New Breakthrough in Model Merging: Interference Bounds and Curvature-Guided Training

A new theory formalizes model composability, characterizes merging loss with an interference functional, and proposes curvature-guided regularization.
This study formalizes the "composability" of merging multiple fine-tuned model parameters as a property of a set of task vectors, proving that the excess loss of the merged model is controlled by an interference functional that measures cross-task curvature overlap and signed alignment of updates. In the quadratic case, the interference functional is exactly equal to the excess loss, providing a necessary and sufficient criterion based on curvature rather than Euclidean distance, explaining why models that appear similar in Euclidean space can suffer a sharp performance drop after merging. Based on this, the authors propose a curvature-guided regularizer that uses curvature statistics maintained by the optimizer to constrain parameter displacement, bounding the interference functional inversely proportional to the penalty coefficient under the Loewner comparison assumption while preserving the O(1/√T) convergence rate. Experiments validate the theoretical bounds, but find that on Transformers the rank correlation between the interference functional and merging degradation does not exceed that of the Euclidean displacement norm; under strong penalties the first-order term dominates, suggesting a trade-off between curvature control and first-order bias.
New Breakthrough in Model Merging: Interference Bounds and Curvature-Guided Training
Now that fine-tuning has become a standard practice in AI applications, efficiently merging multiple independently fine-tuned models remains a central concern for both industry and academia. A recent theoretical study offers a new theoretical tool for this practical problem: it formalizes "composability" as a property of a set of task vectors and derives that the excess loss of a merged model is controlled by an interference functional. This functional measures cross-task curvature overlap and the signed alignment of update directions.

