CRAFT: Conflict-Resolved Aggregation for Federated Training
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "CRAFT: Conflict-Resolved Aggregation for Federated Training".
Jane: Here's the gist of the paper. The paper proposes CRAFT, a new aggregation framework that treats global updates as a geometric correction problem to resolve conflicts in federated learning under heterogeneous data distributions The authors derive a closed-form expression for the constrained optimization problem, avoiding iterative solvers CRAFT promotes a common-descent structure and mitigates conflicts through its projection geometry.
Tom: Conflict Resolution Mechanism CRAFT reformulates federated aggregation as a reference-anchored constrained least-squares problem to find the global update closest to a prespecified reference direction while satisfying conflict-free alignment constraints The core formulation involves finding an update g closest to a reference direction gˆt while satisfying the constraint Utg = ρt.
Jane: This leads to a minimum-norm correction problem where the update is gt = gˆt + ∆t, with ∆t = U†ρt − Utgˆt.
Tom: The reference direction gˆt is proposed as the normalized global update from the previous round, gˆt = U(gt−1).
Paper summary: Tom: So we’re looking at this paper called "CRAFT: Conflict-Resolved Aggregation for Federated Training," and what they're saying is that the way we combine updates from different clients in federated learning, especially when their data isn't all the same—heterogeneous data distributions—is a major problem.
Jane: Exactly. The core idea here is that simple averaging can create a global update that’s good for everyone overall, but it might actually hurt some specific clients because those updates conflict with what they need, causing their local accuracy to drop >
Lu: They propose CRAFT as a new framework that treats the global update like a geometric correction problem, which they say helps them solve these conflicts by finding the update that's closest to a reference direction while making sure it respects certain conflict-free alignment constraints >
Meng: That sounds complicated. So, what’s the big win here? Is this just another trick to make the global model better without hurting anyone? We need to know how much of a problem this actually solves at scale >
Tom: Well, what they claim is they’ve derived a closed-form expression for solving that optimization problem directly. That means instead of running tons of iterative solvers every round, you can calculate the solution in one go, which cuts down on computational overhead >
Jane: And they use something called a layer-wise adaptation too, which addresses conflicts at different levels of feature detail. They solve the correction separately for each layer >
Lu: That layer-wise approach is important because it lets them handle conflicts flexibly. It prevents a big conflict in one part of the model from ruining useful parts in another >
Meng: Okay, so if I'm thinking about deploying this to production, how much extra computation is this framework adding per round? Is it manageable for our existing infrastructure >
Tom: The paper states the server-side computational complexity is O(d m squared plus Q m cubed) per round >
Jane: That’s a specific measure of the cost. But beyond just the math, they provide theoretical guarantees that this approach promotes a common-descent structure and uses projection geometry to handle those conflicts >
Lu: The analysis shows a convergence bound under full client participation and local gradient-descent training, specifically when you pick your learning rate η g equal to the cube root of one over T >
Paper summary: Meng: A convergence bound is fine on paper, but does it actually hold up when we introduce real-world noise or data drift? What about the practical deployment side of that theorem >
Tom: They also show that this layer-wise analysis keeps the same asymptotic form for convergence, but the drift constant becomes a sum of terms specific to each layer >
Jane: So, what does this mean for someone just using federated learning, maybe without reading all the math? It sounds like it tackles the core issue of those conflicting updates directly >
Lu: They reformulate aggregation as a reference-anchored constrained least-squares problem that finds the global update closest to a prespecified reference direction while satisfying conflict-free alignment constraints where g is close to g hat t while satisfying the constraint U t g equals rho t >
Meng: Wait, they use the previous global update from the last round as that reference direction, which is also used in their layer-wise approach >
Tom: Yes, they define that reference direction as g hat t equal to U of g sub t minus one >
Jane: But what about the performance on actual datasets? The experiments show CRAFT improves mean accuracy while simultaneously reducing client-level disparity across various datasets and models >
