acceptodds
Under review as a conference paper at ICLR 2027

Optimizing over Mode Connectivity for Multi-Task Learning and Machine Unlearning

Abstract

Mode connectivity refers to the phenomenon that independently trained neural networks can be connected by continuous low-loss paths in the parameter space, revealing a family of high-quality candidate models rather than isolated optima. While existing theoretical studies have focused exclusively on the single-task learning setting, the role of mode connectivity in more challenging learning paradigms remains largely unexplored. In this paper, we provide the first theoretical study of mode connectivity-based multi-task learning and machine unlearning. We consider overparameterized fully connected multilayer perceptrons (MLPs) trained from a shared initialization in the Neural Tangent Kernel (NTK) regime, which yields closed-form expressions for the loss along a path up to a bounded error. We then extend these results to finite-width nonlinear networks. We establish theoretical conditions under which quadratic Bézier curves achieve lower loss than linear interpolation, demonstrating the advantage of optimizing over nonlinear connectivity manifolds when balancing competing learning objectives. Finally, we conduct experiments on both NTK-linearized models and standard finite-width neural networks, and the results closely validate our theoretical predictions.

Primary area:
interpretability and explainable AI
Keywords:
Mode Connectivity, Multi-Task Learning, Machine Unlearning, Neural Network

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_1QDgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper develops a theoretical framework showing that Bézier curves between two trained models can provide better multi-task learning and unlearning performance compared to linear interpolation.

Critical strengths

A theoretical explanation is indeed what the current mode-connectivity (between independently trained models) literature needs. Any account that goes beyond generic overparameterization arguments would make a valuable contribution to the mode-connectivity field.

Critical weaknesses

(1) The scope of this paper is "shared-initialization cross-task near-linear mode connectivity" rather than "mode connectivity across independently trained models".

The author claims that "mode connectivity refers to the phenomenon that independently trained neural networks..." (in the abstract). However, their theoretical explanation is built upon "models are trained from a common random initialization". Mathematically, if two models are independently obtained, then the layer-wise cosine similarity of each layer's weights should be zero. However, if you start training from the same initialization, layer-wise cosine similarity is not zero [2].

Empirical results are also obtained with an identical pretrained model (pretrained ViT-tiny). This is not mode connectivity, either. This falls within the scope of linear mode connectivity, see [1]: "linear mode connectivity has been consistently observed in model spawning", "model spawning: A network is randomly initialized, trained for a small number of epochs, and then spawned into two copies which continue to be independently trained using different SGD randomnesses"

(2) The authors' results are trivial in the Linear Mode Connectivity (LMC) context.

LMC works already show that linear interpolation has low-loss values between two permuted/spawned modes [1]. Given this context, the authors further show that the Bezier curve is better than linear interpolation. One easy assumption is that the Bezier curve bends the linear interpolation a little bit towards the "low-loss" direction.

Figure 1, the midpoint of the Bezier curve is obtained with training on the full dataset (union of two subsets, Line 404), while the left and right terminal points are obtained with a partial dataset. This naturally causes "Bezier paths substantially outperform direct linear interpolation across all partitions" (Line 391) because the Bezier path is trained on the full dataset.

[1] https://arxiv.org/abs/2307.08286

[2] https://arxiv.org/abs/1912.05671

Questions for authors

I am not sure I understand Equation 11. Is this a standard approach to unlearning? The objective L1(θ)−L2(θ)L_1(\theta) - L_2(\theta) is unbounded and can be minimized by driving L2(θ)L_2(\theta) towards ∞\infty.

8 comments
AuthorsOpenReview ↗

Response to Reviewer's comments (Questions Part)

We thank the reviewer for the detailed review, in particular for identifying the need to state the theoretical scope more precisely and to clarify the optimization objective. We have uploaded a revised version of the manuscript which incorporates reviewers' comments.

We shall first directly answer the reviewer's questions. Due to limited space, we shall address the weaknesses pointed out by the reviewer in a later comment.

  1. Explanation of Equation 11.

Answer: We thank the reviewer for raising this concern. The loss in Equation 11 can indeed be unbounded below without regularization. In our original analysis, the Bézier control point is obtained by minimizing the regularized\textit{regularized} cumulative unlearning loss specified in Section 4.1. This objective includes the penalty ∥θ−θ(0)∥22\|\theta-\theta^{(0)}\|_2^2, and we explicitly assume that the resulting objective is strictly convex.

