A PRM-guided adaptive test-time compute framework outperforms uniform scaling with large gains on MATH-500 and several-fold improvements on AIME24 and AMO-Bench by dynamically selecting strategies and pruning low-reward paths.
4.NUMERIC RISK: If the problem involves non-trivial arithmetic, numeric bounds, probabilities, or quantitative calculations→ includenumeric verifierAFTER the main reasoning tool
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
What If We Allocate Test-Time Compute Adaptively?
A PRM-guided adaptive test-time compute framework outperforms uniform scaling with large gains on MATH-500 and several-fold improvements on AIME24 and AMO-Bench by dynamically selecting strategies and pruning low-reward paths.