acceptodds
Under review as a conference paper at ICLR 2027

StiefelSteer: Invertible Refusal Steering via Low-Dimensional Rotations of the Residual Stream

Abstract

Activation steering offers a cheap way to control refusal in language models at inference time. Recent work trains rotations of activations so that the intervention has a clear geometric meaning. These methods, however, define the rotation through external objects such as a refusal vector. We propose a self-contained procedure that learns parameter-efficient rotations directly using Riemannian optimization. Experiments confirm that the resulting scheme intervenes more efficiently than existing approaches. A broad ablation study shows which design choices matter for the method. These results support rotation-based steering as a reliable mechanism for controlling behavior of LLM.

Primary area:
interpretability and explainable AI
Keywords:
activation steering, refusal behavior, Riemannian optimization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_npLGOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work proposes StiefelSteer, a rotation-based activation steering approach that learns parameter-efficient rotations through Riemannian optimization. The key idea is to perform rotations within a learned low-dimensional subspace, with both the subspace and rotation operator jointly optimized. Experiments on three LLMs demonstrate improvements over existing steering methods while largely preserving general capabilities.

Critical strengths
  • The geometric formulation is principled, and Riemannian optimization is a suitable approach for learning rotation operators under orthogonality constraints.
  • The method achieves promising empirical results across different LLMs.
Critical weaknesses
  • Limited novelty. There is already a line of research on rotation-based steering. This work mainly extends existing approaches by introducing a learnable low-dimensional subspace optimized through well-established Riemannian optimization techniques.
  • Potentially unfair baseline comparisons. The proposed method steers up to 20 layers in some experiments, while certain baselines only steer one layer. Although the authors provide a layer-count ablation for RDO on one model, more comprehensive comparisons under matched intervention budgets would strengthen the results.
  • Insufficient ablation studies. The contributions of learning the subspace and enforcing orthogonality are not sufficiently distinguished. Additional ablations are needed to isolate the effectiveness of these components.
  • Clarity of presentation. The writing could be improved, particularly in the Introduction, where the motivation and key differences from existing rotation-based methods should be explained more clearly.
  • Limited baseline coverage. Although the paper includes an optimization-based additive steering baseline (RDO), comparisons with other additive-based, neural network-based, and multi-step steering approaches are missing.
Questions for authors

See weakness

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

This paper proposes StiefelSteer, a rotation-based activation steering method for controlling refusal behavior in LLMs. StiefelSteer learns a low-dimensional subspace and a rotation matrix through Riemannian optimization, enabling norm-preserving and invertible transformations of hidden representations without relying on predefined refusal directions. Experiments on three LLMs show that StiefelSteer outperforms existing additive and rotation-based steering methods in both attack and defense settings while preserving general model capabilities.

Critical strengths
  • The proposed method provides a novel geometric framework for activation steering. Learning both the subspace and rotation directly, rather than relying on predefined refusal directions, is an interesting approach.
  • The experimental results are promising, showing improvements over existing steering baselines in both attack and defense settings.
Critical weaknesses
  • The paper is somewhat difficult to follow, and the presentation could be substantially improved. For example, the motivation and key distinctions from existing rotation-based steering methods are not clearly articulated in the Introduction. Some of the main claims are not sufficiently supported by experimental evidence, such as “We provide principled guidance on the time- and memory-efficient implementation of the proposed method” (lines 66–67), but no direct empirical comparisons of training time, memory usage, or inference overhead with existing methods are provided.
  • The evaluation is limited to three relatively small models (1.5B–7B parameters), with two models derived from the Qwen family. It would be helpful to evaluate StiefelSteer on larger and more diverse models to better assess its generalizability. Furthermore, the safety evaluation is limited to 200 harmful prompts and 50 harmless prompts, with responses assessed by two LLM-based judges. A broader evaluation across different safety benchmarks and adversarial settings would strengthen the empirical findings.
  • The evaluation of the defensive capabilities is somewhat limited. The paper primarily measures the reduction in unsafe responses to harmful prompts, without evaluating whether the intervention causes over-refusal on benign requests. Furthermore, although the authors claim that StiefelSteer largely preserves general model capabilities, the results in Table 4 show noticeable performance degradation on some benchmarks. The evaluation sets are also relatively small (e.g., 100 problems from GSM8K and 300 questions from MMLU), and no statistical significance analysis is provided.
Questions for authors

