- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper asks whether several trained LoRAs can be converted into activation-space intervention vectors that control multiple behaviors at once: it writes model merging as a weight update parameterized by a behavior-strength vector , expands a layer's hidden state in around zero, and averages the resulting Jacobians over inputs to obtain an input-independent steering vector, with a Taylor remainder bound and a variance bound in terms of a mixed-Jacobian norm . The proposed HeRD-Merging adds low-rank interaction correction terms weighted by products of behavior strengths to the linear adapter sum, training the correction matrices by distribution matching and then optimizing their coefficients against distribution matching, curvature, cross-input variation, and magnitude. On Qwen2.5-VL-3B, Mistral-7B-Instruct-v0.3, and Gemma-4-E2B with three target behaviors (analogies, bullet points, exaggerated complex language), single-layer first-order steering averages 46.4% behavioral adherence with HeRD versus 42.6% for task arithmetic and 30.3% for the strongest non-first-order baseline, Linear-AcT, while HeRD weight merging averages 79.9% against 78.5% for DARE task arithmetic.
Critical strengths
The shift in perspective from weight adaptation to activation intervention is of research value. By mapping existing weight updates into activation space through the Jacobian, a concrete mathematical connection is established between model merging and control during inference. There is also a clear least-squares interpretation for averaging the input-dependent Jacobian. The fundamental structure provides a feasible and verifiable starting point for research and is worth further development.
Critical weaknesses
The analysis and the method do not operate on the same object. The error bounds in the main text are derived for a first-order expansion of the hidden state in the behavior-strength vector, but HeRD's correction terms are products of at least two behavior strengths, so both their values and first derivatives vanish at zero and they cannot contribute to the first-order term; as the paper acknowledges, the resulting steering vectors are then equivalent to task arithmetic. To let the corrections matter, the appendix switches to a linearization in the weights instead, which changes what is being approximated: the original analysis controls the curvature of the layer output as a function of the strengths, whereas the accuracy of the generalized intervention described in Appendix A.2 depends on the error incurred by linearizing in the weights, as well as the error from averaging over inputs. The two are not interchangeable, because the chain rule allows the curvature contributions to cancel, so small curvature in strength space does not imply a small weight-linearization error. The assertion that the generalization leaves the Hessian-regularization results unchanged is therefore unsupported, and no condition under which it would hold is stated. This is consequential because the paper's theoretical contribution is precisely the link from Hessian regularization to steering accuracy: as written, the stated bounds do not directly establish an error guarantee for that generalized intervention, and no experiment reports the actual approximation error or ablates the curvature and cross-input variation penalties that are supposed to drive the reported gains.
Questions for authors
-
Please confirm whether the HeRD steering experiments use from Appendix A.2 and specify the full intervention, including how it is averaged over inputs. If this expression is used, please provide applicable bounds for both the weight-linearization error and the input-averaging error, and explain how the regularization terms defined through derivatives with respect to control these errors.
-
Using the actual intervention implemented in the experiments, how do the activation approximation error and behavioral adherence change when is removed, when is removed, and when both are removed? Please keep the correction-matrix parameterization, first-stage training, second-stage initialization, and optimization budget fixed, changing only the inclusion of these loss terms. Please also report the resulting and values for all variants. These matched comparisons would help distinguish the effects of regularization from those of the additional parameters and training.
3 comments
Rebuttal by Authors (2/2)
Question 2
Using the actual intervention implemented in the experiments, how do the activation approximation error and behavioral adherence change when is removed, when is removed, and when both are removed?
We have now introduced Table 9, which provides the results of the ablation study that you requested. We summarize the results below by taking averages.
| Method | Analogy Use | Bulletpoint Use | Sophisticated Language | Average |
|---|---|---|---|---|
| First-Order Steering (HeRD-Merging) | 19.8 | 65.9 | 53.6 | 46.4 |
| First-Order Steering (Task Arithmetic) | 20.2 | 63.5 | 44.2 | 42.6 |
| Without | 17.3 | 63.9 | 47.4 | 42.9 |
| Without | 16.4 | 62.3 | 52.4 | 43.7 |
| Without and | 17.4 | 63.9 | 47.4 | 42.9 |
Notably, removal of one or both of the residuals noticeably reduces the performance of the method, suggesting that Hessian regularization indeed contributes to the performance of first-order steering with HeRD-Merging.
We have also computed the average approximation error and relative hessian norms (defined in Section C.2) before/after the Levenberg-Marquardt Algorithm (LMA) for each model.
| Qwen | Relative Hessian Norm (%) | Average Relative Approximation Error (%) |
|---|---|---|
| Before LMA | 35.45 | 16.31 |
| After LMA | 23.41 | 11.66 |
| Task arithmetic | 20.57 | 10.46 |
| Mistral | Relative Hessian Norm (%) | Average Relative Approximation Error (%) |
|---|---|---|
| Before LMA | 72.73 | 33.13 |
| After LMA | 52.59 | 25.27 |
| Task arithmetic | 45.00 | 21.28 |
| Gemma | Relative Hessian Norm (%) | Average Relative Approximation Error (%) |
|---|---|---|
| Before LMA | 42.60 | 23.65 |
| After LMA | 39.67 | 22.49 |
| Task arithmetic | 34.56 | 20.44 |
This suggests that the LMA is contributing meaningfully to reducing the first-order steering error, despite the initial error accumulated from the addition of the interaction matrices . We have added these results to Table 10 in our updated paper.