Lu: They show that CRAFT consistently achieves the highest mean accuracy across metrics like FEMNIST and CIFAR-ten with relative gains of seven point eight percent over the second best FedAvgM on FEMNIST and fifty-three point five percent over ConFIG on CIFAR-ten >
Meng: That's a big jump in accuracy numbers. But does it hold up when we look at fairness? The paper shows it improves both the best-ten percent and worst-ten percent client accuracies simultaneously >
Jane: They also show that for deeper networks, the standard deviation of client accuracy is lowest with CRAFT, meaning the final results are tighter across all clients >
Lu: The convergence analysis under full participation shows that if you use η g equal to Theta one over square root of T, the bound for the squared gradient norm is one over T summed over t equals zero nabla F theta t squared is O one over square root of T plus O η l K max minus one >
Tom: That's a specific rate for how fast things settle down, showing it converges at that rate under those conditions >
Paper summary: Jane: It’s important to remember what the authors say about limitations too. They mention that they use the previous global update as the reference direction, and they suggest that using moving averages over multiple previous rounds or adaptive references based on recent conflicts could improve that >
Meng: And there’s another thing: they note that the equality constraint might only be satisfied approximately in implementation, which leads to a residual term R t which is bounded in Corollary A one two >
Lu: They also mention that the analysis doesn't rely on assumptions about the underlying data distributions and it accommodates heterogeneous local update steps Ki across clients >
Jane: So, what does this all mean for someone who only cares about the final model performance, not the deep math? It seems like a robust way to handle messy, real-world federated training scenarios >
Tom: CRAFT is designed as a generic aggregation operator that can be used as a drop-in replacement for standard weighted averaging in existing personalized federated learning pipelines without changing the original personalization objective >
Lu: It strengthens personalized FL methods by improving the internal aggregation step without altering the personalization objective itself, and it’s not sensitive to initial direction because the reference is just used as a soft prior >
Jane: So, if you’re thinking about how to improve your existing federated setup—maybe you're using something like FedAvgM—CRAFT offers a way to plug this new aggregation step in to get better results without rewriting your whole personalization strategy >
Meng: From an engineering standpoint, the server-side complexity is O d m squared plus Q m cubed per round, which gives us a concrete idea of the computational load we’d be looking at >
Tom: And it’s important to stress that while they show great empirical results on CIFAR-ten and FEMNIST, those gains are measured against existing methods like ConFIG and FedAvgM > <ref:2605.21317#pg1>
Jane: So, to wrap up this discussion on CRAFT: the paper proposes a framework that fixes conflict resolution by treating it as a geometric correction problem with a layer-wise twist, offering better accuracy while keeping things fairer for different clients >
Lu: It’s about moving away from simple averaging toward something more structured when data is non-IID >
Meng: And the big takeaway for deployment is that it has measurable complexity and provides concrete performance gains over established methods on these specific tests >
Tom: We'll stick with the name CRAFT, as they’ve shown it consistently improves mean accuracy while simultaneously reducing client-level disparity >
Conclusion: Tom: So, to wrap up this discussion on CRAFT: Conflict-Resolved Aggregation for Federated Training, we're talking about a new way to handle those messy disagreements when different clients send updates in federated learning.
Jane: It basically takes those conflicting updates and treats them like a geometric correction problem to find the best global update that respects certain alignment rules.
Lu: The authors derived a closed-form expression for this optimization problem, which means they don't have to run these iterative solvers every time they aggregate updates.
Meng: That’s huge for deployment because it cuts down on the computational work on the server side of things per round.
Tom: Yeah, that’s right. And they show that this approach promotes a common-descent structure, which is good for getting those models to converge steadily without getting stuck in bad local minima.
Jane: It also uses something called layer-wise adaptation, which lets the framework handle conflicts at different levels of detail in the data layers.
Lu: That layer-wise setup allows them to solve the correction separately for each layer, which keeps the resolution flexible and prevents a big conflict in one part of the model from messing up another part.
Tom: And empirically, they’ve seen it improve mean accuracy while actually reducing that disparity between clients.