In the NTK\mathrm{NTK}-linearized model, the Hessian of this control objective is

I+\frac{2}{15} \left( \lambda\Phi_1^\top\Phi_1-\Phi_2^\top\Phi_2 \right) \right],$$ where $\Phi_1$ and $\Phi_2$ are the retain-task and forget-task Jacobians at the common initialization. The factor $2/15$ follows from integrating the squared coefficient of the control point: $\int_0^1[2\alpha(1-\alpha)]^2\,d\alpha=2/15$. The stated strict-convexity assumption is therefore equivalent to requiring the matrix in brackets to be positive definite. Under this condition, the regularized control objective is bounded below and has a unique finite minimizer. We recognize that the connection between Equation 11 and the regularized objective used to obtain the control point could be clearer. We will clarify the role of regularization when introducing the unlearning loss and state the positive-definiteness condition explicitly in Section 4.1.
AuthorsOpenReview ↗

Response to Reviewer's comments (Weaknesss Part)

Next we address the weaknesses pointed out by the reviewer.

  1. The scope of this paper is "shared-initialization cross-task near-linear mode connectivity" rather than "mode connectivity across independently trained models".

Answer: We thank the reviewer for highlighting this distinction. Our theoretical results concern models trained on different tasks from a common initialization under the NTK\mathrm{NTK} approximation. The ViT experiments likewise use task-specific models fine-tuned from a shared pretrained checkpoint. We agree that the phrase "independently trained'' can obscure this scope, and we will revise the abstract and introduction to describe the shared initialization explicitly.

This setting is closely related to the shared-initialization and model-spawning settings discussed in [1,2]. In particular, [1] describes linear mode connectivity as a form of mode connectivity, and [2] studies it by training copies of a common model with different SGD randomness. The connectivity properties we study concern the loss along the connecting path and does not require the endpoint weights to have zero layer-wise cosine similarity. Shared initialization is naturally available when a common pretrained model is fine-tuned for different tasks, making this setting directly relevant to combining task-specific models.

Within this setting, our contribution is an explicit analysis of linear and optimized quadratic Bézier paths under different learning objectives. For example, in multi-task learning, θ1\theta_1 and θ2\theta_2 fit T1\mathcal T_1 and T2\mathcal T_2, respectively, but can have nonzero losses on the other task. Shared initialization therefore does not by itself establish that their interpolation performs well on both tasks. Theorem 3.1 and Corollary 3.2 characterize the pointwise and cumulative multi-task losses in terms of these cross-task residuals, identifying when and by how much optimizing the Bézier control improves upon linear interpolation. Our unlearning analysis similarly identifies conditions under which models on the connecting paths more closely match the retain-only reference in terms of the task losses.

These results complement the same-task linear connectivity studied in [1,2] by quantifying how optimized nonlinear paths balance multiple or competing objectives. This provides a theoretical basis for using path optimization in multi-task learning and unlearning within the shared-initialization setting. We will emphasize this contribution while making the scope of our assumptions clearer.

  1. The results are trivial in the Linear Mode Connectivity (LMC) context.

Answer: We thank the reviewer for raising this point. The same-task LMC results in [1] establish low-loss connections between solutions evaluated on a common task. In our multi-task setting, each endpoint fits its own task but can have nonzero loss on the other task. The question is therefore how the connecting path balances these different objectives, and how optimizing the path changes its multi-task performance.

We agree that the inclusion of the linear path in the quadratic family implies a weak cumulative-loss inequality under exact optimization. Our contribution is the quantitative characterization of this improvement. In the NTK\mathrm{NTK}-linearized model, Corollary 3.2 gives the exact relation

