acceptodds
Under review as a conference paper at ICLR 2027

Isotropic Steering of Large Language Models

Abstract

Activation steering modifies intermediate model activations during inference, providing a promising solution to intervene in the behavior of large language models (LLMs) without updating model parameters. However, interventions from existing steering methods often suffer from strong anisotropy, which harms steering robustness and transferability. To address this issue, we propose an Isotropic Steering (TRIS) method for LLMs that calibrates intervention strengths across different steering directions and different activations. In particular, TRIS preconditions the gradient of a steering objective with regularized covariance of activations and obtains normalized interventions in coordinates aligned with the activation distribution. From an optimization perspective, we show that TRIS is equivalent to Steepest Ascent under the Mahalanobis metric. Across four LLMs from the Mistral, Llama, and Qwen families, TRIS improves the primary task metric over ODESteer in 10 of 12 model–task comparisons. In addition, we find that covariance estimated on one task transfers effectively to other tasks within the same model, suggesting that TRIS captures reusable activation distribution and transfers well across tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM alignment, Activation Steering

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

The paper proposes TRIS, an activation-steering method that adjusts gradient updates using regularized activation variances. It formulates the update as steepest ascent under a Mahalanobis metric to calibrate intervention strength across directions and activations. Across four LLMs on truthfulness, helpfulness, and detoxification, TRIS outperforms ODESteer in 10 of 12 model–task comparisons. The experiments also examine sensitivity to steering strength, reuse of covariance estimates across tasks, and computational overhead.

Critical strengths
  1. The method is easy to understand and reasonably motivated. It adjusts steering updates to account for different activation scales, has a clear mathematical interpretation, and adds little computational cost.

  2. The comparison with ODESteer is controlled in some respects. Both methods use the same objective, fitting data, and integration steps. TRIS performs better in 10 of 12 model–task comparisons, providing some evidence that the change is useful.

Critical weaknesses
  1. Parameter selection and robustness to steering strength are not fully clarified. The paper reports the best scores over 50 values of TT and selects the shared TT using the evaluation scores, without describing a separate validation set. Moreover, TT has different units in TRIS and ODESteer, so using the same numerical range does not establish that TRIS is inherently easier to tune.

  2. The experiments do not isolate the benefit of covariance calibration. TRIS makes a relatively small change to ODESteer, modifying both the update direction and its magnitude without separating their effects. The diagonal estimator ignores correlations, so it need not remove the anisotropy shown in Figure 1. The theory establishes steepest ascent under the chosen metric, but does not show that this metric better captures the behavioral effects of steering. Covariance is also estimated at the last input token and applied throughout generation, without checking whether those statistics remain suitable for generated tokens.

  3. The detoxification gains come with higher perplexity. TRIS has higher perplexity than ODESteer on all four models, sometimes near the allowed limit of 30. The results therefore do not establish a better toxicity–quality trade-off at comparable perplexity.

Questions for authors
  • Were TT and the intervention layer selected independently of the final evaluation data? Do the gains and robustness results hold with validation-based selection and steering ranges adjusted for the methods’ different scales?

  • Can you separate the effects of changing direction and magnitude using a baseline that keeps the original gradient direction but normalizes its length under the same Mahalanobis metric? Could strong shrinkage partly explain the similarity of covariance estimates across tasks?

  • Does TRIS achieve lower toxicity at comparable perplexity? Can you provide toxicity–perplexity curves across steering strengths?

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

TRIS extends ODE-based activation steering by preconditioning the objective gradient with a regularized activation covariance and normalizing its Mahalanobis speed. Its practical implementation uses a shrunk, capped diagonal estimator. The paper derives a steepest-ascent interpretation and evaluates four LLMs on truthfulness, helpfulness, and detoxification, reporting improvements over ODESteer in 10 of 12 comparisons, shared-strength robustness, and cross-task covariance reuse. Efficiency measurements show that TRIS retains 87.7–95.3% of unsteered throughput and 98.3–99.7% of ODESteer's throughput.

Critical strengths
  1. Clear formulation: The update has a straightforward steepest-ascent interpretation, with explicit derivations connecting covariance calibration to intervention strength.

  2. Useful empirical coverage: Four-model, three-task evaluation uses shared fitting data and objectives; estimator ablations and covariance-transfer experiments provide evidence for reusable activation statistics.

  3. Measured efficiency: The diagonal estimator requires O(d) storage and updates. Figure 6 reports throughput over three timing repeats, showing little additional overhead relative to ODESteer.

