Virtual Logic Depth: new paper adds a fourth scaling axis by reusing weights to boost reasoning

qinzytech · x · 2026-09-04

An arXiv paper proposes Virtual Logic Depth (VLD) as a fourth scaling dimension beyond depth, width, and parameter count: reusing weights to grow effective algorithmic depth without adding parameters.

Key findings: knowledge capacity stays flat under VLD at fixed parameters; properly implemented VLD substantially improves reasoning without more parameters, decoupling reasoning from size; and the gains are robust across architectures and reuse schedules. The authors ask whether superintelligence requires ever-larger models or can be achieved by parameter reuse and deeper logic.

Original post →

More from Research

Research channel →