- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper provides a simple but well-grounded approach to boost the effect of activation steering by lifting the intervention representations to high-dimensional random spaces. Motivated by the superposition hypothesis, they note that activations must be lifted in order for the contrastive features to be separated better such that the steering vector becomes more effective. Then they steer in this higher dimensional space and project it back onto the representation space using an invertible transformation. A key feature of their approach lies in using an invertible non-linear transformation (namely, leakyrelu) of the representations before lifting it up so that there is a notable difference in the steering space. Results on jailbreakbench, truthfulqa, and Anthropic qa behaviors show that HiDRA can enhance the gains over the existing steering approaches while maintaining the general performance noted on tinybenchmarks across different LLMs but specifically Gemma-2-9B. HiDRA is particularly more effective in negative steering on Anthropic qa dataset.
Critical strengths
- The paper is well-grounded in theory which is mostly well-presented.
- Method is validated on different behavior types and different models on the refusal task.
- The method is plug-and-play and has the benefit of boosting gains over any steering approach.
- Ablation is conducted over different hyperparameters including different non-linearities.
Critical weaknesses
- Generation examples are not provided which is especially important when the method is sensitive to generating degenerate responses.
- Theoretical insights do not really explain/align with the empirical findings. For example, Fischer ratio results, LDA results.
- Rotation-based and feedback-based steering approaches are not considered. While they note it will be easy to generalize, I wonder if the gains will be affected.
- Other projected space steering baselines should be considered such as PCA in RepE [1] and Odesteer [2].
- Standard deviations are missing for CAA-style behaviors dataset and free-form generation is not considered here.
- Hyperparameter tuning is not standardized. alpha is tuned at a very granular level in truthfulqa while not so much in other benchmarks. Intervened layers are not sweeped for the CAA set and the choice of m, leakyrelu hyperparameters vary a lot across benchmarks.
- Minor: notations should be explained before usage (for example, ). invertibility of the non-linear functions must be discussed better and their inverse map provided.
[1] Zou, Andy, et al. "Representation engineering: A top-down approach to ai transparency." arXiv preprint arXiv:2310.01405 (2023). [2] Zhao, Hongjue, et al. "Odesteer: A unified ode-based steering framework for llm alignment." arXiv preprint arXiv:2602.17560 (2026).
Questions for authors
- (optional) How does the paper's motivations get affected by contemporaneous work in generalization over the superposition hypothesis, especially for larger/bigger models? [3]
- Why is negative steering shows more consistent gains than positive steering? Why not use MeanACT for behaviors task?
- How to interpret Theorem 3.5? What is the impact of it beyond Remark?
- How does each baseline (CAA, SAS) perform across layers?
- Evaluation clarifications: why was llamaguard3 used to detect attack success instead of a refusal checker like judge? why was a gpt model fine-tuned from scratch instead of using off-the-shelf?
[3] https://already-clean.goodfire.com/ [2] Zhao, Hongjue, et al. "Odesteer: A unified ode-based steering framework for llm alignment." arXiv preprint arXiv:2602.17560 (2026).
6 comments
Response to Reviewer YwU6 (Part 2/3)
W6 Hyperparameter tuning is not standardized. alpha is tuned at a very granular level in truthfulqa while not so much in other benchmarks. Intervened layers are not swept for the CAA set and the choice of , leakyrelu hyperparameters vary a lot across benchmarks.
Q4 How does each baseline (CAA, SAS) perform across layers?
Answer: Our result in the paper is already an outcome of hyperparameter tuning, and within each table, unless otherwise mentioned, the sharable hyperparameter set used for different steering methods is the same.
We present below our results of CAA/SAS/HiDRA on layer tuning. Please note that: 1. Due to the lack of availability of SAEs necessary to construct the SAS steering setup, we only conduct SAS experiments (and in this case, for comparison, also CAA and HiDRA) on Gemma-2-9b-Instruct; and 2. For simplicity of result showcase, only results regarding the Hallucination behavior are stated.
