acceptodds
Under review as a conference paper at ICLR 2027

Directional Curvature Governs Task-Vector Transfer: Theory and Curvature Compensation

Abstract

Amid rapid iteration of pretrained models, task vectors—the parameter difference between a base model and its fine-tuned counterpart—offer a training-free transfer rule: adding the task vector learned on one model to another to impart similar task capabilities. However, prior empirical studies are mixed: task-vector transfer can succeed in some settings but fail dramatically in others. In this paper, under suitable regularity conditions, we theoretically characterize both when vanilla transfer is exact and how its suboptimality is controlled. Specifically, we prove that vanilla transfer is exact if and only if the model displacement lies in the null space of the path-averaged Hessian , i.e., the directional curvature condition . We further show that the violation magnitude controls the suboptimality of transfer and provide proxy-based empirical support for this relationship in neural networks. To overcome the intrinsic limitations of vanilla transfer, we propose the Curvature-compensated Task Vector (CurvTV), which enhances task-vector transfer by compensating for parameter deviations induced by violations of the directional curvature condition. Theoretically, CurvTV retains a quadratic optimality-gap envelope while admitting a sharper quartic refinement in the local regime, and this local gain persists under controlled Hessian estimation error. Extensive experiments across RoBERTa and Llama-family models demonstrate that CurvTV yields consistent improvements in task-averaged performance over vanilla task-vector transfer.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Task Vector, Transfer Learning, Directional Curvature

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_K3vnOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper theoretically characterizes when a task vector can be transferred exactly from one model to another under a proximal fine-tuning formulation. Specifically, exact transfer occurs if and only if the parameter difference between the two original models lies in the null space of the task-loss Hessian averaged along the path connecting their fine-tuned counterparts, under suitable assumptions. The authors further show that the magnitude of the violation of this condition bounds the transfer loss gap, and empirically demonstrate a strong correlation between a K-FAC-based proxy for this violation and the loss gap relative to direct fine-tuning. They also propose CurvTV, which introduces a curvature-based correction to vanilla task-vector transfer, and derive a tighter quartic upper bound on its optimality gap in the local regime, compared with the quadratic upper bound for vanilla transfer. Empirically, CurvTV outperforms vanilla task-vector transfer, Re-Basin, and GradFix on RoBERTa-family models, and improves upon vanilla task-vector transfer on Llama-3.1-8B-based models.

Critical strengths
  1. Clear motivation and relevance. The paper addresses a well-motivated and practically relevant problem: task-vector transfer works well in some settings but fails in others, and understanding when and why this happens is valuable.

  2. Insightful theoretical characterization. The paper provides an elegant theoretical characterization of when task-vector transfer is exact, offering a useful explanation for why task-vector transfer can succeed or fail under the stated assumptions.

  3. Theory-guided method. The proposed CurvTV method is well motivated by the theoretical analysis, directly addressing the identified source of transfer error. The additional theoretical guarantees on its optimality gap are also appreciated.

  4. Solid empirical evaluation. The experiments demonstrate the effectiveness of the proposed method across multiple models and downstream tasks, including language models with up to 8B parameters.

Critical weaknesses
  1. Mismatch between theoretical assumptions and practical fine-tuning. The theoretical analysis assumes proximal regularization toward the original pretrained model, α2∣θ−θi∣2\frac{\alpha}{2}\\|\theta-\theta_i\\|^2, whereas the experiments use standard weight decay, which corresponds to an L2 penalty λ2∣θ∣2\frac{\lambda}{2}\\|\theta\\|^2 under ordinary SGD. This distinction is consequential: replacing proximal regularization with standard L2 regularization changes the stationarity condition to (λI+Hˉ)(θ2⋆−θ1⋆)=0(\lambda I+\bar H)(\theta_2^\star-\theta_1^\star)=0. Under the nonsingularity assumption, this implies θ1⋆=θ2⋆\theta_1^\star=\theta_2^\star, so exact task-vector transfer would require θ1=θ2\theta_1=\theta_2, rather than the proposed directional curvature condition Hˉ(θ2−θ1)=0\bar H(\theta_2-\theta_1)=0.

    This mismatch also affects the proposed CurvTV method. In theory, α1\alpha_1 and α2\alpha_2 represent the proximal regularization strengths used during fine-tuning. However, the practical fine-tuning procedure does not use these proximal regularizers, and the authors instead select α1\alpha_1 and α2\alpha_2 through empirical performance tuning without establishing their connection to the actual fine-tuning procedure. This weakens the connection between the theoretical derivation and the proposed method. Although the authors acknowledge this limitation, a clearer theoretical justification or additional experiments using explicit proximal regularization would help bridge the gap between theory and practice.

  2. Local nature of the theoretical guarantee. The theoretical advantage of CurvTV over vanilla task-vector transfer is local, as acknowledged by the authors. Specifically, the quartic upper bound provides a stronger asymptotic guarantee only when the displacement between the two anchor models is sufficiently small. This limits the scope of the theoretical improvement, as task-vector transfer may also be applied between models with substantial parameter differences.

    A minor additional concern is the comparison of theoretical upper bounds between CurvTV and vanilla TV. The authors argue that CurvTV offers a theoretical advantage by comparing its quartic upper-bound branch in the local regime with the quadratic upper bound of vanilla TV in terms of their asymptotic orders. However, since explicit upper bounds are available for both methods, directly comparing these bounds would be more informative than comparing their asymptotic orders alone. Such a comparison yields an additional threshold on the anchor displacement, below which CurvTV's quartic branch provides a tighter upper bound than vanilla TV's quadratic bound. Beyond this threshold, the quartic bound is not necessarily tighter. Although this does not change the conclusion that CurvTV admits a tighter upper bound in a sufficiently local regime, explicitly discussing this threshold would make the theoretical comparison more precise.

  3. Minor figure presentation issues. In Figure 1(a), several data points are obscured by the legend. In Figure 5, some data points are obscured by the Pareto frontier curve.

Questions for authors

In Figure 2(a), only the values of α2\alpha_2 are explicitly indicated in the legend, although the text discusses trends with respect to both α1\alpha_1 and α2\alpha_2. How are the different values of α1\alpha_1 represented in the figure? It would be helpful to clarify this so that the claimed trends with respect to α1\alpha_1 can be verified.

2 comments

Q1: Proximal regularization versus practical fine-tuning

Thank you for this careful derivation, which is exactly why we use the proximal term to model the overall effect of standard fine-tuning in practice. Under L2 regularization alone, fine-tuning from any initialization would converge to the same solution, so the transfer problem would not arise. However, practical fine-tuning does not converge to the same solution. Techniques such as early stopping, a limited number of training steps and small learning rates keep the solution close to its initialization, which approximately acts as proximal regularization centered at the initialization [1]. The proximal objective captures this key effect while keeping the analysis simple, a modeling choice shared by related work such as Daheim et al. [2].