Jane: They show gains over existing methods like FedAvgM on datasets like FEMNIST and CIFAR-ten and they even show improvements for both the best ten percent and the worst ten percent of clients.
Meng: So, if you’re looking at a production system, this suggests a way to get better results without having to completely overhaul your existing aggregation pipeline.
Tom: Exactly. It’s designed as a drop-in replacement for weighted averaging in personalized federated learning pipelines without changing the original personalization objective.
Jane: But they do have some caveats they mention, like using the previous global update as a reference direction, which they suggest might be improved with moving averages over several rounds.
Lu: They also flag that the equality constraint in their math might only be satisfied approximately in real-world implementation, leading to a residual term that's bounded.
Tom: It means we have to watch out for those implementation details and see how much that residual error actually hurts the final performance on our actual traffic.
Jane: Still, it’s a solid framework because they don't rely on strict assumptions about the underlying data distributions or client update steps.
Lu: And that flexibility is key because it works even when clients have different local update steps, which isn't always true in practice.
Tom: So, CRAFT seems to be a very structured way to tackle the conflict issue in federated learning by turning it into a geometric problem with some smart layer-wise tuning.
Jane: It moves us toward aggregation methods that are more robust when dealing with heterogeneous data distributions across clients.
Ziqi Wang, Qiang Liu, Nils Thuerey
Friedrich-Alexander-Universität Erlangen-Nürnberg · Technical University of Munich · Munich Center for Machine Learning
cs.LG
Submitted: 2026-05-20
Updated: 2026-11-09
Comments: Accepted by NeurIPS 2026. Funded by the European Union (Horizon Europe MSCA project ModConFlex, grant number 101073558)
Code: https://github.com/tum-pbs/CRAFT
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 88/100
The gist: The paper proposes CRAFT, a new aggregation framework that treats global updates as a geometric correction problem to resolve conflicts in federated learning under heterogeneous data distributions…
Key concepts
- Conflict Resolution Mechanism
- CRAFT reformulates aggregation as finding the global update closest to a reference direction while satisfying constraints. It calculates an update correction ($\Delta t$) based on the difference between the required constraint alignment and the current update, effectively minimizing conflicts through a minimum-norm correction problem.
- Layer-wise Adaptation
- The framework addresses conflicts at different feature granularities by solving a separate correction problem for each layer ($q$). This allows for flexible conflict resolution, ensuring that severe conflicts in one layer do not negatively impact useful components in other layers, leading to better overall performance.
- Common-Descent Structure
- The framework establishes a common-descent structure, which is a mathematical property ensuring that the objective function generally decreases during training. This provides theoretical guarantees for convergence and helps ensure that the aggregation process moves towards a stable global optimum.
- Reference Direction ($\hat{g}_t$)
- The reference direction is proposed as the normalized global update from the previous round. It serves as a soft prior for the current update, guiding it toward a desirable direction while maintaining adherence to conflict-free alignment constraints.