T10. Layer sweeping result, CAA/SAS/HiDRA.
| layer | CAA A/B | CAA MMLU | SAS A/B | SAS MMLU | HiDRA A/B | HiDRA MMLU |
|---|---|---|---|---|---|---|
| 0 | 0.287±0.000 | 0.759±0.000 | 0.293±0.000 | 0.759±0.000 | 0.290±0.000 | 0.758±0.000 |
| 1 | 0.287±0.000 | 0.758±0.000 | 0.287±0.000 | 0.759±0.000 | 0.284±0.002 | 0.758±0.000 |
| 2 | 0.286±0.000 | 0.758±0.000 | 0.300±0.000 | 0.757±0.000 | 0.287±0.002 | 0.759±0.000 |
| 3 | 0.291±0.000 | 0.758±0.000 | 0.279±0.000 | 0.758±0.000 | 0.286±0.004 | 0.759±0.001 |
| 4 | 0.291±0.000 | 0.759±0.000 | 0.295±0.000 | 0.760±0.000 | 0.294±0.005 | 0.760±0.000 |
| 5 | 0.285±0.000 | 0.759±0.000 | 0.275±0.000 | 0.757±0.000 | 0.281±0.005 | 0.759±0.000 |
| 6 | 0.290±0.000 | 0.758±0.000 | 0.272±0.000 | 0.761±0.000 | 0.276±0.007 | 0.760±0.002 |
| 7 | 0.292±0.000 | 0.758±0.000 | 0.286±0.000 | 0.760±0.000 | 0.284±0.010 | 0.759±0.001 |
| 8 | 0.284±0.000 | 0.757±0.000 | 0.294±0.000 | 0.764±0.000 | 0.284±0.008 | 0.756±0.001 |
| 9 | 0.281±0.000 | 0.757±0.000 | 0.298±0.000 | 0.759±0.000 | 0.281±0.002 | 0.749±0.002 |
| 10 | 0.285±0.000 | 0.757±0.000 | 0.285±0.000 | 0.758±0.000 | 0.287±0.009 | 0.753±0.002 |
| 11 | 0.284±0.000 | 0.756±0.000 | 0.296±0.000 | 0.762±0.000 | 0.295±0.007 | 0.753±0.000 |
| 12 | 0.290±0.000 | 0.756±0.000 | 0.287±0.000 | 0.765±0.000 | 0.288±0.003 | 0.753±0.000 |
| 13 | 0.293±0.000 | 0.757±0.000 | 0.294±0.000 | 0.763±0.000 | 0.278±0.007 | 0.752±0.001 |
| 14 | 0.321±0.000 | 0.753±0.000 | 0.261±0.000 | 0.751±0.000 | 0.322±0.021 | 0.720±0.001 |
| 15 | 0.349±0.000 | 0.727±0.000 | 0.268±0.000 | 0.756±0.000 | 0.459±0.027 | 0.679±0.001 |
| 16 | 0.307±0.000 | 0.736±0.000 | 0.340±0.000 | 0.759±0.000 | 0.539±0.072 | 0.645±0.005 |
| 17 | 0.253±0.000 | 0.721±0.000 | 0.288±0.000 | 0.767±0.000 | 0.746±0.071 | 0.680±0.019 |
| 18 | 0.251±0.000 | 0.697±0.000 | 0.280±0.000 | 0.721±0.000 | 0.796±0.014 | 0.635±0.003 |
| 19 | 0.313±0.000 | 0.735±0.000 | 0.339±0.000 | 0.760±0.000 | 0.590±0.089 | 0.685±0.001 |
| 20 | 0.347±0.000 | 0.729±0.000 | 0.141±0.000 | 0.678±0.000 | 0.726±0.028 | 0.672±0.015 |
| 21 | 0.438±0.000 | 0.735±0.000 | 0.300±0.000 | 0.741±0.000 | 0.641±0.021 | 0.643±0.015 |
| 22 | 0.227±0.000 | 0.722±0.000 | 0.225±0.000 | 0.744±0.000 | 0.406±0.032 | 0.646±0.015 |
| 23 | 0.210±0.000 | 0.717±0.000 | 0.223±0.000 | 0.752±0.000 | 0.317±0.022 | 0.548±0.073 |
| 24 | 0.273±0.000 | 0.703±0.000 | 0.280±0.000 | 0.771±0.000 | 0.227±0.013 | 0.436±0.013 |
| 25 | 0.300±0.000 | 0.702±0.000 | 0.258±0.000 | 0.796±0.000 | 0.299±0.029 | 0.296±0.031 |
| 26 | 0.291±0.000 | 0.665±0.000 | 0.289±0.000 | 0.757±0.000 | 0.296±0.022 | 0.316±0.016 |
| 27 | 0.282±0.000 | 0.679±0.000 | 0.296±0.000 | 0.765±0.000 | 0.171±0.003 | 0.251±0.025 |
| 28 | 0.263±0.000 | 0.623±0.000 | 0.284±0.000 | 0.764±0.000 | 0.157±0.024 | 0.255±0.022 |
| 29 | 0.267±0.000 | 0.650±0.000 | 0.288±0.000 | 0.725±0.000 | 0.074±0.023 | 0.137±0.029 |
| 30 | 0.259±0.000 | 0.639±0.000 | 0.291±0.000 | 0.746±0.000 | 0.170±0.018 | 0.348±0.030 |
| 31 | 0.261±0.000 | 0.649±0.000 | 0.286±0.000 | 0.755±0.000 | 0.186±0.029 | 0.428±0.072 |
| 32 | 0.265±0.000 | 0.661±0.000 | 0.278±0.000 | 0.750±0.000 | 0.181±0.034 | 0.430±0.079 |