=\frac{3}{8} \int_0^1 \ell_+^{\mathrm{NTK}}(\theta_I(\alpha)) d\alpha,$$ with strict improvement whenever at least one cross-task residual is nonzero. Theorem 3.1 further establishes that the optimized Bézier path has no greater multi-task loss at every $\alpha\in[1/5,4/5]$. These results specify both the magnitude of the cumulative improvement and an entire interval of pointwise improvement, neither of which follows from the inclusion of the linear path alone. For finite-width networks, the comparison additionally requires the predicted improvement to exceed the stated linearization errors. Regarding Figure 1, we indeed optimize the Bézier control point using the union of the two task datasets while keeping the endpoints fixed. This reuses the original training data and requires additional data access and optimization compared with constructing the linear path. We agree that this additional training contributes to the observed improvement. Figure 1 demonstrates the performance of the resulting family of models, including improvements over broad portions of the path, but does not by itself separate the benefit of path optimization from that of further training on the joint dataset. [1] https://arxiv.org/abs/2307.08286 [2] https://arxiv.org/abs/1912.05671
  1. Mode Connectivity vs Linear Mode Connectivity

I believe we have reached agreement that the paper is more closely related to the Linear Mode Connectivity (LMC) literature than to the Mode Connectivity literature. I would appreciate it if the authors could revise the abstract and, if feasible, the title, since “mode connectivity” currently appears in the title. However, changing the title may be difficult at this stage.

There is one point I would like to clarify. You state:

"the connectivity properties we study concern the loss along the connecting path and do not require the endpoint weights to have zero layer-wise cosine similarity."

Connecting models whose corresponding layers have zero cosine similarity is generally more difficult. When the cosine similarity is nonzero (between 0 and 1), the problem is easier. This is also why model permutation is required in LMC. Therefore, the statement that your method “does not require the endpoint weights to have zero layer-wise cosine similarity” is somewhat imprecise.

  1. Thank you for the clarification. I agree that your theoretical contribution provides new insights into model merging. However, this contribution is not sufficiently reflected or demonstrated in the empirical evaluation.
Reviewer_8zSWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Suppose we have two models that are trained on two different tasks (starting from the same initialization). The paper studies mode connectivity: what is the loss on both tasks for models whose parameters are on the line connecting the parameters of the two original models? As a further variant of this, we can also not just move along the line, but we can optimize a quadratic Bezier curve to interpolate between the models. The paper gives closed form expressions for the multi-task loss in a specific setting (L-layer fully connected ReLU networks, NTK regime) that become exact as the width tends to infinity. The Bezier curve performs better than the line. The paper also uses these ideas in a machine unlearning setting.

Critical strengths

The result is quite concrete and exact. An explicit rigorous derivation. The machine unlearning application is a nice addition.

Critical weaknesses

The differences between the line and the Bezier curve (and the fact that the optimized Bezier curve beats the line) is unsurprising. The line is a special case of the Bezier curve. To determine the control points of the Bezier curve also appears to need data and so can incorporate more information. I take the point that there isn't just a single point on the curve that is good (indeed, if a model exists that achieves zero loss on both tasks simultaneously, we could presumably make the curve go through that point), but that large parts of the curve work well. Nevertheless, the advantage over the line is a given. But maybe that is not meant to be the main message of the paper in any case.

Questions for authors

What happens if you compare against a single model that is trained on both tasks?

1 comment
AuthorsOpenReview ↗

Response to Reviewer's comments

We thank the reviewer for the positive assessment of the explicit derivation and for clarifying which aspect of the contribution needs stronger emphasis. We have uploaded a revised version of the manuscript which incorporates reviewers' comments.

We shall first directly answer the reviewer's question.

  1. What happens if you compare against a single model that is trained on both tasks?

Answer: We thank the reviewer for suggesting this important comparison. Under our full-row-rank assumption on the stacked task Jacobians, there exists a joint model θ1,2\theta_{1,2} that interpolates both tasks in the NTK-linearized model and therefore satisfies ℓ+NTK(θ1,2)=0\ell_+^{\mathrm{NTK}}(\theta_{1,2})=0. We do not claim that the optimized Bézier curve achieves a lower training loss than this optimal joint model. Our contribution is to characterize a family of models connecting fixed task-specific endpoints: Theorem 3.1 describes its pointwise loss and favorable interpolation interval, while Corollary 3.2 quantifies its cumulative-loss improvement over linear interpolation.

Next we shall address the weaknesses pointed out by the reviewer.