Terminology
Summary
The paper proposes CRAFT, a new aggregation framework that treats global updates as a geometric correction problem to resolve conflicts in federated learning under heterogeneous data distributions The authors derive a closed-form expression for the constrained optimization problem, avoiding iterative solvers CRAFT promotes a common-descent structure and mitigates conflicts through its projection geometry. p. 0 We derive a closed-form expression
for the constrained optimization problem, avoiding the computational overhead of
iterative solvers. Furthermore, we use a layer-wise adaptation to address conflicts
at varying feature granularities.Where the paper says it
Conflict Resolution Mechanism
CRAFT reformulates federated aggregation as a reference-anchored constrained least-squares problem to find the global update closest to a prespecified reference direction while satisfying conflict-free alignment constraints The core formulation involves finding an update g closest to a reference direction gˆt while satisfying the constraint Utg = ρt. This leads to a minimum-norm correction problem where the update is gt = gˆt + ∆t, with ∆t = U†ρt − Utgˆt. p. 4 Let ∆ := g − gˆt denote the correction applied to the reference direction. 3) is equivalently
written as the following minimum-norm correction problem. Since this correction
lies in the row space of Ut, the update modifies gˆt only along directions spanned by the participating
client gradients, yielding the closest conflict-free update to the reference direction. 3 Choice of reference direction
We propose using the normalized global update from the previous round as the reference direction.Where the paper says it
Where the paper says it
Layer-wise Adaptation and Geometry
The framework incorporates a layer-wise adaptation to address conflicts at varying feature granularities For each layer q, the server constructs a normalized update matrix Uq t and reference direction gˆq t. The correction for each layer is solved separately as gq t = gˆq t + ∆qt, with ∆qt = Uq†ρt − Uqtgˆqt. p. 4 For each layer q ∈ [Q], we construct the normalized update matrix U
q
t
and reference
direction gˆ
q
t
. We then solve the correction problem separately for each layer as
g
q
t = gˆ
q
t + ∆
q
t
, with ∆
q
t =
U
q
t
†ρt − U
q
tgˆ
q
t
. We then solve the correction problem separately for each layer as
g
q
t = gˆ
q
t + ∆
q
t
, with ∆
q
t =
U
q
t
†ρt − U
q
tgˆ
q
t
. 7)
This layer-wise approach allows for more flexible conflict resolution, preventing severe conflicts
in one layer from suppressing useful components in another.Where the paper says it
Where the paper says it
Theoretical Guarantees and Convergence
The theoretical analysis establishes a common-descent structure and provides finite-time convergence guarantees Theorem A.4 yields a convergence bound for the squared gradient norm under full client participation and local gradient-descent training. p. 5 Here, we
briefly summarize the main convergence result. For exposition, the main theorem is stated under full
client participation with local gradient-descent training. 4 yields
1
T
T
X−1
t=0
∥∇F(θt)∥
2 = O
1
√
T
+ O
ηl(Kmax − 1)
(4. 8)
under the standard choice ηg = Θ(1/
√
T). We further
introduce a layer-wise implementation that resolves conflicts at the granularity of model
layers, making the method suitable for deep neural networks. The accompanying theoretical
analysis characterizes the projection geometry, establishes its common-descent structure,
and provides finite-time convergence guarantees.Where the paper says it
Where the paper says it
Where the paper says it
Empirical Performance and Fairness
Extensive experiments demonstrate that CRAFT improves mean accuracy while maintaining strong performance for both head and tail clients Table 4 shows that CRAFT consistently achieves the highest mean accuracy across various datasets and models. Furthermore, CRAFT improves best-10% and worst-10% client accuracies, indicating simultaneous gains for head and tail users. p. 6 Importantly, the
improvement is not driven solely by clients that already achieve high accuracy. CRAFT also raises
both the best-10% and worst-10% client accuracies, indicating simultaneous gains for head and tail
users. These gains indicate that CRAFT
does not simply trade tail users for head users or vice versa. This is also reflected in the standard