Moreover, our experiments already test the theory under standard fine-tuning. On RoBERTa, θ1⋆\theta_1^\star and θ2⋆\theta_2^\star are trained with weight decay, without a proximal term, for a fixed number of epochs [3]. In this setting, the proxy for the violation of the directional-curvature condition is highly correlated with the loss gap (rˉ=0.959\bar r=0.959), and CurvTV raises the average gain from 0.353 to 0.465 without increasing the native gap. This shows that the proximal model also explains standard fine-tuning.

Q2: Connection between α1,α2\alpha_1,\alpha_2 and fine-tuning

Thank you for this question. We would like to clarify that αi\alpha_i models the strength of the implicit regularization in the fine-tuning on each side, which is determined by factors such as the learning rate and the number of training steps. For least squares, for example, gradient flow run up to time tt has a risk close to that of ridge regression with coefficient 1/t1/t [1]. For deep networks this correspondence is only qualitative and αi\alpha_i cannot be computed in advance, so we treat α1\alpha_1 and α2\alpha_2 as hyperparameters.

Q3: Improvement beyond small displacements

Thank you for raising this point. We agree that the fourth-order improvement of CurvTV is local. However, an improvement that does not depend on the size of the displacement follows directly from our existing results. Let u=θ2⋆−θ1⋆u=\theta_2^\star-\theta_1^\star, and let H~\tilde H be the curvature used by CurvTV. Comparing the CurvTV upper bound at θ~CurvTV\tilde\theta_{\mathrm{CurvTV}} with the vanilla lower bound given by the identity in Thm. 3.3 shows that, under the conditions of that theorem, if H~⪰0\tilde H\succeq0 and ∣(Hˉ−H~)u∣<α/(α+β)∣Hˉu∣,\\|(\bar H-\tilde H)u\\|<\sqrt{\alpha/(\alpha+\beta)}\\|\bar Hu\\|, then L2(θ~CurvTV)<L2(θtv)\mathcal L_2(\tilde\theta_{\mathrm{CurvTV}})<\mathcal L_2(\theta_{\mathrm{tv}}). This condition does not involve the size of the displacement. It only requires the curvature estimate to approximate the path-averaged curvature along uu with relative error below α/(α+β)\sqrt{\alpha/(\alpha+\beta)}. Therefore, even when the two anchor models differ substantially, CurvTV strictly outperforms vanilla TV as long as the curvature changes little along the path. In the quadratic case with the exact Hessian, H~=Hˉ\tilde H=\bar H and CurvTV exactly recovers θ2⋆\theta_2^\star for displacements of any size, so the condition holds whenever vanilla TV is not exact (Hˉu≠0\bar Hu\neq0). In the revision, we will state this sufficient condition after Thm. 4.4.

Q4: Direct comparison of the upper bounds

Thank you for this helpful suggestion. We directly compare the fourth-order branch of Thm. 4.4 with the vanilla upper bound in Thm. 4.3, which gives

\iff \\|\theta_2-\theta_1\\|<\frac{2\\|\psi_\alpha(\bar H)\\|}{\rho\Upsilon_\alpha}.$$ That is, the CurvTV upper bound is strictly smaller than the vanilla upper bound whenever the error of CurvTV caused by the change of the Hessian along the path is small compared with the error of vanilla TV caused by ignoring curvature. Hence, the more slowly the Hessian changes along the path, the larger the displacement for which CurvTV has the smaller upper bound. Beyond the threshold, one can instead use the sufficient condition in our response to Q3, which does not depend on the size of the displacement. In the revision, we will state this threshold after Thm. 4.4.

Q5: Figure presentation

Thank you for the careful reading of our figures. We would like to clarify that each curve in Fig. 2(a) fixes one value of α2\alpha_2 and connects five points for α1∈0.05,0.1,0.5,1,5\alpha_1\in\\{0.05,0.1,0.5,1,5\\} in increasing order. Thus, the increase of the native gap with α1\alpha_1 for fixed α2\alpha_2 corresponds to moving right along a single curve, whereas its decrease with α2\alpha_2 for fixed α1\alpha_1 is seen by comparing corresponding points across curves. In the revision, we will label the α1\alpha_1 values in the figure.

Thank you also for pointing out the obscured data points. In the revision, we will move the legend of Fig. 1(a) away from the data points, and draw the Pareto frontier in Fig. 5 beneath the data points and make it semi-transparent.

References

[1] Ali et al. A Continuous-Time View of Early Stopping for Least Squares. AISTATS 2019.

[2] Daheim et al. Model Merging by Uncertainty-Based Gradient Matching. ICLR 2024.

[3] Yu et al. Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch. ICML 2024.

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

The paper studies the transfer problem of applying a task vector learned on one model to a new model of the same architecture. Under a proximal fine-tuning model, it shows that transfer is exact when the path-averaged Hessian cancels the displacement between the two base models at hand. It proposes CurvTV, a transfer methodology that corrects the transferred model using a K-FAC estimate of the source Hessian, and proves that for CurvTV has a smaller optimality-gap bound than vanilla transfer: O(∣∣d∣∣4)\mathcal O(||d||^4) instead of O(∣∣d∣∣2)\mathcal O(||d||^2), where d=θ2−θ1d = \theta_2 - \theta_1. Experiments on RoBERTa and Llama-3.1-8B show gains over vanilla task-vector transfer.

Critical strengths
  1. The question is relevant and interesting, and prior empirical results are mixed.
  2. Mathematical derivations are correct. Ablations are interesting and correctly framed.
  3. The paper is clearly written. The 2D example (Fig. 3) and the "Newton step + merge" view (App. C.2) give useful intuition.
Critical weaknesses

Theory

T1. The "exactness" condition is presented too strongly. Hˉ\bar H is averaged on the path from θ1⋆\theta_1^\star to θ2⋆\theta_2^\star, so checking Hˉ(θ2−θ1)=0\bar H(\theta_2 - \theta_1) = 0 requires θ2⋆\theta_2^\star, the model transfer is meant to avoid. The condition is also equivalent to ∇ℓτ(θtv)=∇ℓτ(θ1⋆)\nabla \ell_\tau(\theta_{\mathrm{tv}}) = \nabla \ell_\tau(\theta_1^\star) (i.e., to θtv\theta_{\mathrm{tv}} being stationary for the target). I see this as a useful reformulation, and would suggest to present it as an interpretation (e.g., "transfer works when the loss is flat along the displacement"), keeping the "if and only if" language for the quadratic case, which is stronger.

T2. CurvTV is a first-order linearization (this should be stated). With α1=α2=α\alpha_1 = \alpha_2 = \alpha, CurvTV equals

θCurvTV=θ1⋆+α(αI+H1⋆)−1(θ2−θ1).\theta_{\mathrm{CurvTV}} = \theta_1^\star + \alpha(\alpha I + H_1^\star)^{-1}(\theta_2 - \theta_1).

