acceptodds
Under review as a conference paper at ICLR 2027

SVM-STEER: Steer Clear of Biases with a Little Support

Abstract

Activation steering enables targeted control of large language model (LLM) behavior through interventions in internal representations. Existing approaches primarily construct steering vectors using unsupervised methods and often rely on brute-force searches to identify intervention layers. In this work, we propose a supervised alternative based on Support Vector Machines (SVMs). We evaluate our approach across two case studies spanning bias mitigation and censorship control, using six datasets and five open-weight LLMs. Our method reduces gender bias by an average of 29.3% and racial bias by 39.7% across models and steering options. For censorship evasion, steering decreases refusal probabilities by an average of 65% . For censorship enhancement, it increases the average refusal probability on XSTest from 0.064 to 0.83. We further show that steering effectiveness varies across model families and identify mode collapse under large steering coefficients, highlighting a trade-off between intervention strength and behavioral control. Our results demonstrate that supervised steering provides a simple, effective, and broadly applicable framework for controlling model behavior through internal representations.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Activation Steering, Large Language Models, Representation Learning, Model Editing, Supervised Steering, AI Safety, Bias Mitigation, Censorship Control

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

The paper proposes SVM-STEER, an activation steering framework that trains a linear SVM on per-layer hidden activations and extracts steering vectors from the resulting decision boundary. Three vector variants are considered (difference of class means, the hyperplane normal, and difference of support-vector means), together with a layer-selection score (RMSE divided by SVM accuracy) that picks the top 10 intervention layers without brute-force search, and a projection adjustment applied at inference time. For unlabeled settings, pseudo-labels are derived from the model's own logits. The method is evaluated on bias mitigation (gender, racial dialect) and censorship control (evasion and enhancement, measured on JailbreakBench, SorryBench, and XSTest) across five open-weight models in the 7B to 12B range, with a supervised case study on TruthfulQA and ToxiGen in the appendix. Steered models reduce bias scores by roughly 30 to 40 percent and can move refusal rates substantially in either direction, relative to the unsteered model.

Critical strengths

The evaluation is reasonably broad for a steering paper: five model families, two distinct application domains, and an appendix ablation of the projection component. The layer-selection criterion is a sensible idea, since brute-force layer search is a genuine practical annoyance in this literature.

Critical weaknesses
  1. There is no comparison to any existing steering method. The paper states explicitly that "our goal is to evaluate the efficacy of SVM-based steering rather than compare it with other approaches" (Sec. 1 and Sec. 3), and every table uses only the unsteered model as baseline. But the motivation of the paper is that supervised SVM-derived directions should improve on unsupervised mean-difference constructions. Without a head-to-head comparison against CAA-style difference-in-means or the unsupervised method of Cyberey and Evans (2025) under identical data and evaluation, the central claim is untested, and what remains ("steering can change model behavior") is already well established. This alone places the contribution below the acceptance bar for me.

  2. The experimental evidence is too fragile to support the conclusions even on its own terms. All numbers are single-run point estimates with one fixed seed (Appendix D), with no variance reported. There is no evaluation of general capability after intervention (no MMLU or similar), so the cost of steering is unknown. And the paper's own Appendix F shows that the projection adjustment, presented as part of the method, is actively harmful for some model-direction pairs (for Llama with C1, removing it changes JailbreakBench refusal from 0.58 to 0.008), meaning a core component does not behave consistently across the very models tested. The steering strength lambda also requires per-model manual tuning (OLMo becomes numerically unstable beyond lambda of 10), with no selection procedure given.

Questions for authors
  1. Can you provide a quantitative comparison with at least one unsupervised steering baseline, e.g. difference-in-means/CAA or Cyberey and Evans (2025), using the same data, layers, and benchmarks? Absent this, how should a reader conclude that the SVM construction is responsible for the reported gains?

  2. Can you report variance across random seeds for at least one representative setting, and any measurement of general capability (e.g. MMLU) under the lambda values used in Tables 1 and 2?

  3. In Sec. 3.1 the text says C1 "leads to a slight increase in disparity (-6.1 percent relative change)" for Qwen1.5 on race, but Table 1 shows disparity decreasing from 0.5154 to 0.4371 relative to the unsteered baseline. Please confirm the -6.1 percent is relative to the N direction and revise the text, which currently reads as an increase over baseline.