deviation metric shown in Figure 5, where CRAFT achieves the lowest standard deviation on deeper
networks, providing a tighter distribution of final client accuracy.Where the paper says it
Where the paper says it
Extensibility and Practicality
CRAFT is designed as a generic aggregation operator that can serve as a drop-in replacement for weighted averaging in existing personalized FL pipelines It can strengthen personalized FL methods by improving their internal aggregation step without changing their personalization objective. The framework is not sensitive to the initial direction because the reference is used as a soft prior, and it has an intrinsic self-correction mechanism that mitigates early-stage sampling bias. The server-side computational complexity is O(dm2 + Qm3) per round. p. 4 Solving the Q corresponding pseudoinverse problems costs
O(Qm3
). Thus, the conservative per-round server complexity is O(dm2 + Qm3
), typically dominated by O(dm2
) when d ≫ m.Where the paper says it
Limitations and Future Directions
A limitation noted is that CRAFT uses the previous global update as the reference direction, which could be improved by using moving averages over multiple previous rounds or adaptive references based on recent client conflicts. Another area for future work is making the alignment target ρt adaptive based on training progress, loss, gradient conflicts, or fairness requirements. p. 8 More adaptive targets could reweight clients based on training progress, loss,
gradient conflicts, or fairness requirements. Understanding how these reference and alignment
choices affect accuracy, stability, and client-level fairness is an important direction for future work. The first term is the usual non-convex optimization rate,
and the second is the deterministic bias induced by multiple local steps, which vanishes when ηl
is
also chosen to decay with T. Notably, the analysis does not rely on assumptions about the underlying
data distributions and accommodates heterogeneous local update steps Ki across clients.Where the paper says it
Where the paper says it
Acknowledgments
Ziqi Wang acknowledges funding from the European Union’s Horizon Europe MSCA project ModConFlex (grant number 101073558) Qiang Liu acknowledges support from the China Scholarship Council (No. 202206120036) The authors are grateful to Enrique Zuazua for invaluable discussions. p. 9 202206120036). We are grateful to Enrique Zuazua for invaluable discussions. We are grateful to Enrique Zuazua for invaluable discussions. We
acknowledge the scientific support and HPC resources provided by the Erlangen National High Performance Computing Center (NHR@FAU) of the Friedrich-Alexander-Universität Erlangen-Nürnberg
(FAU). We
acknowledge the scientific support and HPC resources provided by the Erlangen National High Performance Computing Center (NHR@FAU) of the Friedrich-Alexander-Universität Erlangen-Nürnberg
(FAU). The hardware is partially funded by the German Research Foundation (DFG).Where the paper says it
Where the paper says it
Where the paper says it
References
The paper cites numerous related works, including FedAvg [16], FedProx [11], and ConFIG [15]. The references cover topics such as heterogeneity-mitigating methods, conflict mitigation techniques like FedFV [27] and FedLF [20], and multi-objective optimization approaches like AFL [17] and FedMGDA+ [7]. p. 5 We simulate non-IID
data using highly heterogeneous Dirichlet partitioning [6] with a concentration parameter α = 0. We compare CRAFT with representative methods covering simple averaging (FedAvg
[16]), drift mitigation (FedProx [11], FedNova [26]), server momentum/adaptive optimization (FedAvgM [6], FedAdam/AdaGrad/Yogi [22]), fairness-aware objectives (AFL [17], q-FedAvg [12],
FedMGDA+ [7]), and conflict-aware aggregation (FedFV [27], FedLF [20], ConFIG [15]). We simulate non-IID
data using highly heterogeneous Dirichlet partitioning [6] with a concentration parameter α = 0. We compare CRAFT with representative methods covering simple averaging (FedAvg
[16]), drift mitigation (FedProx [11], FedNova [26]), server momentum/adaptive optimization (FedAvgM [6], FedAdam/AdaGrad/Yogi [22]), fairness-aware objectives (AFL [17], q-FedAvg [12],
FedMGDA+ [7]), and conflict-aware aggregation (FedFV [27], FedLF [20], ConFIG [15]). Similar gradient-conflict issues also arise in multi-task learning (MTL) [29], continual
