Test-time Compute

Instead of making the model bigger, this approach improves performance by using more computation at the moment of answering (inference time). Letting the model think longer, try multiple solution paths, and check its own answers — all of this counts as extra computation at inference time.

This emerged as a second axis after "scaling by training more" ran into debates about its limits. Think of it as extending exam time (inference) instead of studying more (training). The performance curves of reasoning models — graphs of "thinking time vs. accuracy" — are all about this.