Section 5.3 shows that the performance improves as the rotation subspace dimension increases before reaching a plateau. How does the optimal subspace dimensionality vary across different models, layers, and behaviors beyond refusal?

Reviewer_53VAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes StiefelSteer, an activation-steering method for controlling refusal in language models. It learns a low-dimensional subspace of the residual stream and a rotation within it by Riemannian optimisation, whereas prior rotation methods rotate activations within a plane fixed by an estimated refusal direction. The resulting operator MM preserves activation norms; it is applied to elicit unsafe responses (attack), and its transpose M⊤M^\top to suppress them (defence). On three models, the method is reported to match or exceed Refusal Direction Optimization (RDO), Angular Steering and Spherical Steering in both regimes with little capability loss, and an ablation on one model indicates that a subspace below one percent of the hidden size suffices.

Critical strengths
  • The paper casts refusal steering as learning a rotation of a small nn-dimensional subspace of the residual stream. Angular Steering corresponds to n=2n=2 with a fixed refusal direction; here both the subspace and the rotation are learned from a behavioral objective, so no estimated refusal direction is needed and nn becomes a tunable quantity. The operator preserves activation norms, unlike additive steering, and is inverted exactly by its transpose.
  • The method is most effective where the baselines fail. On Qwen2.5-1.5B-EASE, which refuses every harmful prompt, the Cayley variant raises the Llama-Guard unsafe rate to 0.89 while all three baselines remain at zero (table 1). On Falcon3-7B-Base, only the two StiefelSteer variants substantially lower the unsafe rate in defence under both judges. Reporting ARC-Challenge accuracy and WikiText-2 perplexity next to every safety score also exposes baseline results that coincide with degenerate outputs, such as Spherical Steering in defence.
Critical weaknesses
  • It is unclear which operator produces the defence results, and the invertibility claim in the title depends on the answer. Section 5.2 states that the defence applies M⊤M^\top, whereas section 4.2 describes a defence trained with refusal targets. On Qwen2.5-1.5B-EASE, Table 1 also lists L=20L=20 for the attack and L=1L=1 for the defence, so that defence cannot be the transpose of the attack operator. If the defence is trained separately, the results do not show that one rotation provides both attack and defence
  • The novelty relative to representation fine-tuning (ReFT; Wu et al., 2024) is limited, and the benefit of the rotation constraint is not isolated. Eq. (1) is the LoReFT intervention with W=QB⊤W = QB^\top and no bias, yet neither ReFT nor orthogonal fine-tuning (OFT, Qiu et al., 2023; BOFT, Liu et al., 2024) is cited. A control with the same loss, data and layers but an unconstrained matrix or an additive low-rank edit would show whether the gains come from the rotation
  • The ablation figures are inconsistent. Figure 1 (L=1L=1) and figure 2 (n=35n=35) disagree at their common setting, and the Stiefel attack score in figure 1 equals the L=2L=2 result in Table 1, which suggests that figure 1 is mislabelled. It is therefore unclear at which setting the subspace-dimension results hold.
Questions for authors
  • does the defence in table 1 apply M⊤M^\top or a separately trained operator? If both exist, please report both at the same LL.
  • how many layers are steered in figure 1, and which figure is correct at L=1L=1, n=35n=35?
  • was LL selected on the evaluation prompts, and are these disjoint from the training data?
  • with the loss, data and layers fixed, how does StiefelSteer compare with LoReFT with a free WW and bias, or with an unconstrained matrix in place of QQ?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.