Please also report the resulting and values for all variants.
We provide the residuals with the same ablations below. Including the regularization terms generally reduces the corresponding residuals, although the differences are small for some models.
| Qwen | ||
|---|---|---|
| First-Order Steering (HeRD-Merging) | 0.6589 | 0.0329 |
| Without | 0.7078 | 0.0346 |
| Without | 0.6873 | 0.0342 |
| Without | 0.7078 | 0.0346 |
| Mistral | ||
|---|---|---|
| First-Order Steering (HeRD-Merging) | 1.7121 | 0.0686 |
| Without | 1.7125 | 0.0686 |
| Without | 1.7121 | 0.0686 |
| Without | 1.7125 | 0.0686 |
| Gemma | ||
|---|---|---|
| First-Order Steering (HeRD-Merging) | 3.0069 | 0.2474 |
| Without | 3.1135 | 0.2498 |
| Without | 3.0069 | 0.2474 |
| Without | 3.1135 | 0.2498 |
Here, we note that even if the values of some residuals and are the same between ablations (e.g. without vs. with all residuals), it is still possible for the overall behavioral adherence to change, as displayed in Table 9. This is because the computed values of are different between ablations, which produces different steering vectors and therefore produces different generations under greedy decoding. For reference, we include the computed values of for each variant and model below.
| Mistral | ||||
|---|---|---|---|---|
| All residuals | 0.4431 | 0.9254 | 0.1421 | 0.4232 |
| Without | 0.2072 | 0.2022 | 0.2212 | 0.0000 |
| Without | 0.2048 | 0.2027 | 0.2230 | 0.0000 |
| Without | 0.2072 | 0.2022 | 0.2212 | 0.0000 |
| Gemma | ||||
|---|---|---|---|---|
| All residuals | 0.8487 | 0.8382 | 0.7217 | 0.7455 |
| Without | 0.3531 | 0.8137 | 0.4957 | 0.0000 |
| Without | 0.9159 | 0.5082 | 1.0000 | 0.0000 |
| Without | 0.3531 | 0.8136 | 0.4956 | 0.0000 |
| Qwen | ||||
|---|---|---|---|---|
| All residuals | 0.0000 | 0.8835 | 0.4467 | 0.0029 |
| Without | 0.3449 | 0.5642 | 0.6515 | 0.3385 |
| Without | 0.3308 | 0.5592 | 0.6453 | 0.3731 |
| Without | 0.3449 | 0.5642 | 0.6515 | 0.3386 |
Request for Follow-Up Feedback on Rebuttal
Dear Reviewer QZpD,
Thank you once again for your feedback on our paper. We have responded to the points that you have identified in our rebuttal, and we would like to ask if our responses have sufficiently addressed your concerns. Please do not hesitate to let us know if there are any additional points which remain unclear. We greatly appreciate the opportunity to improve our paper.
Rebuttal by Authors (1/2)
Dear Reviewer QZpD,
We sincerely appreciate the thorough and constructive feedback you have provided on our paper. We have carefully incorporated your feedback and have responded to your suggestions below. Please let us know if there are any outstanding comments or concerns.
Question 1
Yes, the HeRD steering experiments use JWℓh(x,Wbase)[ΔW(s)], and we now explained this with more detail in Section 4.2 and Section A of our paper. Namely, we clarify that the approximation used to motivate first-order steering with HeRD-Merging is
hℓ′(x,s)≈hℓ(x)+JWℓhℓ′(x,0)[ΔW(s)].
The averaging over inputs is done by generating a steering function (which generalizes the steering matrix A∗ from Section 3) defined by
A∗(s)=M1∑m=1MJWℓhℓ′(xm,0)[ΔWℓ(s)]
so that the final intervention is
hℓ′(x,s)=hℓ(x)+A∗(s).
This information is now clearly outlined in Section 4.2 and Appendix A of our updated paper.
Because we previously clarified the intervention from first-order steering in the appendix rather than in the main paper, we did not use this notation to describe the residuals. Following your comments, we realized that this is ambiguous. We have therefore corrected this in Section 4.4 by replacing the residuals as follows.
Let s∗ be a unit vector constructed by sampling uniformly from {−1,1}n and then scaling by 1/n. We define D∗=ΔWℓ(s∗) so that
r2=∥JWℓhℓ′(x,Wℓbase)[D∗]∥∥JWℓ(JWℓhℓ′(x,Wℓbase)[D∗])[D∗]∥
and
r3=E[∥JWℓhℓ′(x,Wℓbase)[D∗]∥2]∥Covx∼D[JWℓhℓ′(x,Wℓbase)[D∗]]∥F.
Importantly, the residuals used for fitting the reported HeRD-Merging adapters do not use derivatives with respect to s. The applicable error bounds which motivate these residuals are now shown in Section 4.2 and Section A (Proposition 4).