This is the first-order Taylor expansion of the fine-tuning map θ↦θ⋆(θ)\theta \mapsto \theta^\star(\theta), whose Jacobian is α(αI+H)−1\alpha(\alpha I + H)^{-1}. Vanilla transfer is the same formula with the Jacobian replaced by II.

The robustness result (Cor. 4.5) keeps this rate only if the Hessian error shrinks with ∥d∥\|d\|, i.e., ∥Δ∥≤ϵ∥d∥\|\Delta\| \le \epsilon \|d\|. K-FAC is in practice computed at θ1⋆\theta_1^\star and never sees θ2\theta_2 (respecting the experimental protocol), hence its error does not shrink. With a fixed error, the rate is O(∥d∥2)O(\|d\|^2), so the same as vanilla. I would state the linearization view and the robustness result for a fixed error.

T3. The bounds are loose. The bounds in Thm 3.3 differ by a factor of (α+βα)3\left(\frac{\alpha+\beta}{\alpha}\right)^3 and assume local convexity. CurvTV's quadratic bound (Thm 4.4) can be larger than vanilla's.

Experiments

E1. It is unclear whether curvature drives the gains.

  • The Identity ablation (Table 3) replaces K-FAC with II. With (α1,α2)=(0.1,0.5)(\alpha_1, \alpha_2) = (0.1, 0.5) this gives θ≈0.67 θ1⋆+0.33 θ2\theta \approx 0.67\,\theta_1^\star + 0.33\,\theta_2, which is close to weight averaging and recovers most of CurvTV's improvement over TV. Curvature helps with the "native gap", not with the gain.
  • The reported configuration (α1≠α2\alpha_1 \neq \alpha_2) is not covered by the theory, which assumes α1=α2\alpha_1 = \alpha_2. The shared-α\alpha variant has more gain but less native ability, so the native-gap advantage comes from α1≠α2\alpha_1 \neq \alpha_2 (which is a heuristic).

I would suggest to compare against vanilla TV with λ\lambda selected under the same native-gap constraint used for CurvTV and explain the role of α1≠α2\alpha_1 \neq \alpha_2.

E2. Comparison with standard adaptation using the same data. CurvTV uses N=1000N = 1000 labeled examples, plus a validation split for tuning, and needs forward-backward passes. Collecting K-FAC statistics costs roughly as much as two fine-tuning epochs on those examples. With the same data and compute, one could fine-tune the target directly. What are the pros and cons in this case? Fine-tuning coul be done in different ways (e.g., linear probing, tuning the last few blocks, LoRA, TV + fine-tuning, etc).

E3. The LLM experiments are limited. Only vanilla TV is compared, and the choice of α\alpha's is not described. Native ability is not measured.

Questions for authors

None

5 comments

Q1: Presentation of the exactness condition

We agree, and thank you for the suggestion. In the revision, we will refer to the directional-curvature condition as a characterization of exact transfer rather than a criterion, and keep the "if and only if" statement only for the quadratic case (Hˉ=H\bar H=H). The equivalent form you suggested is easier to check in practice, and we will add it to Sec. 3.1. At the same time, the directional-curvature form remains useful because it defines an approximately computable violation measure that we use to analyze transfer and to derive CurvTV.

Q2: Linearization view of CurvTV

Thank you for this useful perspective. Fine-tuning from θ\theta can be viewed as a map θ↦θ⋆(θ)\theta\mapsto\theta^\star(\theta), whose Jacobian at θ1\theta_1 is α(αI+H1⋆)−1\alpha(\alpha I+H_1^\star)^{-1} by the stationarity condition and the implicit function theorem. Vanilla TV is then the zeroth-order expansion of the task vector Δ(θ)=θ⋆(θ)−θ\Delta(\theta)=\theta^\star(\theta)-\theta, i.e., θ2⋆≈θ1⋆+(θ2−θ1)\theta_2^\star\approx\theta_1^\star+(\theta_2-\theta_1), whereas CurvTV with α1=α2\alpha_1=\alpha_2 is exactly the first-order expansion of θ⋆(θ)\theta^\star(\theta), i.e., θ2⋆≈θ1⋆+α(αI+H1⋆)−1(θ2−θ1)\theta_2^\star\approx\theta_1^\star+\alpha(\alpha I+H_1^\star)^{-1}(\theta_2-\theta_1). Higher-order expansions would require derivatives of the Hessian, which are unaffordable for large models, so CurvTV is the highest-order correction that is computable in practice. In the revision, we will state this view as a Remark in Sec. 4.1.

Q3: Robustness under a fixed Hessian error

Thank you for this careful observation. We agree that with a fixed error, the upper bound of Cor. 4.5 is O(∣d∣2)O(\\|d\\|^2), the same as for vanilla TV, so comparing rates alone cannot tell which method is better. However, a sufficient condition follows directly from our existing results. Let u=θ2⋆−θ1⋆u=\theta_2^\star-\theta_1^\star, and let H~\tilde H be the curvature used by CurvTV. Comparing the CurvTV upper bound at θ~CurvTV\tilde\theta_{\mathrm{CurvTV}} with the vanilla lower bound given by the identity in Thm. 3.3 shows that, under the conditions of that theorem, if H~⪰0\tilde H\succeq0 and ∣(Hˉ−H~)u∣<α/(α+β)∣Hˉu∣,\\|(\bar H-\tilde H)u\\|<\sqrt{\alpha/(\alpha+\beta)}\\|\bar Hu\\|, then L2(θ~CurvTV)<L2(θtv)\mathcal L_2(\tilde\theta_{\mathrm{CurvTV}})<\mathcal L_2(\theta_{\mathrm{tv}}). This condition does not require the error to shrink with ∣d∣\\|d\\|, only that the curvature estimate approximates the path-averaged curvature along uu with relative error below α/(α+β)\sqrt{\alpha/(\alpha+\beta)}. In the quadratic case with the exact Hessian, H~=Hˉ\tilde H=\bar H and CurvTV exactly recovers θ2⋆\theta_2^\star, so the condition holds whenever vanilla TV is not exact (Hˉu≠0\bar Hu\neq0). In the revision, we will explicitly state the fixed-error result and this sufficient condition.

Q4: Tightness of the bounds

Thank you for pointing this out, and these bounds can indeed be tightened. For Thm. 3.3, if the violation ∣Hˉ(θ2−θ1)∣\\|\bar H(\theta_2-\theta_1)\\| is replaced by the transfer error ∣ψα(Hˉ)(θ2−θ1)∣\\|\psi_\alpha(\bar H)(\theta_2-\theta_1)\\|, the same proof gives upper and lower bounds that differ only by a factor of (α+β)/α(\alpha+\beta)/\alpha. We state the theorem in terms of the violation because its purpose is to show that violating the directional-curvature condition controls the loss gap, which the transfer-error form does not show.

For Thm. 4.4, the quadratic branch exceeding the vanilla upper bound does not imply that CurvTV is worse. Since comparing two upper bounds cannot tell which method is better, we instead compare the CurvTV upper bound with the vanilla lower bound, which gives a sufficient condition for CurvTV to strictly outperform vanilla TV (see our response to Q3).