Reviewer_zmLwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SVM-Steer, a supervised and self-supervised activation steering framework that leverages Support Vector Machines (SVMs) to construct steering vectors and select intervention layers for controlling large language model (LLM) behavior without fine-tuning. The method explores three steering directions derived from SVM decision boundaries and support vectors, and evaluates their effectiveness in bias mitigation and censorship control across five open-weight LLMs. Experimental results demonstrate reductions in measured gender and racial bias, effective manipulation of refusal behavior, and the sensitivity of steering performance to intervention strength.

Critical strengths

1 The proposed method leverages lightweight SVM classifiers to construct steering vectors and identify intervention layers without modifying the pretrained model parameters, making it relatively straightforward to implement and computationally efficient.

2 The paper evaluates SVM-Steer on five LLMs from different model families and multiple benchmarks covering gender bias, racial bias, and censorship control, demonstrating its applicability across different behavioral objectives.

3 The paper investigates three steering directions, different intervention strengths, and the effect of projection adjustment. The observations regarding model-dependent sensitivity and mode collapse provide useful insights into the limitations of activation steering.

Critical weaknesses

1 The proposed framework mainly applies SVM classifiers to construct steering directions and select intervention layers. However, closely related work, such as LLM-VA [1], has already explored SVM-based identification of behavior-related directions and layer selection. Although the intervention mechanisms differ, the paper does not sufficiently clarify its methodological novelty beyond existing supervised representation-based steering approaches.

[1] Zhang H, Wang D, Liu Y, et al. LLM-VA: Resolving the Jailbreak-Overrefusal Trade-off via Vector Alignment[C]//Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026: 5760-5776.

2 The experiments only compare SVM-Steer against unsteered models, without including established activation steering baselines or recent approaches such as AlphaSteer [2], PID Steering (ICLR 2026), and Energy Landscape Steering [3]. Consequently, the reported improvements demonstrate that the proposed method can modify model behavior, but do not establish its effectiveness or competitiveness relative to existing techniques.

[2] Sheng L, Shen C, Zhao W, et al. Alphasteer: Learning refusal steering with principled null-space constraint[C]//International Conference on Learning Representations. 2026, 2026: 25861-25904.

[3]Jiang E H, Ou W, Liu R, et al. Mitigating Over-Refusal in Aligned Large Language Models via Inference-Time Activation Energy[C]//Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026: 37930-37950.

3 The effectiveness of the projection adjustment is highly inconsistent across models and steering directions, and removing it sometimes leads to substantially better results. Furthermore, the theoretical derivation only characterizes the modified SVM classification score, without establishing corresponding guarantees for the LLM's generated behavior. More systematic ablations and analysis are needed to clarify when and why the proposed intervention is beneficial.

Questions for authors

1 Will the source code and datasets be made publicly available in the future?

Reviewer_WZa6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SVM-Steer, an activation steering framework that uses SVMs trained on LLM hidden representations to construct steering directions. The method uses model-generated pseudo-labels to separate representations into positive and negative classes, then derives steering vectors from the SVM decision boundary.

Critical strengths

The proposed approach is relatively simple and computationally lightweight, as it does not require fine-tuning the underlying language models. The experiments also cover multiple model families and behavioral objectives, providing some empirical evidence that SVM-derived directions can influence model outputs.

Critical weaknesses
  1. My primary concern is the extremely limited novelty of the proposed approach. Training simple classifiers on LLM hidden representations to distinguish behavioral concepts is already known in a lot of literatures. Linear probes and other elementary classifiers, such as logistic regression, SVMs, and MLPs, have long been used to analyze representations in hidden space. Similarly, activation steering through linear directions is not a new idea. Neither the use of SVMs nor activation steering is novel, and their combination does not constitute a sufficiently substantial methodological contribution for ICLR.

  2. The paper provides no convincing evidence that SVMs offer any particular advantage over simpler alternatives. The authors state in Section 3 that they do not aim to compare against existing steering approaches, and use only unsteered models as baselines. This is a major limitation. Without comparisons against other steering methods, it is impossible to establish whether the proposed approach offers any meaningful improvement.

  3. The related work discussion is also insufficient. The paper does not adequately discuss how the proposed approach differs from prior work on probing and manipulating hidden representations.

  4. Another major concern is the reliance on the model's own outputs to construct pseudo-labels. The paper does not adequately justify why these model-generated labels should be treated as reliable indicators of the target behavioral properties.

Questions for authors

Could the authors clarify Equation (4)? The indicator appears to compare the sign of the projected activation with the continuous score sxs_x, rather than sign⁡(sx)\operatorname{sign}(s_x). I am not sure whether I misunderstand.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.