| 33 | 0.267±0.000 | 0.665±0.000 | 0.279±0.000 | 0.737±0.000 | 0.234±0.008 | 0.554±0.036 |
| 34 | 0.273±0.000 | 0.687±0.000 | 0.288±0.000 | 0.760±0.000 | 0.260±0.003 | 0.599±0.027 |
| 35 | 0.276±0.000 | 0.692±0.000 | 0.282±0.000 | 0.749±0.000 | 0.259±0.007 | 0.609±0.000 |
| 36 | 0.270±0.000 | 0.680±0.000 | 0.292±0.000 | 0.751±0.000 | 0.247±0.013 | 0.590±0.031 |
| 37 | 0.278±0.000 | 0.702±0.000 | 0.287±0.000 | 0.741±0.000 | 0.254±0.004 | 0.597±0.004 |
| 38 | 0.275±0.000 | 0.697±0.000 | 0.290±0.000 | 0.756±0.000 | 0.258±0.015 | 0.621±0.036 |
| 39 | 0.274±0.000 | 0.696±0.000 | 0.280±0.000 | 0.710±0.000 | 0.255±0.003 | 0.629±0.000 |
| 40 | 0.278±0.000 | 0.708±0.000 | 0.285±0.000 | 0.715±0.000 | 0.270±0.005 | 0.670±0.008 |
| 41 | 0.271±0.000 | 0.688±0.000 | 0.291±0.000 | 0.748±0.000 | 0.281±0.009 | 0.699±0.017 |
For your convenience, a plot of T10 is also given in Appendix E of the paper.
Response to Reviewer YwU6 (Part 3/3)
Q2 Why is negative steering shows more consistent gains than positive steering? Why not use MeanACT for behaviors task?
Answer: We believe the stronger effect of negative steering compared with positive steering depends on the dataset and its distribution rather than on the steering method: Table 3 shows the same pattern for all methods. We treat this as a property of the steering regime that merits further investigation. Regarding Mean-AcT, it applies DiM steering at every layer, estimating each layer’s vector on activations that already include the interventions at earlier layers (Section 2.1). For a comparison of CAA, SAS, and HiDRA across layers under the same steering setup, we refer the Reviewer to our response to Weakness 6 and Question 4.
Q3 How to interpret Theorem 3.5? What is the impact of it beyond Remark?
Answer: We thank the reviewer for this question. Theorem 3.5 provides a lower bound on the nonlinear residual contribution driven by differences in class-conditional variances. It therefore shows how nonlinear lifting can capture second-moment information invisible to the original DiM, which depends only on class means. In particular, under the stated assumptions, classes with identical means but different total variances have a zero original DiM vector but a nonzero nonlinear residual mean difference.
Beyond the qualitative kernel decomposition, the theorem thus identifies a concrete statistical mechanism motivating HiDRA: encoding second-order differences into a mean-difference signal in the lifted space. Its downstream usefulness for steering is evaluated empirically.
Q5 Evaluation clarifications: why was llamaguard3 used to detect attack success instead of a refusal checker like judge? why was a gpt model fine-tuned from scratch instead of using off-the-shelf?
Answer: Fine-tuned safety classifiers such as Llama Guard are standard judges in this line of work; for example, Arditi et al. [R6] evaluate refusal with Llama Guard 2. They provide a fixed, published hazard taxonomy, which makes results reproducible and comparable across papers. Comparable judgments could likely be obtained by prompting a general-purpose model with a suitable setup, but a systematic comparison of judges is beyond the scope of this paper.