  1. What's the point of including the linear path as a special case?

Answer: We agree that the weak cumulative inequality follows immediately from the fact that setting the control to (θ1+θ2)/2(\theta_1+\theta_2)/2 recovers linear interpolation. Our intended contribution is the quantitative description of the optimized path. In the linearized model, the cumulative loss decreases from

\quad\text{to}\quad \frac18\left(\|R_{12}^{\mathrm{NTK}}\|^2+\|R_{21}^{\mathrm{NTK}}\|^2\right),$$ with equality between the two cumulative losses only when both cross-task residuals vanish. We also obtain a pointwise comparison throughout $[1/5,4/5]$. These constants, the equality characterization, and the pointwise interval are valuable results that do not follow merely from inclusion of the linear path in the quadratic family. For finite-width networks, the comparison requires the linearized improvement to exceed the approximation errors. 2. Data use and the purpose of control optimization. Answer: We agree that learning the Bézier control requires access to the original task datasets and additional optimization compared with constructing the linear path from endpoint weights. We will make these requirements explicit. Our contribution is to characterize how this data-dependent optimization changes the loss throughout the path. In particular, the control is optimized to minimize the average loss along the entire curve, rather than its own standalone loss. The reviewer's observation that a curve can be made to pass through a jointly optimal model is correct. However, this does not by itself minimize the cumulative loss. For example, in the NTK-linearized model, let $\theta_{1,2}$ interpolate both tasks. Choosing the control as $2\theta_{1,2}-(\theta_1+\theta_2)/2$ makes the curve pass through $\theta_{1,2}$ at $\alpha=1/2$. The resulting cumulative multi-task loss is $$\frac{2}{15} \left( \|R_{12}^{\mathrm{NTK}}\|^2 +\|R_{21}^{\mathrm{NTK}}\|^2 \right),$$ whereas the optimized curve in Corollary 3.2 achieves the smaller coefficient $1/8$. This comparison illustrates the distinction between ensuring that one point on the curve is jointly optimal and optimizing the loss across the entire family of models. Our analysis quantifies the latter objective and identifies the regions where the resulting curve improves over linear interpolation.
Reviewer_o11JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the mode connectivity-based multi-task learning and machine unlearning. They consider the problem on MLPs under the Neural Tangent Kernel regime. Furthermore, they theoretically show that the quadratic Bezier curves achieve lower loss than linear interpolation. They then conduct experiments to verify their findings.

Critical strengths
  1. In task arithmetic and machine unlearning, the linear interpolation is widely used and analyzed. The use of quadratic Bezier curve is new to me. The analysis and proposed method provide a novel insight for this area.
  2. The proposed method is inspired by theoretical analysis, which provide strong evidence. The experiments also verify the great performance when using the quadratic Bezier curve.
Critical weaknesses
  1. The main limitation of the proposed method is that the analysis and the method can only be applied in two-task scenarios. The generalization to multi-task settings (T > 2) is not considered in this paper.
  2. Even in the two-task setting, optimizing the quadratic Bézier curve requires additional data and a trainable control point compared with linear interpolation. The improvement can also be obtained with the additional training. Reporting the performance of an MTL model is also necessary.
  3. The novelty of this paper seems to be overstated. The authors claim that they provide the first theoretical study of mode connectivity-based MTL and unlearning. However, the NTK analyses of multi-task combination and model unlearning have already been proposed in [1,2]. The authors should include a discussion about these related works.

Reference
[1] Guillermo Ortiz-Jiménez, Alessandro Favero, Pascal Frossard. Task Arithmetic in the Tangent Space: Improved Editing of Pre-Trained Models. NeurIPS 2023.
[2] Kotaro Yoshida, Yuji Naraki, Takafumi Horie, Ryosuke Yamaki, Ryotaro Shimizu, Yuki Saito, Julian J. McAuley, Hiroki Naganuma. Mastering Task Arithmetic: τJp as a Key Indicator for Weight Disentanglement. ICLR 2025.

Questions for authors
  1. Does the proposed novel method achieve better performance than multi-task and single-task models?
  2. How can the framework be generalized to the multi-task setting with T>2T > 2?
2 comments
AuthorsOpenReview ↗

Response to Reviewer's comments (Questions Part)

We thank the reviewer for recognizing the value of our theoretical analysis and for identifying important questions about our scope and related literature.

