Conjecture on machine-native neural architectures for reasoning

Investigate and test the conjecture that closing the reasoning gap in neural computers does not require biologically inspired, brain-like designs but instead benefits from explicitly machine-native neural architectures that incorporate discrete operations, compositional structures, and verifiable computation.

Background

After noting that video-based prototypes show early I/O control but limited reasoning reliability, the authors propose a design-direction conjecture.

They argue that rather than mimicking human cognition, neural computers may benefit from architectures native to machines, emphasizing discrete and compositional computation and verifiability within neural systems.

References

We emphasize that the following is a conjecture rather than a conclusion drawn from our experiments. Closing the reasoning gap may not require designing neural networks that more closely mimic animal cognition or the human brain.

— Neural Computers  (2604.06425 - Zhuge et al., 7 Apr 2026) in Section 4 (Position: Toward Completely Neural Computers) — Additional Thoughts, A hypothesis: machine-native neural architectures