Q5: Role of α1≠α2\alpha_1\neq\alpha_2

Thank you for this suggestion. We would like to clarify that Sec. 4.2 uses equal coefficients only to keep the theoretical insights easy to follow, and the derivation extends directly to α1≠α2\alpha_1\neq\alpha_2, including the CurvTV method, Thm. 4.4 and Cor. 4.5. It suffices to replace α\alpha by α2\alpha_2 and θ2−θ1\theta_2-\theta_1 by the effective displacement d~:=θ2−θ1−(1−α1/α2)(θ1⋆−θ1)\tilde d:=\theta_2-\theta_1-(1-\alpha_1/\alpha_2)(\theta_1^\star-\theta_1). In the revision, we will give this extension in the appendix.

As for its role, αi\alpha_i models the strength of the implicit regularization by which techniques such as early stopping, a limited number of training steps and small learning rates keep the fine-tuned solution close to its initialization. Even under the same procedure, this strength depends on the model and the loss landscape around its initialization, so forcing α1=α2\alpha_1=\alpha_2 would limit the applicability of CurvTV.

Q6: Role of curvature beyond weight averaging

Thank you for the careful analysis. As you note, replacing K-FAC with II reduces CurvTV to 11+α2θ1⋆+α21+α2θ2+α11+α2(θ1⋆−θ1)\frac{1}{1+\alpha_2}\theta_1^\star+\frac{\alpha_2}{1+\alpha_2}\theta_2+\frac{\alpha_1}{1+\alpha_2}(\theta_1^\star-\theta_1), which for (α1,α2)=(0.1,0.5)(\alpha_1,\alpha_2)=(0.1,0.5) is 23θ1⋆+13θ2+115(θ1⋆−θ1)\frac23\theta_1^\star+\frac13\theta_2+\frac{1}{15}(\theta_1^\star-\theta_1), close to weight averaging.

However, transfer cannot be judged by gain alone, since otherwise directly using θ1⋆\theta_1^\star would be best, even though it loses most of the native capability. Following your suggestion, we therefore sweep λ\lambda for vanilla TV and for linear interpolation (LI) λθ1⋆+(1−λ)θ2\lambda\theta_1^\star+(1-\lambda)\theta_2, and compare both metrics. The table below reports averages over the 24 target–task pairs on RoBERTa.

θ1⋆\theta_1^\starLI (λ=0.2\lambda=0.2)LI (λ=0.4\lambda=0.4)LI (λ=0.6\lambda=0.6)LI (λ=0.8\lambda=0.8)TV (λ=0.6\lambda=0.6)TV (λ=0.8\lambda=0.8)TV (λ=1.0\lambda=1.0)CurvTV
Gain ↑\uparrow0.5250.0340.1710.3830.5010.2210.2960.3530.465
Native gap ↓\downarrow0.3950.0000.0250.1010.2520.0310.0810.1510.146

Vanilla TV stays well below the gain of CurvTV even at λ=1.0\lambda=1.0, where its native gap already exceeds that of CurvTV. LI has a lower gain than CurvTV at λ=0.6\lambda=0.6 and a much larger native gap at λ=0.8\lambda=0.8. These sweeps show that, without curvature, tuning a single scalar coefficient cannot match CurvTV on both metrics, whereas curvature lets CurvTV transfer more task knowledge while preserving more native capability. The effect of curvature is even clearer on Llama-3.1, where, for the math task vector transferred to Llama-3.1-abliterated, Identity has a much larger native gap than K-FAC (0.093 vs. 0.038) on the Korean benchmark KMMLU [1], which tests Korean knowledge across 45 subjects, mostly unrelated to the transferred math task. In the revision, we will include these new results.

Q7: Comparison with direct fine-tuning

Thank you for this comment. We would like to clarify that our implementation collects the K-FAC statistics in a single pass over the NN samples without optimizer updates, so its cost is at most that of one fine-tuning epoch on these samples. More importantly, CurvTV computes the K-FAC statistics only once. They are computed at θ1⋆\theta_1^\star and depend only on the source task, so they can be reused for every target model, including models released later. In contrast, although directly fine-tuning the target, including with parameter-efficient fine-tuning (PEFT), usually performs better, it must be rerun and retuned for every new target model, so its total cost grows as models iterate. CurvTV therefore trades a low one-time cost, further amortized across target models, for substantial gains (the average gain rises from 0.353 to 0.465 on RoBERTa).

Q8: LLM baselines and native capability

Thank you for these valuable suggestions. Following them, we add GradFix as a baseline on Llama and measure native capability. Re-Basin is not directly applicable to Llama, because its permutation scheme assumes that the query, key and value projections have the same number of heads, which does not hold under the grouped-query attention of Llama. We transfer the math task vector to Llama-3.1-abliterated and measure native capability by the macro-average accuracy on the Korean benchmark KMMLU, which tests Korean knowledge across 45 subjects, mostly unrelated to the transferred math task. For all methods, we reuse the hyperparameters selected on RoBERTa without retuning.

Vanilla TVGradFixCurvTV
MATH-500 gain ↑\uparrow−0.07−0.060.13
GSM8K gain ↑\uparrow−0.470.180.30
KMMLU native gap ↓\downarrow0.100−0.0130.038

GradFix keeps only the entries of the task vector whose signs agree with the gradient of the target model, which is overly conservative. It therefore preserves native capability but transfers much less of the math capability, with a negative gain on MATH-500. CurvTV achieves the highest gains on both benchmarks while keeping the native gap small. In the revision, we will include these new results and describe the hyperparameter setting in App. D.

References

[1] Son et al. KMMLU: Measuring Massive Multitask Language Understanding in Korean. NAACL 2025.

Thank you for the detailed rebuttal and the new experiments. Several of my concerns are addressed, and the proposed changes to the theory are welcome. The new sweeps also show that, at a similar native gap, vanilla TV gains less than CurvTV, so I withdraw the statement that scaling the task vector explains the gains.

However, I keep my score. The revised framing, which I appreciate, also makes the scope of the contribution clearer. Once the directional-curvature condition is presented as a characterization rather than a criterion, and CurvTV as the first-order linearization of the fine-tuning map, the theory appears as a clean application of standard tools to cross-model transfer, built on the same proximal model of fine-tuning used in prior work (e.g., Daheim et al., ICLR 2024). I find this interesting but incremental.

Additionally, the specific benefit of curvature remains experimentally modest: CurvTV is above the (estimated) linear interpolation frontier by a moderate margin at N=1000N = 1000 and slightly below it at N≤100N \le 100. Against the Identity variant, curvature adds about 0.02 in gain and reduces the native gap by about 0.01 on RoBERTa. The clearer effect on Llama comes from a single target model and task, and native ability is still not measured for the other Llama targets. Finally, the practical setting is narrow: the method needs source-side task data and storing K-FAC factors whose size is not reported (easily bigger than the model), and, as the authors acknowledge, fine-tuning (including PEFT) is usually better when target data is available. A comparison with fine-tuning on the same NN samples is still missing.