We start by directly answering the two questions raised by the reviewer. Due to limited space, we shall address the weaknesses pointed out by the reviewer in a later comment. We have also uploaded a revised version of the manuscript which incorporates reviewers' comments.

  1. Does the proposed novel method achieve better performance than multi-task and single-task models?

Answer: In the multi-task learning setting, Theorem 3.1 provides explicit expressions for the multi-task loss along the linear and Bézier paths, allowing a precise comparison with the single-task endpoints in the NTK-linearized model. In particular, evaluating the Bézier loss at α=2/5\alpha=2/5 and 3/53/5 gives

=\frac1{25}\|R_{21}^{\mathrm{NTK}}\|^2,\qquad \ell_+^{\mathrm{NTK}}\bigl(\theta_B(3/5)\bigr) =\frac1{25}\|R_{12}^{\mathrm{NTK}}\|^2.$$ The multi-task losses of the single-task models $\theta_2$ and $\theta_1$ are $\|R_{21}^{\mathrm{NTK}}\|^2$ and $\|R_{12}^{\mathrm{NTK}}\|^2$, respectively. Thus, if both cross-task residuals are nonzero, the best Bézier model achieves strictly smaller aggregate training loss than both single-task endpoints. For comparison with a jointly trained multi-task model, our full-row-rank assumption implies that a joint model $\theta_{1,2}$ can achieve $\ell_+^{\mathrm{NTK}}(\theta_{1,2})=0$. We therefore do not claim a lower training loss than the optimal joint model. Our results characterize a family of models connecting the fixed task-specific endpoints and quantify its improvement over linear interpolation. For machine unlearning, $\theta_1$ serves as the retain-only reference and satisfies $\ell_{\simeq 1}(\theta_1)=0$. Our objective is therefore to approach this reference. In Setting 1, Theorem 4.1 shows that the optimal Bézier model achieves a retain-task similarity loss no greater than that of the optimal linear model, while Corollary 4.2 characterizes when this improvement is strict. 2. How can the framework be generalized to the multi-task setting with $T>2$? Answer: Our paper focuses on two task-specific endpoints. The cumulative-loss analysis can be extended to $T$ tasks, $\mathcal T_1,\ldots,\mathcal T_T$, by connecting their $T$ endpoint models using a quadratic Bézier simplex. For nonnegative weights $t=(t_1,\ldots,t_T)$ satisfying $\sum_{j=1}^T t_j=1$, define $\theta_I(t)=\sum_{j=1}^{T} t_j\theta_j$ and $\theta_B(t)=\sum_{j=1}^{T} t_j^2\theta_j+2\sum_{1\le i < j\le T}t_it_j\theta_{ij}^{\ast}$. Each vertex corresponds to a task-specific model $\theta_j$, while $\{\theta_{ij}^*\}_{1\le i<j\le T}$ are the trainable control models. We extend the multi-task loss to \$ \ell_+^{\mathrm{NTK}}(\theta) :=\sum_{i=1}^T\ell_i^{\mathrm{NTK}}(\theta), \$ where $\ell_i^{\mathrm{NTK}}$ is the training loss on task $\mathcal T_i$ in the NTK-linearized model. We choose the control models to minimize $\mathbb E_t[\ell_+^{\mathrm{NTK}}(\theta_B(t))]$, where $\mathbb E_t$ denotes expectation with respect to the uniform probability distribution over the simplex of admissible weights. This generalizes the uniform averaging over $\alpha\in[0,1]$ used in the two-task analysis. Under the full-row-rank assumption on the Jacobian matrix obtained by stacking all $T$ task Jacobians, the optimal control residuals and cumulative losses admit closed-form expressions. In particular, for $T\ge3$, $$\mathbb E_t \left[ \ell_+^{\mathrm{NTK}}(\theta_B(t)) \right] \le \frac{4}{(T-1)(T+2)} \mathbb E_t\!\left[ \ell_+^{\mathrm{NTK}}(\theta_I(t)) \right].$$ Thus, the average multi-task loss over the optimized Bézier simplex is at most $4/[(T-1)(T+2)]$ times that over the linear simplex. Since this factor is strictly less than one for $T\ge3$, the Bézier construction achieves a strictly smaller average loss unless every endpoint already fits all tasks. For $T=2$, the exact cumulative-loss expressions recover the coefficients $1/3$ and $1/8$ in Corollary 3.2 under the original interpolation assumptions. As in Theorem 3.1, transferring the comparison to finite-width networks additionally requires the endpoints and controls to remain in the common NTK neighborhood and the predicted improvement to exceed the approximation errors. We have added this construction and cumulative comparison to Section 3, with the exact identities and proof in Appendix A.4. \end{enumerate}
AuthorsOpenReview ↗

Response to Reviewer's comments (Weaknesses Part)

Next we attempt to address the weaknesses raised by the reviewer.

