- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper improves activation steering through interval calibration, setting activations to target coordinates, and jointly controlling correlated directions in a transformed coordinate system. Experiments examine single- and multi-attribute control, including warmth and sycophancy, on Llama and Qwen models.
Critical strengths
- Setting target coordinates gives steering a useful operational meaning: activations can move different distances toward the same target. Joint control removes dependence on intervention order, with a simple construction and valid reconstruction guarantees.
- The experiments show a concrete benefit: average sensitivity for highly correlated traits falls from 1.508 to 0.827 on Llama. Open-ended generation and Qwen tests add useful evidence, including cases where the benefit is limited.
Critical weaknesses
- Order-independent coordinate updates do not guarantee independent behaviors: changing coordinate j still moves along the original direction u_j. The abstract should match the more careful limitations. Mixed Qwen results and Table 4 also show that orthogonalization's empirical advantage is conditional.
- The disclosed setup mismatch prevents a direct ranking against MAT-Steer. A clearer distinction from the cited K-Steering would help establish the method's place. These issues limit comparative claims, but do not invalidate the controlled additive/parameterized/orthogonalized comparisons.
- Appendix F's judge ignores correctness, completeness, and truncation. The open-ended results support trait control, but leave output quality unclear. Appendix C.2 should also avoid implying that changing a small subspace necessarily has negligible collateral effects.
- Endpoint coverage and slope do not guarantee a monotonic or linear interval. Clipping is therefore an empirical calibration heuristic. Likewise, zero denotes a midpoint between contrastive means, not necessarily neutral behavior; claims about universally interpretable strength need qualification.
- Minor corrections: Replace the Llama 2 citation “(?)” on page 3 and remove the doubled subject in “which ... it sets them” in the abstract. Theorem 1 should specify that invertibility requires retaining the residual, as the appendix does.
Questions for authors
- Can you state precisely which independence properties are guaranteed and which behavioral improvements are only observed empirically?
- What explains the differing benefit of orthogonalization across Llama, Qwen, and Table 4, and when should a user prefer it over parameterization alone?
- For the existing open-ended outputs, how does correctness or completeness change at comparable levels of target-trait control?
- How often do selected intervals violate monotonicity, and what interpretation should users give to zero and intermediate parameter values?
3 comments
Q3: Refined LLMAJ evals
For the existing open-ended outputs, how does correctness or completeness change at comparable levels of target-trait control?
Open-ended evaluation at intermediate steering levels. To examine trait control at intermediate settings of the second trait, we extended the LLM-as-a-judge evaluation to and . For each fixed , we measure the change in the fraction of responses judged to exhibit the trait as varies between and .
The results at are mixed: across the five trait pairs shown, the average ranges for parameterized and orthogonal parameterized steering are 2.0 and 1.6 percentage points at , and 1.5 and 1.2 percentage points at , respectively. These results concern the effectiveness of reaching the target behavior at comparable target-trait control.
Alpha-trait percent-steered range at
Assuming , here is the corrected table with it included for both beta settings:
| α Trait | β Trait | Corr. | β = −0.5: AI | β = −0.5: P | β = −0.5: OP | β = −0.5: ΔAI | β = −0.5: ΔP | β = +0.5: AI | β = +0.5: P | β = +0.5: OP | β = +0.5: ΔAI | β = +0.5: ΔP |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| coordination | power | 0.870 | 1.0% | 0.5% | 0.5% | −0.5% | 0.0% | 1.0% | 1.0% | 1.0% | 0.0% | 0.0% |
| power | coordination | 0.870 | 4.0% | 2.5% | 1.5% | −2.5% | −1.0% | 1.0% | 1.0% | 2.0% | 1.0% | 1.0% |
| corrigibility | coordination | 0.810 | 1.5% | 2.5% | 1.5% | 0.0% | −1.0% | 2.5% | 3.5% | 1.5% | −1.0% | −2.0% |
| coordination | corrigibility | 0.810 | 0.5% | 0.5% | 3.5% | 3.0% | 3.0% | 0.5% | 0.0% | 0.5% | 0.0% | 0.5% |
| power | wealth | 0.730 | 4.5% | 4.0% | 1.0% | −3.5% | −3.0% | 1.0% | 2.0% | 1.0% | 0.0% | −1.0% |
| Average | 2.3% | 2.0% | 1.6% | −0.7% | −0.4% | 1.2% | 1.5% | 1.2% | 0.0% | −0.3% |
Our existing multiple-choice evaluations provide complementary evidence on correctness. In Table 4, orthogonal parameterized steering improves the fraction of preferred labeled answers from 63.0% to 78.5% on TruthfulQA and from 59.5% to 69.5% on BBQ relative to the unsteered baseline, although parameterized steering scores higher on both. These evaluations do not measure correctness or completeness of the open-ended responses. We will make that limitation explicit and aim to extend the LLM-as-a-judge evaluation to assess correctness and completeness in the final version.