Reviewer_b7oaOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates task-vector transfer across pretrained models and establishes a directional-curvature condition for exact transfer under proximal fine-tuning assumptions. The authors propose CurvTV, which uses K-FAC-based curvature compensation to correct transferred parameters. Theoretical optimality-gap bounds and experiments on RoBERTa and Llama-3.1 models support the proposed approach.

Critical strengths

The paper provides a principled theoretical characterization of task-vector transfer, establishing a necessary and sufficient directional-curvature condition for exact transfer under explicit assumptions and deriving bounds on transfer suboptimality. CurvTV translates the theoretical analysis into a practical compensation method, demonstrating improvements across RoBERTa and Llama-3.1 models, particularly in challenging transfer settings where vanilla task vectors perform poorly.

Critical weaknesses

W1. The theory assumes proximal regularization centered at pretrained weights, whereas experiments use weight decay; the authors should theoretically justify this modeling choice or empirically compare the directional-curvature condition under proximal and standard fine-tuning.

W2. Table 3 shows that identity-based compensation achieves comparable gains; the authors should compare K-FAC with simpler curvature approximations under matched computational budgets, particularly on Llama-3.1, reporting both transfer performance and computational overhead.

W3. Although curvature-estimation sample sizes are varied, the authors should further investigate how approximation accuracy and effective curvature rank affect transfer performance, identifying the minimum second-order information required for effective compensation.

W4. Although Appendix B.4 derives a residual-corrected condition, the authors should measure proximal stationarity residuals in actual fine-tuned models and examine whether these residuals explain transfer errors or affect the applicability of the theoretical guarantees.

W5. Limited evaluation across modalities.The experiments focus exclusively on language models; the authors should evaluate CurvTV on vision tasks using compatible pretrained architectures to assess whether its effectiveness extends beyond language models.

Questions for authors

Please address all concerns raised in Weaknesses W1–W5.

6 comments

Q1: Justification of the proximal objective

Thank you for raising this point. We would like to clarify that the proximal term models the overall effect of standard fine-tuning in practice, rather than weight decay alone. Under weight decay alone, fine-tuning from any initialization would converge to the same solution. However, practical fine-tuning does not, because techniques such as early stopping, a limited number of training steps and small learning rates keep the solution close to its initialization. This approximately acts as proximal regularization centered at the initialization [1]. The proximal objective captures this effect while keeping the analysis simple, a modeling choice shared by related work such as Daheim et al. [2].

Moreover, our experiments already test the theory under standard fine-tuning. On RoBERTa, θ1⋆\theta_1^\star and θ2⋆\theta_2^\star are trained with weight decay, without a proximal term, for a fixed number of epochs. In this setting, the proxy for the violation of the directional-curvature condition is highly correlated with the loss gap (rˉ=0.959\bar r=0.959), and CurvTV raises the average gain from 0.353 to 0.465 without increasing the native gap. This shows that the proximal model also explains standard fine-tuning.

Q2: Simpler curvature approximations on Llama-3.1

Thank you for this valuable suggestion. Following it, we compare K-FAC with the simplest curvature approximation, the identity, on Llama-3.1. We transfer the math task vector to Llama-3.1-abliterated and measure native capability by the macro-average accuracy on the Korean benchmark KMMLU [3], which tests Korean knowledge across 45 subjects, mostly unrelated to the transferred math task.

Vanilla TVIdentityCurvTV (K-FAC)
MATH-500 gain ↑\uparrow−0.070.090.13
GSM8K gain ↑\uparrow−0.470.260.30
KMMLU native gap ↓\downarrow0.1000.0930.038
Curvature statistics——64.5 min
Transfer time17.6s22.2s30.2s

Identity already avoids the negative transfer of vanilla TV, but it loses almost as much native capability as vanilla TV. In contrast, K-FAC achieves higher gains on both benchmarks while reducing the native gap by more than half. As for the overhead, the transfer step of all three methods takes at most 30.2s, and the extra cost of K-FAC lies in collecting the curvature statistics, which takes 64.5 min on two A100 GPUs with a peak memory of 79 GiB per GPU. Importantly, these statistics are computed at θ1⋆\theta_1^\star and depend only on the source task, so they are computed once and reused for every target model, including models released later. This one-time cost is thus amortized across target models and is worthwhile for the higher gains and the substantially smaller native gap. In the revision, we will include these new results.

Q3: Approximation accuracy and curvature rank

Thank you for this suggestion. We agree that directly measuring the approximation accuracy would be informative, but it requires the exact Hessian, which is unfortunately out of reach for deep networks. For RoBERTa alone, the Hessian over its 85M encoder parameters has about 7×10157\times10^{15} entries, or about 29 PB in single precision. Instead, we vary the number of curvature samples NN, since a smaller NN gives a less accurate K-FAC estimate. On RoBERTa, the gain of CurvTV increases steadily from 0.387 to 0.465 as NN grows from 50 to 1000, so a more accurate curvature estimate leads to better transfer. As for the minimum second-order information, N=50N=50 samples already suffice to outperform vanilla TV on both metrics (gain 0.387 vs. 0.353, native gap 0.129 vs. 0.151), whereas removing curvature altogether is not enough to preserve native capability (see our response to Q2).

Regarding the effective rank, we would like to clarify that K-FAC already represents the curvature of each layer by two small Kronecker factors, which is itself a structured low-cost approximation. Its main cost lies in collecting activation and gradient statistics over the NN samples, so truncating the rank of the factors would not reduce this cost but would discard curvature information.

Q4: Proximal stationarity residuals

We appreciate this suggestion. We would like to clarify that the proximal stationarity residual ri=∇ℓτ(θi⋆)+αi(θi⋆−θi)r_i=\nabla\ell_\tau(\theta_i^\star)+\alpha_i(\theta_i^\star-\theta_i) is not uniquely defined for practically fine-tuned models. As noted in our response to Q1, the proximal objective is our model of practical fine-tuning rather than the objective it actually optimizes, so rir_i depends on the chosen αi\alpha_i and its absolute size cannot indicate whether the theory applies. What App. B.4 shows is that only the difference r1−r2r_1-r_2 enters the transfer condition.

In fact, our experiments also indicate that the residuals neither dominate the transfer error nor undermine the applicability of the theory. In the correlation analysis, θ1⋆\theta_1^\star and θ2⋆\theta_2^\star are not trained to stationarity, yet the proxy for the violation of the directional-curvature condition, which ignores the residuals, is highly correlated with the loss gap (rˉ=0.959\bar r=0.959). Moreover, CurvTV, which is derived under exact stationarity, still substantially improves the average gain over vanilla TV on RoBERTa.