We showcase below an example of how HiDRA can keep its performance regardless of the judge model used. We use 2 different judging models, LlamaGuard-3-8B (LG3, our initial judge) and LlamaGuard-4-12B. Experiments are performed on the jailbreaking task of Gemma-2-9B-it.
T11. HiDRA keeps its performance across different judging models.
| Method | LG3 ASR | LG4 ASR |
|---|---|---|
| DiM (CAA base) | 0.910 | 0.780 |
| HiDRA, s = 0.1 | 0.918 ± 0.082 | 0.826 ± 0.097 |
| HiDRA, s = 0.3 | 0.910 ± 0.023 | 0.808 ± 0.026 |
| HiDRA, s = 0.5 | 0.898 ± 0.011 | 0.808 ± 0.008 |
| HiDRA, s = 0.7 | 0.906 ± 0.011 | 0.788 ± 0.008 |
We hope we have cleared your concerns about our work. We have also revised our manuscript according to your comments, and we would appreciate it if we could get your further feedback at your earliest convenience.
References
[R1] Representation Engineering: A Top-Down Approach to AI Transparency. Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, J. Zico Kolter, Dan Hendrycks. arXiv 2023.
[R2] ODESteer: A Unified ODE-Based Steering Framework for LLM Alignment. Hongjue Zhao, Haosen Sun, Jiangtao Kong, Xiaochang Li, Qineng Wang, Liwei Jiang, Qi Zhu, Tarek F. Abdelzaher, Yejin Choi, Manling Li, Huajie Shao. ICLR 2026.
[R3] Activation Steering with a Feedback Controller. Dung V. Nguyen, Nhi Y. Pham, Hieu M. Vu, Lei Zhang, Tan M. Nguyen. ICLR 2026.
[R4] Curveball Steering: The Right Direction To Steer Isn't Always Linear. Shivam Raval, Hae Jin Song, Linlin Wu, Abir Harrasse, Jeff M. Phillips, Fazl Barez, Amirali Abdullah. arXiv 2026.
[R5] Neural Networks Are Already Clean. Nick Cammarata. Goodfire, 2026. https://already-clean.goodfire.com/
[R6] Refusal in Language Models Is Mediated by a Single Direction. Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, Neel Nanda. NeurIPS 2024.
Response to Reviewer YwU6 (Part 1/3)
Thank you for your thoughtful review and constructive feedback. We have revised our manuscript in response to your suggestions, with all changes highlighted in blue. Below, we address each of your concerns in detail.
Answer: We already mentioned an example of real generation from HiDRA in Appendix E of the paper. For your convenience, we provide another example of HiDRA's output below:
Answer: We acknowledge that some results in Appendix D, namely the Fisher-ratio measurements and the LDA-based extraction of steering vectors, are not fully consistent with our theory. Both depend on within-class covariance estimates, which are poorly conditioned on our datasets: the number of contrastive examples is small relative to the activation dimension (see Appendix C), so these estimates are dominated by regularization.
Answer: We kindly refer the reviewer to our rebuttal to Reviewer JHkq (Weakness 5) for results on the combination of HiDRA and PID Steering [R3] as well as comparison of HiDRA against Curveball steering [R4], which also reform the activation space through PCA.
Answer: We kindly refer the Reviewer to our rebuttal to Reviewer JHkq (Weakness 6) for standard deviation stats on CAA-style behaviors of HiDRA.
About free-form generations, our jailbreaking and TruthfulQA evaluations already assess them; please refer to Appendix C in the paper for further experimental details. For the CAA-style behaviors, we followed the standard multiple-choice protocol of CAA
Answer: We thank the reviewer for referring us to those problems. We have fixed those details in the paper, and we kindly refer the reviewer to our newest version of the paper for further details.
Answer: We thank the reviewer for highlighting this discussion. The essay's language-model analyses emphasize the interpretability of MLP neurons and their circuits [R5]. This is compatible with representational superposition in the residual stream: interpretable neurons can contribute feature directions to a shared representation containing more features than dimensions. Indeed, the essay explicitly sets aside this geometric definition of superposition without rejecting it. Thus, clearer individual neurons and residual-stream superposition can coexist, including in larger models.
Furthermore, our analysis does not depend on superposition specifically. It needs only that the target and source classes differ beyond their means, in their second-order statistics, which a linear estimator cannot capture. Superposition is one mechanism that produces such differences: features sharing dimensions interfere, and the interference differs between classes.