Critical weaknesses
  1. Contribution and gain attribution: TRIS changes both update direction and the scale of T. Appendix A.4 shows that C=mI only rescales ODESteer, yet this scalar-only baseline is missing. Together with a normalization-only control, it would establish whether directional preconditioning adds value beyond strength retuning; identical numerical T grids do not ensure comparable effective interventions.

  2. Selection and statistical evidence: The paper scans 50 strengths and reports the best scores without specifying an independent tuning set. Layer selection uses TruthfulQA, while shared T is optimized over the evaluated model–task scores. Objective cross-fitting alone does not resolve selection bias. Held-out selection and paired confidence intervals are needed to assess the reported gains and robustness.

  3. Geometric interpretation: Unit speed in the chosen diagonal metric does not imply isotropy of the actual correlated activation distribution, nor output-level robustness. Figure 1's PCA spectrum does not isolate marginal-scale imbalance as the relevant problem. Residual-anisotropy measurements and integration-resolution checks would better connect the theoretical argument to the implemented method.

  4. Safety–utility trade-off: PPL≤30 is a feasibility threshold, not a quality-matched comparison. On Qwen3-14B, improved toxicity accompanies a PPL increase from 21.42 to 28.36 relative to ODESteer. Matched-quality trade-off curves and broader utility/over-refusal evaluation are needed; the toxicity-score aggregation rule should also be stated explicitly.

Questions for authors
  1. Are layers, task-specific T, and shared T selected independently of evaluation labels? Please specify the selection splits and report held-out results with paired uncertainty estimates.

  2. How do scalar-only C=mI and normalization-only controls perform under comparable effective-strength ranges and matched output quality?

  3. Does diagonal calibration measurably reduce the relevant anisotropy, and do conclusions persist with increased K? Please also clarify the toxicity aggregation rule and its quality trade-off.

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

TRIS is built on ODESteer's interesting view of activation steering as a continuous trajectory: rather than making one fixed adjustment, it moves an activation through small steps following the normalized gradient of a learned objective. The paper argues that because activations are highly anisotropic, unit Euclidean steps correspond to uneven intervention strengths across directions and activations. TRIS therefore changes the geometry of these steps, using regularized activation variances to adjust their direction and magnitude. Its update is steepest ascent under a Mahalanobis metric, implemented with an inexpensive regularized diagonal covariance estimate. Across four models and three tasks, TRIS improves over ODESteer in 10 of 12 primary-metric comparisons, with promising results when sharing steering strength across tasks and models and when reusing covariance across tasks.

Critical strengths
  1. A simple and mathematically justified modification with little additional computational cost

TRIS derives its update from a clear optimization problem, resulting in a closed-form update (eq 8) has unit Mahalanobis speed and is equivalent to normalized gradient ascent in corresponding transformed coordinates. These are coherent mathematical results under the stated assumption. With the diagonal covariance estimate, the update adds only O(Bd) work per integration step beyond gradient evaluation (B is the batch size and d the hidden dimension).

  1. Promising empirical improvements across models and tasks, including shared-strength settings.

TRIS improves over ODESteer in 10 of 12 primary-metric comparisons and achieves the best score among the evaluated methods in 9 of 12 (table 1). When each model uses one strength shared across tasks, TRIS outperforms ODESteer in all 12 model-task comparisons. Under a globally shared strength, mean loss relative to each method's own task-specific optima decreases substantially. These results suggest that TRIS may simplify strength selection in practice.

Critical weaknesses
  1. The source of the improvement is not sufficiently isolated.

The motivation relies on anisotropy across principal components, but the practical regularized diagonal estimator does not generally whiten the actual activation distribution. More importantly, the paper (appendix A4) shows that C^=mI\hat C=mI reduces TRIS to scalar-rescaled ODESteer, but does not include this baseline (table 3). Without it, the gains cannot be clearly attributed to directional calibration rather than overall scale normalization. Using the same numerical TT grid also does not ensure comparable intervention ranges.

  1. The parameter selection and statistical uncertainty need clearer reporting.

For ODESteer and TRIS, the paper reports the best scores from a 50-value T sweep without clearly identifying an independent validation set. Layer selection also uses TruthfulQA, while th selection of λ,κ\lambda, \kappa and other baselines' tuning budgets is insufficiently documented. Task results have no confidence invervals making smaller gains difficult to assess.

Questions for authors
  1. Could you add the λ=1(C^=mI)\lambda =1 (\hat C = mI) baseline to the main and shared-strength experiments, with T ranges rescaled accordingly?

  2. Which data were used to select the intervention layer, T,λT, \lambda and κ\kappa, and were evaluation examples excluded?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.