Q5: Vision experiments

Thank you for this suggestion. Following it, we add experiments with CLIP ViT-B/16, using the model pretrained on CommonPool-L as the source and the model pretrained on DataComp-L as the target. We use the source models fine-tuned on DTD, EuroSAT, GTSRB and SVHN released by Rinaldi et al. [4], and all methods reuse the hyperparameters selected on RoBERTa without retuning. The table below reports the top-1 accuracy (%) of the transferred model.

MethodDTDEuroSATGTSRBSVHNAvg.
θ2\theta_2 (zero-shot)47.6149.3043.2246.1346.56
Vanilla TV47.2341.4137.9523.8237.60
Re-Basin47.3454.0943.6446.3047.84
GradFix49.1871.5754.0161.1758.98
CurvTV53.5172.0877.3069.2868.04

Vanilla TV falls below the zero-shot target on all four datasets, and Re-Basin improves it only slightly on average. CurvTV achieves the highest accuracy on every dataset and raises the average accuracy from 46.56 to 68.04, compared with 58.98 for GradFix. This shows that CurvTV also applies beyond language models. In the revision, we will include these new results.

References

[1] Ali et al. A Continuous-Time View of Early Stopping for Least Squares. AISTATS 2019.

[2] Daheim et al. Model Merging by Uncertainty-Based Gradient Matching. ICLR 2024.

[3] Son et al. KMMLU: Measuring Massive Multitask Language Understanding in Korean. NAACL 2025.

[4] Rinaldi et al. Update Your Transformer to the Latest Release: Re-Basin of Task Vectors. ICML 2025.

Thank you for your detailed response and additional experiments. I appreciate the authors' efforts, but I believe the following concerns remain insufficiently addressed.

W1. While I understand that the proximal objective is intended to approximate practical fine-tuning, the current explanation does not sufficiently justify this approximation for general nonconvex neural networks under finite-step optimization. I would appreciate a more direct empirical validation. For example, under the same RoBERTa tasks and model pairs, the authors could compare proximal fine-tuning and standard fine-tuning in terms of the directional-curvature violation proxy, transfer loss gap, and the relationship between these two quantities. This would help clarify whether the theoretical characterization remains applicable under standard fine-tuning.

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

This paper investigates the theoretical conditions under which task vectors learned from one pretrained model can be transferred to another.

The authors derive a necessary and sufficient condition for exact transfer in terms of the path-averaged Hessian and propose CurvTV, a curvature-compensated task-vector transfer method. Under local regularity assumptions, they establish a refined fourth-order upper bound on the optimality gap and report empirical improvements on RoBERTa and Llama-based models.

Critical strengths

The paper offers an interesting geometric interpretation of task-vector transfer through directional curvature.

The theoretical formulation is mathematically motivated, and the exact-transfer condition provides useful insight into why naive task-vector addition may fail. The proposed correction is derived systematically, and the experiments cover multiple pretrained models, tasks, and ablation settings.

Critical weaknesses

There are a few critical weaknesses from my point of view:

  1. The analysis relies on explicit proximal objectives, exact stationarity, local positive-definiteness of the regularized Hessian, and Lipschitz continuity of the Hessian. These assumptions are quite strong and difficult to verify for modern neural networks, particularly large language models. The paper does not sufficiently establish how its theoretical setting relates to practical fine-tuning procedures or whether the required conditions are satisfied in realistic model-transfer scenarios.

  2. The theoretical results concern the regularized proximal objective, whereas the experiments evaluate unregularized task losses and downstream performance. Improvement in the former does not necessarily imply improvement in the latter, even for simple convex losses. Furthermore, the implementation replaces the true Hessian with an empirical-Fisher/K-FAC surrogate without establishing the approximation accuracy required by the theory. This is particularly concerning because empirical Fisher matrices can differ substantially from true Hessians, and such approximation errors can invalidate the claimed fourth-order scaling.

  3. The fourth-order optimality-gap bound is an upper bound, not a guarantee that CurvTV achieves a smaller loss than vanilla task-vector transfer. Counterexamples exist even for smooth convex objectives with exact Hessians. Consequently, the theoretical results provide a useful local error characterization but do not fully justify the broader interpretation that curvature compensation necessarily improves transfer.

  4. The identity-matrix ablation achieves performance close to the full method, with average transfer gains of 0.448 versus 0.465. Since the identity variant effectively reduces to a simple affine combination of model checkpoints, the experiments do not convincingly isolate the contribution of curvature information. Stronger comparisons against equally tuned checkpoint-mixing methods and direct-gradient, second-order correction baselines are needed.

  5. Although K-FAC avoids constructing the full Hessian, curvature estimation and the associated matrix operations introduce substantial computational and memory overhead. The reported RoBERTa transfer time increases from approximately 1.25 seconds for vanilla transfer to 59.71 seconds for CurvTV. Moreover, the paper does not provide corresponding runtime and memory measurements for its 8B-model experiments. Given the modest incremental improvements over the identity ablation, the practical cost–benefit trade-off remains unconvincing. The evaluation of native capability preservation is also limited to masked-token prediction accuracy rather than the target models' actual downstream capabilities.

Overall, the theoretical perspective is interesting, but the restrictive assumptions, substantial theory–practice gap, insufficient isolation of curvature-specific benefits, and unresolved computational trade-offs significantly weaken the paper's broader claims and practical significance.

Questions for authors
  1. The analysis relies on exact proximal stationarity, local positive-definiteness, and Hessian Lipschitz continuity. Can the authors provide theoretical or empirical evidence that these assumptions approximately hold for practical fine-tuning procedures, particularly for large language models? How sensitive are the conclusions to violations of these assumptions?

  2. The theoretical guarantees concern a proximal-regularized objective, whereas the experiments evaluate unregularized task losses and downstream performance. What justifies transferring the theoretical conclusions to these empirical metrics? Can the authors establish a connection or provide additional experiments that directly validate the theoretical guarantees?

  3. CurvTV replaces the true Hessian with an empirical-Fisher/K-FAC approximation, but the robustness guarantee requires sufficiently small Hessian estimation error. Can the authors quantify this approximation error or demonstrate that the required conditions hold in practice? How robust is CurvTV when the curvature estimate is inaccurate?

  4. The fourth-order upper bound does not establish that CurvTV always outperforms vanilla task-vector transfer. Can the authors provide sufficient conditions for actual improvement? Furthermore, given the relatively small performance difference between CurvTV and the identity-matrix ablation (0.465 vs. 0.448 average gain), can the authors compare against equally tuned checkpoint-mixing and direct-gradient second-order baselines to isolate the contribution of curvature?

  5. CurvTV introduces substantial computational overhead compared with vanilla transfer. Given the approximately 48-fold increase in transfer time reported for RoBERTa, can the authors provide end-to-end runtime and peak-memory measurements for the 8B models? More importantly, do the improvements justify this additional cost compared with simpler transfer methods, particularly when curvature estimation must be repeated across tasks or models?

