The shape of intelligence matters as much as its scale.
As models grow, so does the cost of building and using them. We study architectures that direct computation where it matters, reuse what a system already knows, and deliver useful capability within practical resource limits.
The central question.
When can selective computation outperform always-on scale?
Questions that follow
- How can structured reuse reduce repetition during learning and inference?
- Which measures capture capability per unit of compute, memory, energy, and cost?
How we study it
A fair comparison uses clear baselines, matched tasks, and stated resource budgets. It considers task quality alongside latency, memory use, and total compute, so an efficiency claim remains meaningful in practice.
What meaningful progress looks like
A useful improvement holds across relevant tasks, makes its resource trade-offs visible, and can be reproduced under stated conditions.
Explore all research programsRead the work behind these questions.
These independent papers and official guidance offer useful starting points for this field. Follow each link to review the original methods, evidence, and limits in context.