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

A team of researchers from multiple institutions has introduced Group Cognition Learning (GCL), a novel governed collaboration paradigm designed to overcome two persistent issues in centralized multimodal learning: modality dominance (where optimization favors the strongest modality, ignoring weaker but informative signals) and spurious modality coupling (overfitting to incidental cross-modal correlations). The paper, accepted at ICML 2026, proposes a two-stage protocol applied after modality-specific encoding. In Stage 1 (Selective Interaction), a Routing Agent proposes directed interaction routes, while an Auditing Agent assigns sample-wise gates to emphasize exchanges that yield positive marginal predictive gain and suppress 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, keeping each modality as a specialized channel.

Extensive experiments on three standard multimodal benchmarks—CMU-MOSI (sentiment regression), CMU-MOSEI (sentiment classification), and MIntRec (intent recognition)—show that GCL establishes new state-of-the-art results across both regression and classification tasks. Analysis experiments further validate the design choices, demonstrating clear mitigation of dominance and coupling effects. The work represents a shift from end-to-end fusion pipelines to governed multi-agent collaboration, potentially influencing future multimodal AI systems in areas like emotion recognition, human-computer interaction, and autonomous decision-making. The full manuscript is available on arXiv (ID: 2605.00370).