3 comments

Q1: Validity of the assumptions

Thank you for raising this point. We would like to clarify that these assumptions are standard in theoretical work on task arithmetic. For example, Daheim et al. [1] also model fine-tuning as a proximal objective with stationary solutions, and Tao et al. [2] and Zhou et al. [3] also rely on smoothness and local Taylor expansions. Moreover, prior work suggests that these assumptions approximately hold in practical fine-tuning. For example, early-stopped gradient descent approximates ridge regularization centered at the initialization [4], the Hessian late in training has few and small negative eigenvalues [5, 6], and language-model fine-tuning can stay in a nearly linearized regime where the curvature changes slowly [7].

Q2: Sensitivity to assumption violations

Thank you for this question. Our experiments are themselves conducted in settings where these assumptions do not hold exactly. On RoBERTa, θ1⋆\theta_1^\star and θ2⋆\theta_2^\star are fine-tuned with the standard procedure of Yu et al. [8], which uses weight decay, has no proximal term and trains for a fixed number of epochs, so they are not exact stationary points. In this setting, the relation predicted by the theory still holds (rˉ=0.959\bar r=0.959), and CurvTV raises the average gain from 0.353 to 0.465. On the theory side, App. B.4 relaxes exact stationarity and shows that the residuals enter the transfer condition only through their difference r1−r2r_1-r_2.

Q3: From the proximal objective to empirical metrics

Thank you for this question. We would like to clarify that the proximal objective is a deliberate modeling choice, shared by related work such as Daheim et al. [1]. The proximal term models the overall effect of standard fine-tuning in practice. Techniques such as early stopping, a limited number of training steps and small learning rates keep the solution close to its initialization, which approximately acts as proximal regularization centered at the initialization [4]. The proximal term keeps this key effect, namely that the fine-tuned solution depends on the initialization, while keeping the analysis simple enough to yield useful insights and practical methods. As noted in our response to Q2, we tested these insights on unregularized metrics. The results show that the proxy for the violation of the directional-curvature condition is highly correlated with the unregularized task-loss gap, and that CurvTV substantially improves the average gain over vanilla TV.

Q4: Hessian approximation error and robustness

Thank you for raising this point. We would like to clarify that only the fourth-order rate in Cor. 4.5 requires small Hessian estimation error, while the first bound holds for any fixed error Δ\Delta and grows smoothly with ∣Δ∣\\|\Delta\\|. Moreover, CurvTV strictly outperforms vanilla TV whenever the relative error of the curvature estimate is below α/(α+β)\sqrt{\alpha/(\alpha+\beta)} (see Q5). On the other hand, quantifying ∣Δ∣\\|\Delta\\| directly would require the exact Hessian, which is unfortunately out of reach for deep networks. For RoBERTa alone, the Hessian over its 85M encoder parameters has about 7×10157\times10^{15} entries, or about 29 PB in single precision. This is why Fisher-type curvature is commonly used as a substitute for the Hessian in model merging [1, 9].

Instead, we probe robustness through the number of curvature samples NN, since a smaller NN gives a less accurate K-FAC estimate. The table below reports averages over the 24 target–task pairs on RoBERTa.

N=50N=100N=500N=1000Vanilla TV
Gain ↑\uparrow0.3870.3910.4230.4650.353
Native gap ↓\downarrow0.1290.1390.1210.1460.151

As NN decreases, CurvTV degrades smoothly instead of breaking down, and even at N=50N=50 it still outperforms vanilla TV on both metrics.

Q5: Sufficient condition for improvement over vanilla TV

Thank you for this constructive question. A sufficient condition for CurvTV to strictly outperform vanilla TV follows directly from our existing results. Let u=θ2⋆−θ1⋆u=\theta_2^\star-\theta_1^\star, and let H~\tilde H be the curvature used by CurvTV. Comparing the CurvTV upper bound at θ~CurvTV\tilde\theta_{\mathrm{CurvTV}} with the vanilla lower bound given by the identity in Thm. 3.3 shows that, under the conditions of that theorem, if H~⪰0\tilde H\succeq0 and ∣(Hˉ−H~)u∣<α/(α+β)∣Hˉu∣,\\|(\bar H-\tilde H)u\\|<\sqrt{\alpha/(\alpha+\beta)}\\|\bar Hu\\|, then L2(θ~CurvTV)<L2(θtv)\mathcal L_2(\tilde\theta_{\mathrm{CurvTV}})<\mathcal L_2(\theta_{\mathrm{tv}}). That is, CurvTV strictly outperforms vanilla TV whenever its curvature estimate approximates the path-averaged curvature along uu with relative error below α/(α+β)\sqrt{\alpha/(\alpha+\beta)}. In the quadratic case with the exact Hessian, H~=Hˉ\tilde H=\bar H and CurvTV exactly recovers θ2⋆\theta_2^\star, so the condition holds whenever vanilla TV is not exact (Hˉu≠0\bar Hu\neq0). In the revision, we will state this condition.

Q6: Checkpoint-mixing and gradient-based baselines

Thank you for the suggestion. Following it, we add checkpoint mixing, i.e., linear interpolation (LI) λθ1⋆+(1−λ)θ2\lambda\theta_1^\star+(1-\lambda)\theta_2 over a range of λ\lambda. The table below compares it with the baselines in the paper, averaged over the 24 target–task pairs on RoBERTa.

θ1⋆\theta_1^\starLI (λ=0.2\lambda=0.2)LI (λ=0.4\lambda=0.4)LI (λ=0.6\lambda=0.6)LI (λ=0.8\lambda=0.8)Vanilla TVRe-BasinGradFixCurvTV (Identity)CurvTV (K-FAC)
Gain ↑\uparrow0.5250.0340.1710.3830.5010.3530.3520.3180.4480.465
Native gap ↓\downarrow0.3950.0000.0250.1010.2520.1510.1510.1710.1600.146

Transfer cannot be judged by gain alone, since otherwise directly using θ1⋆\theta_1^\star would be best, even though it loses most of the native capability. Considering both metrics, LI has a lower gain than CurvTV at λ=0.6\lambda=0.6 and a much larger native gap at λ=0.8\lambda=0.8, and vanilla TV, Re-Basin and GradFix are all worse than CurvTV on both metrics. Compared with CurvTV (Identity), which differs only in the curvature, K-FAC achieves both a higher gain and a lower native gap. The contribution of curvature is even clearer on Llama-3.1, where, for the math task vector transferred to Llama-3.1-abliterated, Identity has a much larger native gap than K-FAC (0.093 vs. 0.038) on the Korean benchmark KMMLU [10], which tests Korean knowledge across 45 subjects, mostly unrelated to the transferred math task. This shows that curvature lets CurvTV transfer more task knowledge while preserving more native capability. In the revision, we will include these new results.