  1. The main limitation of the proposed method is that the analysis and the method can only be applied in two-task scenarios. The generalization to multi-task settings (T>2T>2) is not considered in this paper.

Answer: Please refer to \textbf{A2} in the response.

  1. Even in the two-task setting, optimizing the quadratic Bézier curve requires additional data and a trainable control point compared with linear interpolation. The improvement can also be obtained with the additional training. Reporting the performance of an MTL model is also necessary.

Answer: We thank the reviewer for pointing out that optimizing the Bézier control point uses the task datasets and requires additional training. We would like to clarify that this optimization reuses the datasets used to train the endpoints and does not require a new source of samples. We will state the data-access and training requirements explicitly. Our analysis characterizes how optimizing an endpoint-preserving path changes the loss throughout that path.

The comparison with joint training can also be made analytically. Under our stacked full-row-rank assumption, there exists a joint model θ1,2\theta_{1,2} satisfying ℓ+NTK(θ1,2)=0\ell_+^{\mathrm{NTK}}(\theta_{1,2})=0. We therefore do not claim that the optimized curve outperforms the optimal joint model on the training objective. Theorem 3.1 and Corollary 3.2 characterize the pointwise and cumulative losses of a family connecting the fixed task-specific endpoints, including the interval where the optimized Bézier curve improves over linear interpolation.

We agree that a comparison with direct joint training would help assess the practical value of the additional optimization. These baselines are not included in our current experiments, which compare the linear and optimized B'ezier paths. The observed improvements therefore do not isolate the benefit of path optimization from that of further training on the union of the task datasets.

The main contribution of our work is to establish a quantitative theoretical foundation for optimizing connectivity between fixed task-specific models. While a matched joint-training comparison would provide additional evidence about relative test performance and training cost, our current experiments support the comparison with linear interpolation without establishing superiority over joint training. 3. The novelty of this paper seems to be overstated. The authors claim that they provide the first theoretical study of mode connectivity-based MTL and unlearning. However, the NTK analyses of multi-task combination and model unlearning have already been proposed in [1,2]. The authors should include a discussion about these related works.

Answer: We thank the reviewer for identifying these closely related works. [1] characterize weight disentanglement, connect task arithmetic to localized NTK eigenfunctions, and study linearized fine-tuning. [2] analyze the τ\tau-Jacobian product and use it to reduce interference in task addition and negation. Both works provide important theoretical insights into multi-task learning and task negation using tangent-space analysis. However, our work addresses a fundamentally different research question and provides several theoretical results that are not covered by these studies.

First, the research objective of our work differs from prior works. [1, 2] focuses on the understanding of the role of weight disentanglement and task interference when adding or negating task vectors. In contrast, our work focuses on characterizing the pointwise and cumulative losses along optimized paths between fixed endpoint models, which is different from analyzing the effectiveness of task arithmetic alone.

Second, our work quantifies the effect of mode connectivity for different paths with a theoretical comparison. For both multi-task learning and unlearning, our work characterizes the losses for both linear interpolation and quadratic B'ezier curves. Our result quantifies not only whether nonlinear connectivity is beneficial, but also where and by how much it improves upon linear interpolation. Such path-wise performance characterizations are not provided by the existing task arithmetic analyses.

To clarify our novelty and avoid any misunderstanding, we would like to restate our work as \textbf{the first to provide a theoretical analysis of linear and optimized quadratic B'ezier paths for multi-task learning and machine unlearning under the NTK framework, deriving explicit conditions under which nonlinear connectivity improves upon linear interpolation}. We will add both references, clarify the relationship between cross-task Jacobian products and our target-dependent residuals, and revise our novelty statement to emphasize this specific contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.