#TOPAS: A Convergent Neuro-Symbolic Architecture for General Intelligence

3 messages · Page 1 of 1 (latest)

austere barn
#

Abstract & Summary: I just published a paper on Zenodo proposing TOPAS (Theoretical Optimization of Perception and Abstract Synthesis). The architecture attempts to solve the "symbol grounding problem" in LLMs by separating the system into two distinct modules:

Perception (Neural): Handles pattern recognition and messy data.

Synthesis (Symbolic): Handles abstract reasoning and logic enforcement.

Unlike standard Transformers that struggle with long-horizon logic, TOPAS converges these two approaches to maintain coherence without hallucinating.

Links:

📄 The Paper (Zenodo): https://zenodo.org/records/17683673

🤖 Live Implementation: We are testing this architecture in a live agent called BitterBot to see if the theory holds up in the wild: https://www.bitterbot.ai

What I'm looking for: I am looking for feedback specifically on the Synthesis Module described in Section 3. Does the logic enforcement seem robust enough to replace standard RLHF methods? I'm happy to discuss the architecture or the agent implementation!

Zenodo
Theoretical Optimization of Perception and Abstract Synthesis (TOPA...

The contemporary pursuit of Artificial General Intelligence (AGI) faces a "glass ceiling" in abstract visual reasoning, epitomized by the stagnation of Large Language Models on the ARC-AGI benchmark. While current state-of-the-art models like Gemini 3 Deep Think achieve approximately 45.1% on ARC-AGI-2, they lack the capacity for rigorous, multi...

warped parrot
#

I’m working on a small personal project that also uses modular perception, semantic memory and symbolic reasoning.
It’s very early-stage, but it’s nice to see that this direction is taken seriously in research. This is very interesting.

austere barn
#

Amazing - really appreciate the feedback. And I wish you luck on your own project! Keep us posted!