Q7: Cost on 8B models

Thank you for raising this practical concern. The 48-fold ratio arises because vanilla TV requires only weight arithmetic and takes about one second. In absolute terms, CurvTV remains much cheaper than directly fine-tuning the target model. On RoBERTa, CurvTV takes 59.7s per transfer including the K-FAC statistics, whereas fine-tuning takes 43.1 min per target model. On Llama-3.1-8B, collecting the K-FAC statistics takes 64.5 min on two A100 GPUs with a peak memory of 79 GiB per GPU, and the transfer itself takes 30.2s, so CurvTV takes about 65 min end to end (about 2.2 GPU-hours). This is comparable to GradFix (about 64 min) and far below fine-tuning on our three tasks, which takes 17–154 A100 GPU-hours [11].

Moreover, the cost of CurvTV can be amortized, which reduces it further. The K-FAC statistics are computed at θ1⋆\theta_1^\star and depend only on the source task, so they are computed once and reused for every target model, including models released later. In contrast, both fine-tuning and GradFix must be rerun for every new target model, because they rely on training or gradients of that model. CurvTV therefore strikes a good balance: its cost is modest, and its gains are substantial where vanilla TV fails, raising the average gain on CodeBERT from 0.137 to 0.381 and turning the negative transfer of −0.17 on Llama-3.1-abliterated into +0.09.

Q8: Evaluation of native capability

Thank you for this comment. We would like to clarify that our RoBERTa targets are masked language models (MLMs), i.e., encoders continually pretrained on their own domains with masked-token prediction. Therefore, masked-token accuracy on in-domain data (TweetEval, PubMedQA and CodeSearchNet) directly measures their native capability.

References

[1] Daheim et al. Model Merging by Uncertainty-Based Gradient Matching. ICLR 2024.

[2] Tao et al. Task Arithmetic Through The Lens Of One-Shot Federated Learning. TMLR 2025.

[3] Zhou et al. On Task Vectors and Gradients. arXiv 2025.

[4] Ali et al. A Continuous-Time View of Early Stopping for Least Squares. AISTATS 2019.

[5] Sagun et al. Empirical Analysis of the Hessian of Over-Parametrized Neural Networks. arXiv 2017.

[6] Ghorbani et al. An Investigation into Neural Net Optimization via Hessian Eigenvalue Density. ICML 2019.

[7] Malladi et al. A Kernel-Based View of Language Model Fine-Tuning. ICML 2023.

[8] Yu et al. Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch. ICML 2024.

[9] Matena et al. Merging Models with Fisher-Weighted Averaging. NeurIPS 2022.

[10] Son et al. KMMLU: Measuring Massive Multitask Language Understanding in Korean. NAACL 2025.

[11] He et al. MergeBench: A Benchmark for Merging Domain-Specialized LLMs. NeurIPS 2025.

Dear Reviewer CsUQ,

Thank you again for your detailed review. As the discussion period is coming to a close, we would greatly appreciate it if you could let us know whether our responses have addressed your concerns. In particular, we provided a sufficient condition under which CurvTV strictly outperforms vanilla TV, compared CurvTV with linear interpolation and its Identity variant to isolate the contribution of curvature, and reported end-to-end runtime and peak memory on the 8B models. We truly value your feedback and thank you for your time.

Best regards,

The Authors

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

Thank you for your time and constructive feedback. Our paper explains through directional curvature why a task vector can fail to transfer between pretrained models, and derives from this explanation a curvature-compensated transfer method, CurvTV, that improves over vanilla task-vector transfer (vanilla TV). During the discussion, we conducted several new experiments following the reviewers' suggestions. To help you quickly grasp the discussion, we summarize the new evidence below by the concerns shared by multiple reviewers.

Proximal modeling and practical fine-tuning (Reviewers CsUQ, b7oa, K3vn). The proximal term is a modeling choice for practical fine-tuning, shared by related work (e.g., Daheim et al., ICLR 2024). In practice, early stopping, limited training steps and small learning rates keep the fine-tuned solution close to, and thus dependent on, its initialization, which is the effect the proximal term captures. Following Reviewer b7oa's suggestion, we compared proximal and standard fine-tuning on CoLA and SST-2. The correlation between the directional-curvature violation proxy and the transfer loss gap is strong under both protocols, and barely changes when the loss gap includes the proximal term. Moreover, although our theory is built on proximal fine-tuning, the resulting CurvTV remains effective under the standard fine-tuning used in all our experiments, raising the average gain on RoBERTa from 0.353 to 0.465 (a 32% relative increase).

Role of curvature (Reviewers CsUQ, b7oa, ADyy). We swept the coefficient of vanilla TV and of linear interpolation between the fine-tuned source model and the target model. Neither method matches CurvTV in both gain and native gap at any of the coefficients, and Reviewer ADyy accordingly withdrew the statement that scaling the task vector explains the gains.

CurvTV (Identity), which replaces the K-FAC curvature with the identity matrix, is itself a special case derived from our theory, which explains why it is competitive on RoBERTa. Even so, with the same coefficients, CurvTV achieves both a higher average gain and a smaller average native gap than CurvTV (Identity). Curvature matters more in harder transfers. For example, on Llama-3.1-abliterated, where vanilla TV leads to negative transfer, CurvTV raises the MATH-500 accuracy of CurvTV (Identity) from 33.0% to 36.6% (an 11% relative increase), and its KMMLU accuracy, which measures native capability, from 26.0% to 31.5% (a 21% relative increase).

Theoretical guarantees (Reviewers CsUQ, ADyy, K3vn). For small displacements θ2−θ1\theta_2-\theta_1 between the pretrained models and with exact curvature, CurvTV has an O(∣θ2−θ1∣4)O(\\|\theta_2-\theta_1\\|^4) optimality-gap bound, compared with O(∣θ2−θ1∣2)O(\\|\theta_2-\theta_1\\|^2) for vanilla TV. Moreover, comparing the CurvTV upper bound with the vanilla TV lower bound from our existing theorems proves that CurvTV strictly outperforms vanilla TV whenever its curvature estimate is sufficiently accurate along θ2⋆−θ1⋆\theta_2^\star-\theta_1^\star, regardless of how far apart the pretrained models are.

Cost and practicality (Reviewers CsUQ, ADyy). CurvTV does introduce additional computation and storage, but this cost is incurred only once. The K-FAC statistics depend only on the source task, so every target model, including models released later, can reuse them directly without any data or training. In contrast, direct fine-tuning and GradFix must be rerun for every new target model. Following Reviewer ADyy's suggestion, we also fine-tuned the RoBERTa target models directly on the same 1000 samples. CurvTV passes over these samples once for all target models, whereas direct fine-tuning takes 10 epochs per target model, yet CurvTV achieves a higher average gain (0.465 vs. 0.392). CurvTV thus trades a one-time cost for better transfer than the baselines.

All new experiments and clarifications will be incorporated into the revised version. We are sincerely grateful for the reviewers' feedback, which has substantially improved the paper, and we hope this summary is helpful for your evaluation. We would be glad to clarify any point or provide further details during the remaining discussion period.

Kind regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.