Topas: A Convergent Neuro-Symbolic Architecture for General Intelligence (zenodo.org)

TOPAS (Theoretical Optimization of Perception and Abstract Synthesis) is a convergent neuro-symbolic architecture proposed to push past the “glass ceiling” in abstract visual reasoning where large models have stalled. The authors report an Exact Match score exceeding 69% on the ARC-II benchmark—substantially higher than recent SOTA (~45% on ARC-AGI-2)—by combining principled theory (Free Energy Principle and Integrated Information Theory) with modular algorithmic components rather than treating learning as tabula rasa. TOPAS is framed as a “Canonical Unified Model” that explicitly integrates perception, symbolic planning, and memory to enable rigorous multi-hop and counterfactual reasoning. Technically, TOPAS is built from three novel subsystems: the Hebbian Triad separates concerns into ObjectSlots for perception, a NeuroPlanner for symbolic/neural hybrid reasoning, and VSA-based World Models for memory, connected via a type-safe “Sacred Signature.” Arbitration between neural intuition and symbolic candidates is handled by a Hypothesis Market using Hanson’s LMSR, while an Energy-Based Refiner performs Thermodynamic Refinement by minimizing a global free-energy functional to enforce logical consistency. The stack also uses the Muon optimizer for geometric regularization of sparse networks and Test-Time Training with LoRA for OOD generalization. Empirical validation across 121 tasks suggests TOPAS bridges statistical approximation and algorithmic synthesis—offering a reproducible design direction for more robust, explainable AGI-style reasoning.