Q4: Do intervals violate monotonicity
How often do selected intervals violate monotonicity, and what interpretation should users give to zero and intermediate parameter values?
Monotonicity and interpretation of steering strength. As discussed in our response to PUzq, we seek a generally increasing steering curve but do not enforce strict monotonicity. The interval is selected on training data, while Figure 2 shows the resulting curves on held-out data; they are increasing over the selected range. Thus, monotonicity is an empirical goal rather than a guarantee.
After rescaling, and represent the endpoints of the selected steering interval, and is their midpoint. Zero therefore provides a reference within the calibrated interval, but it does not necessarily represent neutral behavior or the unsteered model. For some behaviors, neutrality may not be well-defined; nor would a neutral setting necessarily correspond to 50% behavior-matching answers, since that measure depends on the evaluation data distribution. Intermediate values interpolate between the endpoints, but do not guarantee a proportional change in measured behavior.
We will also correct the missing Llama 2 citation and the doubled subject in the abstract.
Thanks for your response. I have read it and kept my positive score.
Thank you for the feedback and questions. Sorry for delay, we had challenges accessing our shared GPUs.
Q1: Independence Guarantees
Independence guarantees. For linearly independent steering directions, our construction guarantees that, at the intervention layer, assigning one selected transformed coordinate leaves the other selected coordinates unchanged, and assignments to distinct coordinates commute. The procedure is idempotent. Reconstruction also preserves the activation’s component outside the span of the steering directions, provided that the residual is retained.
These are guarantees about activation coordinates, not independent behaviors. The linear representation hypothesis motivates associating the learned directions with concepts, but it does not guarantee that those directions capture the concepts perfectly or that later layers preserve behavioral separation. Whether changing one coordinate affects other measured behaviors is therefore an empirical question.
Our experiments show reduced behavioral interference in some settings, including the power–coordination example in Figure 5, but the results vary across traits and models. We will revise the abstract and Section 5 to distinguish these empirical findings from the coordinate-level guarantees.
Q2: Benefits of orthogonalization (across models)?
Parameterized steering sets an activation to a target coordinate rather than adding a fixed displacement. In the joint construction, assignments to distinct transformed coordinates commute and leave the other selected coordinates unchanged. By contrast, additive α-iterative steering does not gain this benefit from orthogonalization: adding a transformed steering vector and mapping back produces the same update as adding the original steering vector directly, so interference between the directions remains. These coordinate guarantees do not ensure behavioral independence, and the empirical benefit varies across settings.
As shown in Table 2 and illustrated in Figure 5 (right, top), just applying parameterized on two traits has the effect that the second steered trait partially overwrites the effect of steering the first trait.
Comparison with K-Steering. K-Steering also addresses multi-attribute control, using a trained nonlinear multi-label classifier and its gradients to construct interventions. Our approach instead starts from individual steering directions and sets targets in a joint coordinate system. This yields an explicit order-independence guarantee for the selected coordinates, while K-Steering can model nonlinear interactions through its classifier. We have not found an easy way to perform a direct empirical comparison, so we do not claim that either approach is generally more effective.
In Table 4, parameterized steering scores higher than orthogonal parameterized steering on all three benchmarks on the less stable Percent Steered measure. The comparison with MAT-Steer uses unmatched data splits and different baselines, so it does not support a direct ranking against MAT-Steer. Some of our no-steering scores start very high, leaving limited room for improvement on those datasets.