acceptodds
Withdrawn from ICLR 2027

What a Steering Rank Leaves Out: The Paired Response of Reference Directions

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A steering vector's rank among reference directions does not reveal how those directions respond when the injection sign is reversed, or what averaging across questions hides. We study 14 released contrastive activation addition vectors across seven behaviours and two Llama-2-chat sizes, using 39 reference directions per family, each matched to the vector's output-distribution cost at each sign. Pairing each reference direction's responses separates a common component from a sign-changing component. With random-split references, two rank-one cases show different patterns: one vector lowers the readout less than every random-split reference direction at the ranked condition (at the smaller size that family is geometrically concentrated, and the isotropic rank is 7 of 40); the other has a larger sign-changing response than every reference direction. Item averaging also hides structure: on myopic-reward at the smaller size, the common component accounts for 0.11 of the squared paired response after averaging, versus 0.58 before averaging. These shares depend on reference family, readout and amplitude. Failed control and interval-calibration checks limit our conclusions to descriptive findings on the observed sample. We recommend reporting paired, per-item reference responses alongside candidate effects and ranks.

Primary area:
interpretability and explainable AI
Keywords:
activation steering, steering vectors, contrastive activation addition, evaluation methodology, reference directions, control directions, item-level analysis, interpretability evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_pQGiOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper continue the work from some released steering vectors of LLMs by Rimsky et al., 2024. The author is worried that the vectors and report that their pre-registered positive control, bias checks and interval-coverage checks failed. They therefore present all results as descriptive point estimates and recommend reporting paired per-item reference responses.

Critical strengths
  • post-hoc amendments are documented ( few papers report their own negative checks).
Critical weaknesses
  • The practice being critiqued is barely in use, and if then the paper fail to convince me of worrying about what the paper is brining new to that practice.
Questions for authors

NA

Reviewer_J1DcOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies steering vectors, specifically how they are often measured in a misleading fashion. For example, comparing a vector to random vectors can make it look good just by being least bad. As the paper shows “on sycophancy all 39 random-split directions lower the readout, and the vector takes rank one by lowering it least”.

Using 14 CAA vectors released for Llama models in 2024, the authors propose evaluating vectors by splitting the responses into 2 parts: a “common component” and a “sign changing component”. The former stays the same, and the latter changes, when flipping the vector sign. This isolates the effect of any injection (the common part), from the steering component, leading to a less misleading perspective.

Critical strengths

The central observation, that reporting rank can be deeply misleading, is a genuinely good observation that identifies a core issue with prior work. The overall idea of splitting into 2 components is logically compelling.

Critical weaknesses

The scope of the study is quite limited (14 vectors on Llama models). Moreover, while the idea is of interest, I do not think this paper contains a significant novel research contribution. In the author’s own words, they make 3 contributions:

  1. “Different responses behind rank one”
  2. “Structure hidden by item averaging”
  3. “Family, readout and amplitude change the observed response summaries”

1 and 2 are good points, but more observations than genuine research contributions. 3 is really more acknowledging the limits of their evals than a contribution at all.

The main contribution, despite oddly enough not being listed in contributions, is the decomposition into common and steering components, but overall in my view this may be appropriate for a workshop paper, but does not meet the standards of ICLR.

The results themselves are not very conclusive. As the authors say in the abstract “Failed control and interval-calibration checks limit our conclusions to descriptive findings on the observed sample.“ While I commend their honesty in stating this, as opposed to obfuscating this fact somewhere in an appendix, this paper needs more work.

Questions for authors

Given the failed control and interval-calibration checks, and the small nature of the observed sample, how can you be confident that there is any significance in your findings?

Reviewer_6SekOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper argues that a steering vector’s rank against reference directions can hide how those directions respond to opposite injection signs and across individual test items. It examines 14 released CAA vectors on two Llama-2-chat models, comparing each with KL-matched random-split and isotropic reference directions. For each direction, the authors decompose the two signed readout changes into common and sign-changing components, reporting both aggregated and per-item summaries. The rank-one examples for sycophancy and coordinate-other-ais illustrate different response patterns. The registered positive controls and interval coverage checks do not meet their criteria, so the reported analyses are descriptive rather than population-level estimates.

Critical strengths
  • The paper identifies a real weakness in interpreting ranks alone. A candidate can rank highly because of what its reference directions do, rather than because it induces a large or specific behavioral change. The paired-sign comparison exposes this possibility in a way that a single rank cannot.
  • The authors disclose failed controls and interval checks instead of making unsupported inferential claims. The distinction between registered and exploratory analyses is explicit and well documented.
Critical weaknesses
  • I think the evidence is descriptive and the interpretation are not resolved. The registered positive control does not meet its threshold, and the reported uncertainty procedure fails its coverage checks. A key concern is that the multiple-choice readout responds to generic log-odds contraction or an answer-label preference, rather than to a property specific to steering vectors. A within-item answer-order swap and a held-out item replication would be straightforward tests. Without them, the main observation is interesting but hard to generalize.
  • The study is narrow and the presentation makes it harder to assess. It uses one steering method, one layer, two older model sizes, and small item sets. The random-split reference family at 7B is highly concentrated. conclusions should not be phrased as applying to reference directions broadly. I would also like the authors to spell out what a practitioner should do differently after seeing the paired plots. any learned or best practices to derive from this work.
Questions for authors
  • Can you run a within-item label swap (each item shown with the answers in both orders)? That would separate a letter or position preference from contraction and from differences between item subsets. I think this a key experiment that is missing from your evaluation.
  • Can you replicate the main paired-plane contrasts on held-out questions or another dataset, and on at least one recent model family or another layer?
  • At 7B the random-split references are dominated by two activation coordinates. What happens if you project them out or whiten the bank before splitting? Do the sycophancy-7B conclusions hold?
  • The planted gradient direction fails the both-condition control. What would a practitioner gain from your reporting recommendation? Is there a case where the paired response changes a decision about a vector, compared with reporting the effect, the rank and the reference effects at each sign?
  • How does this work differ from Luo et al. (2026, SteerCheck), which also uses cost-matched isotropic and sign-randomized references?

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

Related papers

Discussion (0)

Sign in to comment.