Test-Time Centralism

From the AI Research Compass wiki

Test-Time Centralism: Spending more compute at inference, through sampling, search and longer thinking, is a scaling axis as important as pretraining.

Beliefs

Test-Time Centralism argues that the field spent years scaling only one side of the ledger. Pretraining compute grew by orders of magnitude, while each answer was still produced in a single forward pass per token with no second thoughts. The doctrine moves the budget to inference: sample many candidates and vote, search over reasoning trees, verify and revise, or simply let the model think longer. A smaller model that thinks harder can beat a bigger one that answers at once.

It sits in the upper half, well toward the Language side. The language lean reflects that most test-time methods operate over text reasoning traces. The Monolith lean is real but moderate: inference-time search sometimes uses verifiers and simple scaffolds, yet the core remains one model allotted more compute. Its paradigm level is Steady, one step from Mainstream: inference scaling extends today’s recipe, and became a mainstream axis only recently.

Arguments

  • Sampling multiple reasoning paths and taking the majority answer consistently improves accuracy over a single greedy chain.
  • Tree-structured search over intermediate thoughts solves planning puzzles that single-pass generation fails on.
  • Compute-optimal analyses show that allocating test-time compute adaptively can rival scaling model parameters substantially.

Critiques

  • Inference compute is paid on every query, so the approach shifts cost onto users and deployment.
  • Gains depend on having a reliable verifier or a task where consistency signals correctness.
  • Efficiency-minded rivals argue that on-device and latency-sensitive applications cannot afford to think for minutes.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ −0.70, y ≈ +0.42, paradigm ≈ −0.25.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data