Background: Why Does Parameter Merging Work, and When Does It Fail?
Merging the parameter updates of multiple fine-tuned models (i.e., task vectors) has become an alternative to multi-task training. Its appeal lies in the fact that independent fine-tuning alone can compose multi-task capabilities at nearly zero cost. However, merging does not always work—interference between task vectors can degrade performance. Previously, practitioners mainly relied on empirical rules (such as linear mode connectivity) to judge mergeability, but rigorous theoretical bounds were lacking.
This study formalizes "composability" as a property of a set of task vectors and provides a mathematical characterization of the excess loss of the merged model. In the quadratic case, the interference functional is exactly equal to the excess loss, thereby providing a necessary and sufficient criterion: it measures the angle between task vectors under the curvature metric rather than the Euclidean metric. This means that looking only at the Euclidean distance of parameter updates can be misleading; the local geometry of the loss function must be taken into account.
Task vector is a core concept in model merging: after independently fine-tuning for each task, the difference vector obtained by subtracting the pretrained model parameters from the fine-tuned model parameters is (\tau_i = \theta_i - \theta_0). During merging, these task vectors are typically added with weights and then added back to the base model, for example (\theta_{\text{merged}} = \theta_0 + \sum_i \lambda_i \tau_i). The intuition behind this approach is that the updates brought by fine-tuning mainly encode task-specific knowledge; if the task vectors are "orthogonal" to each other or aligned in direction, then adding them together can preserve multiple capabilities simultaneously. Conversely, if the update directions conflict, adding them will cancel each other out. The task vector method is popular in industry because it decouples multi-task training into multiple independent single-task fine-tuning runs, avoiding the data mixing and computational scheduling challenges of joint training, and can compose multi-task models at extremely low cost.
Interference Functional: The Core Metric for Measuring Merging Loss
The interference functional consists of two parts: cross-task curvature overlap and signed alignment of updates. Intuitively, if two task vectors are aligned in direction and reinforce each other in the curvature-weighted sense, then interference is small; conversely, if their directions conflict or curvature overlap causes mutual cancellation, then interference is large. In the quadratic case, this functional is exactly equal to the excess loss, which provides a clear objective for subsequent training strategies.
One detail: this criterion differs from angles in Euclidean space. Two vectors with a very small angle in Euclidean space may have a large angle under the curvature metric, leading to merging failure. This explains why some seemingly "similar" fine-tuned models suffer a sharp performance drop after merging.
Curvature here refers to the local second-order geometry of the loss function in parameter space, usually characterized by the Hessian matrix or its approximations (such as the Fisher information matrix or the second-moment estimates maintained by the optimizer). If we imagine the loss surface as an undulating terrain, curvature describes the rate at which the slope changes in different directions: in directions with large curvature, a small movement of parameters causes drastic changes in loss; in directions with small curvature, the surface is relatively flat. The angle between two task vectors under the "curvature metric" is equivalent to first weighting and stretching the directions with the Hessian and then computing the angle, so it can reflect whether the updates fall in directions to which each other is sensitive. Loewner comparison is a positive semidefinite ordering of matrices: if matrix (A \preceq B), it means (B-A) is positive semidefinite, i.e., for any vector (v), (v^\top A v \le v^\top B v). The paper uses a single Loewner comparison to relate the base curvature statistics maintained by the optimizer to the true task curvature, thereby transforming the displacement constraint during training into an upper bound on the interference functional.
Curvature-Guided Regularization: A Training Strategy for Controlling Interference
Based on the above theory, the authors propose a curvature-guided regularizer. This regularizer uses the curvature statistics (such as second-moment estimates) maintained by memory-efficient optimizers (such as Adam) to constrain the parameter displacement at each iteration under the base curvature metric to not exceed an upper limit set by a penalty coefficient, while preserving the standard O(1/√T) nonconvex convergence rate.
Under the condition that a single Loewner comparison holds between the base statistics and the task curvature, this control can be transformed into an upper bound on the interference functional, and this upper bound is inversely proportional to the penalty coefficient. In other words, by adjusting the penalty coefficient, the training process can suppress the curvature term of the interference functional within a controllable range.
However, the regularized solution is not stationary with respect to the task loss, and therefore produces a first-order term. This first-order term cannot be reduced in the same way. Experiments show that under strong penalties, the first-order term dominates. This means that over-reliance on strong regularization may backfire, and a trade-off is needed between curvature control and the first-order term.
The nonconvex convergence rate (O(1/\sqrt{T})) is a standard result in stochastic optimization theory: under the conditions of a nonconvex objective and noisy gradient estimates, after (T) iterations, the approximate stationary point found by the algorithm (the expected squared gradient norm) decreases at a rate of (1/\sqrt{T}). The paper emphasizes that the regularizer preserves this rate, meaning that introducing curvature constraints does not break the optimizer's theoretical convergence guarantee, which is important for practical deployment. On the other hand, a stationary point refers to a point where the gradient is zero; the regularized solution optimizes a composite objective of "task loss + penalty term," so its gradient with respect to the original task loss is usually nonzero, and this residual gradient is the first-order term. Strong penalties push the solution farther from the optimum of the task loss, and the first-order term increases accordingly, resulting in a trade-off where "curvature interference is suppressed but first-order bias rises."
Experimental Validation: Theoretical Bounds and Training Dynamics
The authors conducted experiments on exact non-quadratic models, verified the existence of a budget and a ceiling, and observed the dominant role of the first-order term under strong penalties. At the same time, the curvature-guided penalty was slightly better than the isotropic penalty.
In a three-seed study based on Transformers, the interference functional was rank-correlated with merging degradation, but its strength was not greater than that of the Euclidean displacement norm. This result is somewhat unexpected: although the curvature metric is theoretically more precise, in actual Transformer models, the simple Euclidean displacement norm remains a competitive predictor. The authors honestly present this discrepancy, suggesting that there is still a gap between theoretical bounds and practical predictive metrics.
Practical Implications and Limitations
For engineers working on model merging, this study offers several insights:
- The interference functional can serve as a diagnostic tool for merging quality: although its predictive power on Transformers is comparable to the Euclidean norm, it provides theoretical guarantees, especially being necessary and sufficient under the quadratic approximation.
- Curvature-guided training helps control merging loss: by regularizing and constraining displacement, interference can be actively reduced during training, but attention must be paid to the side effect of the first-order term under strong penalties.
- The Euclidean norm remains a simple and effective baseline: in the absence of curvature information, monitoring the Euclidean displacement norm of task vectors is still a good heuristic.
The limitations of the study are: the theoretical guarantees are mainly established under the quadratic case and the Loewner comparison assumption, and the non-quadratic nature of actual models (such as Transformers) may weaken the bounds. In addition, the dominant role of the first-order term under strong penalties suggests that more refined regularization designs are needed.
Overall, this work provides a more rigorous theoretical foundation for model merging and connects training dynamics with post-merging interference. Although the practical predictive metric has not yet significantly surpassed the Euclidean norm, its analytical framework points the way for future research: how to design more effective curvature-aware regularization and how to tighten bounds in non-quadratic settings will be worthwhile questions to explore.
One-line Summary
A new theory formalizes model composability, characterizes merging loss with an interference functional, and proposes curvature-guided regularization.
Paragraph Summary
This study formalizes the "composability" of merging multiple fine-tuned model parameters as a property of a set of task vectors, proving that the excess loss of the merged model is controlled by an interference functional that measures cross-task curvature overlap and signed alignment of updates. In the quadratic case, the interference functional is exactly equal to the excess loss, providing a necessary and sufficient criterion, and its metric is based on curvature rather than Euclidean distance, explaining the phenomenon where models that appear similar in Euclidean space suffer a sharp performance drop after merging. Based on this, the authors propose a curvature-guided regularizer that uses curvature statistics maintained by the optimizer to constrain parameter displacement, and under the Loewner comparison assumption, bounds the interference functional inversely proportional to the penalty coefficient while preserving the O(1/√T) convergence rate. Experiments validate the theoretical bounds, but find that on Transformers the rank correlation between the interference functional and merging degradation does not exceed that of the Euclidean displacement norm; under strong penalties the first-order term dominates, suggesting a trade-off between curvature control and first-order bias.
Related articles

Anthropic's IPO: Is a $1.4 Trillion Valuation AI's Future or the Peak of a Bubble?
Anthropic is about to list on Nasdaq with a valuation of up to $1.4 trillion, potentially raising over $60 billion to become the second-largest IPO in history. This article breaks down its revenue, the truth behind its hundred-billion-dollar book loss, compute costs, control design, and lock-up pressure, and analyzes the IPO's chain effects on the AI sector.

Anthropic Races Toward Nasdaq: Ticker ANT, Can a Trillion-Dollar Valuation Hold?
Anthropic files its S1 for a Nasdaq IPO under ticker ANT, led by Goldman Sachs and Morgan Stanley. We break down revenue, massive compute lease commitments, founder voting control, and the valuation debate.

How Anthropic Uses a Four-Layer Narrative to Prop Up a $2 Trillion Valuation
Anthropic's estimated $2 trillion valuation: a breakdown of its positioning theory and four-layer narrative — civilization-level fear, identity, product, and business — and how founder essays, technical docs, and opinion leaders support the AI safety guardian valuation story.