learning (CL) [4, 23], and physics-informed neural networks (PINNs) [8, 15]. For training PINNs,
ConFIG [15] formulates conflict resolution as solving a linear system. [6] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. Measuring the effects of non-identical
data distribution for federated visual classification. [12] Tian Li, Maziar Sanjabi, Ahmad Beirami, and Virginia Smith. Fair resource allocation in
federated learning. Statistical and system heterogeneity are central challenges in
FL [9]. FedAvg often struggles with non-IID data, leading to slow convergence or biased global
models. In FL, FedFV [27] identifies gradient conflicts as a key source of unfairness
and proposes iteratively removing conflicting components before averaging. FedLF [20] and Fed2 To address this limitation, CRAFT formulates
conflict resolution as a constrained least-squares problem with a momentum-like reference direction,
yielding a more consistent and effective solution. Multi-objective optimization. FedRep [2] decomposes the model
into a shared backbone and client-specific heads, enabling personalization by locally fine-tuning the
client-specific components. Ditto [13], instead of separating model components, learns an additional
personalized model for each client by regularizing it toward the shared global model with a proximal
term. Adaptive server optimizers, such as FedAvgM [6] and FedOpt [22],
have also been explored to improve convergence under heterogeneity [9]. These methods primarily
target global optimization performance, but high mean accuracy does not guarantee reliable accuracy
for every client. [4] Mehrdad Farajtabar, Navid Azizan, Alex Mott, and Ang Li. Orthogonal gradient descent for
continual learning. [8] Youngsik Hwang and Dong-Young Lim. Dual cone gradient descent for training physicsinformed neural networks. Conflict mitigation. In FL, FedFV [27] identifies gradient conflicts as a key source of unfairness
and proposes iteratively removing conflicting components before averaging. We further derive an explicit closed-form
solution based on a Moore–Penrose correction, avoiding iterative projection or quadraticprogramming solvers and keeping the server-side computation lightweight. • Momentum-like reference and layer-wise adaptation. MDFG [19] further develop this direction through layer-wise constraints or fairness-guided gradient
adjustment. Similar gradient-conflict issues also arise in multi-task learning (MTL) [29], continual
learning (CL) [4, 23], and physics-informed neural networks (PINNs) [8, 15]. In contrast, CRAFT anchors aggregation to a momentum-like reference and applies
the minimum correction needed to enforce positive alignment with every participating client. Relation to personalized FL. Lower values indicate a tighter distribution and therefore better fairness. 4 Integration with personalized federated learning
CRAFT provides a new aggregation operator rather than a personalized FL algorithm that maintains
a separate model for each client. We further derive an explicit closed-form
solution based on a Moore–Penrose correction, avoiding iterative projection or quadraticprogramming solvers and keeping the server-side computation lightweight. • Momentum-like reference and layer-wise adaptation. 3
CIFAR-100 & ResNet-110
CRAFT
FedAvg
FedProx
FedNova
FedAvgM
FedAdaGrad
FedAdam
FedYogi
AFL
FedFV
FedMGDA+
qFedAvg
FedLF
ConFIG
Figure 2: Evolution of mean test accuracy over communication rounds across datasets and model
architectures. CRAFT consistently converges faster and reaches higher final accuracy, with especially
stable gains on more challenging datasets and deeper neural networks.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
CRAFT ConFIG FedAvgM FedAdaGrad FedFV FedAdam
The table compares various aggregation methods, showing that CRAFT consistently improves mean accuracy while simultaneously reducing client-level disparity. CRAFT achieves higher mean client accuracy while simultaneously reducing client-level disparity. p. 1 51
CIFAR-10 & ResNet-20
Figure 1: Mean (left, ↑) and standard deviation (right, ↓) of per-client accuracy. CRAFT achieves
higher mean client accuracy while simultaneously reducing client-level disparity.Where the paper says it
Figure 1
Figure 1 shows that CRAFT achieves higher mean client accuracy while simultaneously reducing client-level disparity.
