Group Cognition Learning: Making Everything Better Through Governed Two-Stage Agents Collaboration

2026-04-30

AI Summary

This work addresses the challenges of modality dominance and spurious coupling in multimodal learning by proposing a Group Cognitive Learning framework, which introduces a novel two-stage fusion paradigm governed collaboratively by multiple agents. In the first stage, routing and auditing agents enable selective cross-modal interactions; in the second stage, consensus predictions are generated through shared latent factors and an aggregation agent, while preserving modality-specific representations. The method explicitly models shared factors and integrates sample-level gating with contribution-aware weighting to dynamically regulate modality interactions and fusion weights. Evaluated on CMU-MOSI, CMU-MOSEI, and MIntRec benchmarks, the approach achieves state-of-the-art performance in both regression and classification tasks, significantly enhancing the robustness and effectiveness of multimodal learning.

Abstract

Centralized multimodal learning commonly compresses language, acoustic, and visual signals into a single fused representation for prediction. While effective, this paradigm suffers from two limitations: modality dominance, where optimization gravitates towards the path of least resistance, ignoring weaker but informative modalities, and spurious modality coupling, where models overfit to incidental cross-modal correlations. To address these, we propose Group Cognition Learning (GCL), a governed collaboration paradigm that applies a two-stage protocol after modality-specific encoding. In Stage 1 (Selective Interaction), a Routing Agent proposes directed interaction routes, and an Auditing Agent assigns sample-wise gates to emphasize exchanges that yield positive marginal predictive gain while suppressing redundant coupling. In Stage 2 (Consensus Formation), a Public-Factor Agent maintains an explicit shared factor, and an Aggregation Agent produces the final prediction through contribution-aware weighting while keeping each modality representation as a specialization channel. Extensive experiments on CMU-MOSI, CMU-MOSEI, and MIntRec demonstrate that GCL mitigates dominance and coupling, establishing state-of-the-art results across both regression and classification benchmarks. Analysis experiments further demonstrate the effectiveness of the design.

Problem

• modality dominance
• spurious modality coupling
• multimodal learning
• cross-modal correlation
• fused representation

Innovation

• Group Cognition Learning
• Modality Dominance
• Spurious Modality Coupling
• Two-Stage Agent Collaboration
• Multimodal Fusion