not the order. Order accuracy is 93% to 94%. When no feature turns on or off, per-order terms give each interaction error exactly. Section 5.2 and Appendix J therefore treat this match as a code check. The abstract calls the same match a prediction. When a steered feature turns on or off, order accuracy falls to 50% to 78%. The abstract mentions this drop without a number. My presentation score mainly reflects this use of the rank correlation.

Questions for authors

The paper should test whether order matters in a released state-dependent method. Two FLAS fields are one example.

At strength 0.02 times the hidden-state norm, interaction error is 4% to 7% of final error. The paper should report this share at the largest strength in Appendix H.

The authors should report order accuracy in the abstract. The authors should present the rank correlation as a code check.

Reviewer_h7SFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper examines the discrepancy in applying LLM activation interventions designed for a single layer are applied across multiple layers. Focusing on sparse autoencoders (SAEs), the authors derive a closed-form expression for the representation defect that arises when a perturbation at one layer propagates to the SAE at the next. They argue that this defect is not explained by reconstruction error alone, but is inherent to the design of SAEs and to single-step intervention methods.

Critical strengths

The formalization of “one step steering” and its errors is fairly rigorous and may be useful for further advancements in activation steering theory and applications.

Critical weaknesses
  • The central claim, that one-step steering does not amount to central control, is intuitive, and its key practical consequence (“Sequential steering therefore needs an explicit execution law, event-aware corrections, and direct outcome evaluation” l461-3), has been directly addressed by several activation steering papers (e.g., [1,2,3]). [2], for example, explicitly uses event-aware feedback control which adapts to the realized activations under perturbation. It appears that none of these papers are cited or discussed. As a result, it is unclear what this work contributes beyond existing results, and the omission raises concerns about whether the authors have adequately assessed the relevant literature. The authors should position this contribution relative to these works and explicitly state what is novel about this work.
  • Evaluations are limited to two relatively small models, so it is not clear if the method generalizes to other or larger LLMs. Crucially, the results also do not demonstrate a clear benefit from the proposed approach. In Table 1, it is shown (and acknowledged in prose) that the different controller settings have little impact on logit behavior. This undermines the practical utility of this method, as well as reinforces the significance concerns discussed above. Thus, the paper would be strengthened by extended evaluations to other models and more convincing results which show some practical utility to the method.

[1] Rodriguez et al. "Controlling language…” (2025) [2] Skifstad et al. "Local linearity…” (2026) [3] Kong et al. "Aligning large language…” (2024)

Questions for authors
  • In light of existing closed-loop or multistep steering formulations (see the Critical Weakness for examples), what are some practical consequences of this work which could inform steering controller design?
  • They key metric is designed for and evaluated on exclusively SAEs. The paper acknowledges other formulations of steering, such as contrastive addition. How does the proposed methodology translate to other single-step steering formulations?
  • R_o is frequently referenced but never explicitly defined. What does this notation mean?
Reviewer_G4wbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper evaluates how features determine activation directions, with direction specifies only one step: once the first intervention moves the hidden state, the second must be evaluated at a state where it was never specified. They propose a rule that resolves this the execution law and study it for encoder–decoder controllers, deriving an exact identity showing that re-encoding after each step adds an interaction defect and shows that good reconstruction does not bound this defect. The statement of is not well stated in. my opinion.

Critical strengths

The work is an interesting study and result, but seems more appropriate for a technical journal. There does seem to be some very interesting analysis here, but it is not very well written in order to get a better sense of exactly how useful or extensive the contribution is.

Critical weaknesses

The paper is fairly disorganized and difficult to follow. For instance, the beginning of the conclusion is "Interaction residual, total distance, named-target accuracy, behavior, and safety measure different properties. Order selection changes the first while leaving the larger direct error nearly unchanged." These are difficult sentences to extract the key contributions of the paper and their value. The examples of their methodologies are also quite weak.

Questions for authors

Why limit Table 1 to GPT-2 and Gemma-2-2B? Can you give a more insightful example of what the identify means in terms of execution?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.