Test-Time Scaling in the Wild: Why Exploitation, Not Exploration, Is the Bottleneck

TL;DR - A compute-normalized study of test-time scaling on open-ended tasks finds that generating better candidate pools is not the main limitation; reliably selecting or combining their best outputs is. Current reward models correlate weakly with true quality, making exploitation the bottleneck.

• Across medicine, law, finance, general chat, and creative writing, the best candidate improved steadily as inference compute increased.
• Reward models correlated with true quality at only about (ρ_v \approx 0.12), leaving candidate selection nearly random regardless of budget.
• Tree search worsened the problem through diversity collapse, while refinement produced a clear benefit on only one of five benchmarks.
• Candidate synthesis via Fusion was the only consistently effective approach, but recovered only about 40% of the available quality.