ModSync for Generalized PINNs: prevents capacity failures

Heejo Kong, Beomchul Park, Sung-Jin Kim and Seong-Whan Lee propose Modular-Sparsity Synchronization, or ModSync, to fix a capacity-induced training failure in physics-informed neural networks. The paper was submitted to arXiv on 18 Jun 2026 as arXiv:2606.20156 and is listed as accepted by ICASSP 2026.

What is the capacity-induced failure the authors identify?

The failure is that overparameterized PINNs develop functional modularity: networks self-partition into task-exclusive modules that suppress cross-objective interaction and block convergence to Pareto-stationary points. Earlier conflict-averse optimizers attempt to reduce gradient interference between residual and boundary losses, but the authors show those schemes weaken as model capacity grows, allowing task-exclusive wiring to emerge and undermine multi-objective training.

The paper frames the problem specifically as a capacity-driven breakdown of cross-objective coupling in generalized PINNs, where the intended interaction between residual and boundary objectives is lost when networks become highly parameterized.

How does ModSync work in practice?

ModSync integrates structural optimization into conflict-averse training by penalizing task-exclusive connections while preserving interaction-promoting pathways. In short, the framework adds a structural regularizer that discourages the network from creating connections used exclusively by one objective, keeping pathways that support interaction across objectives.

The authors describe ModSync as a complement to existing conflict-averse optimization schemes: rather than only adjusting gradient signals, ModSync enforces a wiring-level constraint so the network cannot simply segregate itself into separate modules for each loss. The paper states that code is available at the project's URL provided in the submission.

What evidence do the authors give that ModSync works?

The paper reports extensive experiments across diverse PDE benchmarks where ModSync "consistently prevents capacity-driven failures, sustains robust cross-objective coupling, and achieves state-of-the-art accuracy." The abstract presents these results as the main empirical claim, contrasting ModSync with prior conflict-averse schemes whose effectiveness deteriorates with increased model capacity.

The submission also highlights practical signals: the work appears on arXiv as arXiv:2606.20156 and is marked accepted by ICASSP 2026, indicating peer validation and a forthcoming conference presentation.

Why it matters

Physics-informed neural networks aim to encode PDE constraints in training objectives, so losses for residuals and boundary conditions must interact during optimization. If overparameterized networks self-partition, those interactions vanish and solutions stop improving despite larger models. ModSync addresses the architectural side of that problem, shifting focus from only tuning optimizers to shaping network connectivity. That matters for researchers pushing PINNs with bigger models and for practitioners who rely on robust convergence when scaling architectures.

What to watch

Look for the ICASSP 2026 proceedings and the authors' code release to inspect implementation details and benchmark scripts. The next concrete signals will be the conference presentation and community experiments reproducing the reported state-of-the-art accuracy on the described PDE benchmarks.

References and metadata

• Title: Modularity-Free Conflict-Averse Training for Generalized PINNs
• Authors: Heejo Kong, Beomchul Park, Sung-Jin Kim, Seong-Whan Lee
• arXiv ID: arXiv:2606.20156 (submitted 18 Jun 2026)
• Notes: Accepted by ICASSP 2026; code URL included in the submission.