Resolving Oversmoothing with Opinion Dissensus
Keqin Wang, Yulong Yang, Ishan Saha, Christine Allen-Blanchette
TL;DR
This paper introduces a novel GNN model inspired by nonlinear opinion dynamics that resists oversmoothing, outperforming existing methods and maintaining effective node representations over many layers.
Contribution
It proposes the Behavior-Inspired Message Passing (BIMP) GNN, which uses nonlinear opinion dynamics to prevent oversmoothing and improve deep GNN performance.
Findings
BIMP resists oversmoothing beyond 100 layers.
BIMP outperforms existing architectures with oversmoothing mitigation.
BIMP adapts well to different dataset types.
Abstract
While graph neural networks (GNNs) have allowed researchers to successfully apply neural networks to non-Euclidean domains, deep GNNs often exhibit lower predictive performance than their shallow counterparts. This phenomena has been attributed in part to oversmoothing, the tendency of node representations to become increasingly similar with network depth. In this paper we introduce an analogy between oversmoothing in GNNs and consensus (i.e., perfect agreement) in the opinion dynamics literature. We show that the message passing algorithms of several GNN models are equivalent to linear opinion dynamics models which have been shown to converge to consensus for all inputs regardless of the graph structure. This new perspective on oversmoothing motivates the use of nonlinear opinion dynamics as an inductive bias in GNN models. In our Behavior-Inspired Message Passing (BIMP) GNN, we…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The analogy between oversmoothing and consensus dynamics is presented clearly and supported by formal statements, which may help unify prior perspectives on oversmoothing. 2. The paper states assumptions and propositions explicitly, with proofs delegated to the appendix, making the theoretical section readable. 3. The empirical plots demonstrate that the proposed model maintains stable Dirichlet energy and accuracy even when simulated for up to 10^3 layers, which supports the main claim of th
1. Necessity of focusing on oversmoothing is not fully justified. The paper treats oversmoothing as a central obstacle in GNN design, but it is unclear why controlling depth-induced feature collapse is still an impactful problem in practice. Many modern architectures (e.g., graph transformers, attention-based models with residual paths, or shallow-but-expressive methods) already achieve strong performance with 2–4 layers, often outperforming deeper continuous-depth models regardless of oversmoo
The paper presents an innovative connection between GNNs and opinion dynamics. The proposed GNN model is not susceptible to oversmoothing and outperforms state-of-the-art models.
Some relevant related works may be missing from the paper. For example, there is no reference to how this work relates to previous work on oversmoothing, such as *A Note on Over-Smoothing for Graph Neural Networks by Chen Cai and Yusu Wang* or *Residual Connections and Normalization Can Provably Prevent Oversmoothing in GNNs by Michael Scholkemper, Xinyi Wu, Ali Jadbabaie, and Michael T. Schaub*, which appeared in ICLR 2025. Additionally, there is insufficient reference to discrete-time opinion
1. Theoretical originality: The paper establishes a clear and elegant analogy between oversmoothing in GNNs and opinion consensus in nonlinear dynamics, offering a novel theoretical lens for understanding message passing. 2. Well-motivated nonlinear model: The introduction of nonlinear saturation functions, bifurcation-controlled attention, and external inputs provides a principled way to prevent consensus, moving beyond heuristic anti-oversmoothing tricks. 3. Compreh
1. Ablation clarity The authors should include an additional Dirichlet energy ablation experiment (on activation functions and inductive bias) to verify the impact of different functions and modules on oversmoothing. Minor issues * Please provide citations for the baseline models in the main text, and ideally include brief descriptions of these baselines.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsOpinion Dynamics and Social Influence