Table 3
Table 3 reports test accuracy metrics for FEMNIST and CIFAR-10, with additional results provided in Appendix C. CRAFT consistently improves mean accuracy, with relative gains of 7. p. 6 2 Overall performance
Table 1 reports test accuracy metrics for FEMNIST and CIFAR-10, with additional results provided
in Appendix C. CRAFT consistently improves mean accuracy, with relative gains of 7. CRAFT consistently improves mean accuracy, with relative gains of 7. 8% over
the second-best FedAvgM on FEMNIST and 53. 8
Mean client accuracy
x1. 29
CIFAR-10 & ResNet-20
CRAFT
FedNova qFedAvg
FedMGDA+
FedFV
FedAvg
FedProx
FedAdam
FedYogi FedAdaGrad
FedLF ConFIG
AFL FedAvgM
0. 2 Overall performance
Table 1 reports test accuracy metrics for FEMNIST and CIFAR-10, with additional results provided
in Appendix C. CRAFT consistently improves mean accuracy, with relative gains of 7.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Figure 3
Figure 3 illustrates that CRAFT moves most of the mass to the right and reduces the low-accuracy tail on FEMNIST and CIFAR-10, producing a more favorable distribution with improved mean and tail performance. Figure 4 shows that CRAFT keeps the mean and best-client performance high while lifting the worst-client curve, showing a much fairer distribution. p. 7 Several baselines obtain reasonable mean accuracy while showing a large gap between the best and
worst clients, which reflects a polarized client distribution. CRAFT keeps the mean and best-client
performance high while lifting the worst-client curve, showing a much fairer distribution.Where the paper says it
Figure 5
Figure 5 illustrates that CRAFT achieves the lowest standard deviation on deeper networks, providing a tighter distribution of final client accuracy. The comparison shows that CRAFT improves mean accuracy without sacrificing either end of the client distribution. p. 7 Mean accuracy alone can hide severe
disparity because two global models with the same mean may behave very differently across clients. The results in Table 1 show that CRAFT improves mean accuracy without sacrificing either end
of the client distribution.Where the paper says it
Figure 6
Figure 6 reports heatmaps over different server and client learning-rate combinations, demonstrating CRAFT’s strong extensibility by improving mean test accuracy across the grid. The comparison between vanilla Ditto and Ditto with CRAFT yields average relative improvements of 6. p. 8 4
CIFAR-100 & ResNet-20
Figure 6: Comparison between vanilla Ditto and Ditto with CRAFT across various settings. Replacing
Ditto’s average aggregation with the CRAFT operator yields average relative improvements of 6. 8
Mean client accuracy
x1. 29
CIFAR-10 & ResNet-20
CRAFT
FedNova qFedAvg
FedMGDA+
FedFV
FedAvg
FedProx
FedAdam
FedYogi FedAdaGrad
FedLF ConFIG
AFL FedAvgM
0. 1 Experimental setup
Datasets, models, and participation. We use the FEMNIST [1] and CIFAR-10/100 [10] datasets.Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved aggregation mechanism for federated learning by treating it as a geometric correction problem, which is
finding the update closest to a reference direction while satisfying conflict-free alignment constraints.
This allows the system to enforcepositive alignment between the global update and every participating client update,
leading to aconflict-free solution whenever the conflict-free alignment constraints are feasible.
-
Implementation of layer-wise adaptation for conflict resolution, where
conflicts in shallow layers may signal input distribution shifts, while conflicts in deeper layers may stem from label skew.
This allows the framework to resolve conflicts at the granularity of model layers, preventingsevere conflicts in one layer from suppressing useful components in another.
-
Enhanced performance across heterogeneous benchmarks by improving both mean accuracy and client-level fairness. The system is shown to
improve mean accuracy while reducing performance disparity across clients compared to state-of-the-art baselines
and achieves amore favorable distribution with improved mean and tail performance.
Abstract
The aggregation of conflicting client updates remains a fundamental bottleneck in federated learning (FL) under heterogeneous data distributions. Naive averaging can produce a global update that improves the global objective while conflicting with specific clients, causing degradation for those clients. In this work, we propose CRAFT (Conflict-Resolved Aggregation for Federated Training), a new aggregation framework that treats the global update as a geometric correction problem. We formulate aggregation as finding the update closest to a reference direction while satisfying conflict-free alignment constraints. We derive a closed-form expression for the constrained optimization problem, avoiding the computational overhead of iterative solvers. Furthermore, we use a layer-wise adaptation to address conflicts at varying feature granularities. We provide a theoretical analysis showing that CRAFT promotes a common-descent structure and mitigates conflicts through its projection geometry. Extensive experiments on heterogeneous benchmarks demonstrate that CRAFT improves the accuracy of the global model while reducing performance disparity across clients compared with state-of-the-art baselines. The source code for CRAFT is available at https://github.com/tum-pbs/CRAFT.
Sources
- LEAF: A Benchmark for Federated Settings
- Measuring the Effects of Non-Identical Data Distribution for Federated Visual Classification
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis