ANTARA: Adaptive Neural Thinking Architecture for Recursive Autonomy -- A Neuro-Symbolic Framework for Continual Self-Improvement via Recursive Global Workspaces and Latent Predictive Dynamics

ANTARA (Adaptive Neural Thinking Architecture for Recursive Autonomy) is a novel cognitive architecture designed to address catastrophic forgetting, static inference, and the lack of internal agency in contemporary deep learning systems.

This work introduces a unified framework for continual, self-regulated learning by integrating recursive global workspaces, neuromodulatory plasticity, and internal state optimization within a single computational system. Unlike conventional neural networks that operate as fixed optimization graphs, ANTARA actively modulates its internal dynamics in response to uncertainty, predictive surprise, and task interference.

Core Contributions

• Recursive Global Workspace (RGW): A deliberative reasoning mechanism enabling dynamic compute allocation under uncertainty via recursive attention cycles.

• Hierarchical Mixture of Experts (H-MoE): Task-adaptive routing that reduces cross-task interference and supports robust lifelong learning.

• Latent Predictive Dynamics: An entropy-driven neuromodulatory system that regulates learning plasticity based on predictive surprise.

• Internal State Optimization: A reinforcement learning–based controller (REINFORCE) that autonomously adjusts internal affine modifiers to stabilize training and inference.

• Unified Memory System: Holographic associative memory and orthogonal gradient projection to preserve previously acquired knowledge.

Empirical Results

Extensive continual learning experiments demonstrate that ANTARA achieves a 99.3% reduction in catastrophic forgetting, measured via backward transfer (BWT), compared to naive baselines. Ablation studies confirm the necessity of both memory protection and meta-cognitive regulation for stable long-term learning. Additional embodied simulations show that ANTARA preserves early motor skills while acquiring progressively complex behaviors in a robotic control setting.

Scope and Status

This release corresponds to ANTARA v2, a consolidated and experimentally validated version of the architecture building on earlier conceptual work. All core architectural ideas, experimental results, and evaluation protocols are publicly disclosed in this preprint.

This Zenodo record establishes public priority for the ANTARA architecture and its constituent mechanisms and is intended to support open research, reproducibility, and future peer-reviewed publication.