arXiv cs.AI
·
6 hours ago
Researchers introduced ANet Patu-1, a self-organizing consensus protocol for networks of AI agents that dynamically forms coalitions to optimize network value across different scaling regimes. The protocol achieves coordination in O(1) parallel consensus rounds by adaptively combining broadcast (V∝N), mesh (V∝N²), and group-forming (V∝2^N) network architectures. The work shows that heterogeneous agent networks using cheaper models can outperform homogeneous networks of stronger models, and that agents converging on problems naturally reconstruct this optimal protocol structure.
arXiv cs.AI
·
6 hours ago
Researchers developed a formal framework called analytic abduction with a κ-τ apparatus that identifies latent factors explaining complex observations while resisting premature conclusions by maintaining multiple candidate explanations until governance conditions are met. The framework uses a two-level causal cluster architecture (intra-cluster κ*, inter-cluster κ**) to prevent causal misattribution and was demonstrated in epidemiological crisis decomposition and adversarial cyber threat analysis. The approach produces weighted competing explanatory scenarios for human-AI coordination rather than a single imposed answer, enabling decision-makers to act while ambiguity remains unresolved.
arXiv cs.CL
·
6 hours ago
Researchers tested whether large language model agents communicating through latent representations preserve more information than text-based communication, using sparse autoencoder analysis and cross-model alignment. They found that latent channels retained 99.4% probe accuracy at 28-fold compression versus 80.4% for text, and text serialization destroyed 88% of SAE features. However, task-level evaluation showed latent communication provided no practical advantage over text, with lost features encoding mostly surface form rather than task-relevant semantics.
