DNON: A Brain-Inspired Architecture for Multi-Domain Reasoning with Specialized Neural Modules

Munish Singh, Namita Mittal, Arvind Ramachandra

Keywords: Neural orchestration, Information geometry, Brain-inspired architecture, Multi-tier memory, Lyapunov stability, Modular language models

TL;DR: DNON orchestrates specialized language models through information-theoretic routing on Riemannian manifolds, providing both theoretical convergence guarantees and improved performance on reasoning benchmarks.

Abstract:

Dynamic Neural Orchestration Networks (DNON) is introduced as a brain-inspired
architecture that composes specialized language models via information-theoretic
routing. DNON comprises four modules—Perception, Memory (short-term, long-term,
deep subconscious), Reasoning, and Executive—whose interactions are dynamically
regulated by mutual-information signals on information manifolds. The framework
provides a principled path for modular cognition and offers Lyapunov-style
convergence guarantees under reasonable assumptions. The implementation leverages
frozen foundation models (Claude Sonnet 4.5 and 3.7 and Mistral Pixtral Large-2502) while
training only the routing mechanisms and memory dynamics via gradient-based
optimization of mutual-information objectives. Empirically, DNON demonstrates
strong performance across diverse reasoning benchmarks, including arithmetic,
multi-hop inference, and adversarial compositional tasks, while reducing inference
cost relative to baselines and retrieval-augmented methods. Ablation studies
highlight the importance of the three-tier memory, as removing STM, LTM, or DSM
significantly degrades performance. DNON thus combines theoretical rigor and
practical gains for modular, interpretable large-model reasoning.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 741