Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
More inference compute only helps when it is spent in the right shape. Follow a difficulty-aware router that chooses between independent samples, sequential revisions, and process-reward search, then read the fourfold efficiency result and the boundary on the hardest problems.
Reading focus: Why model-specific difficulty predicts whether extra compute should buy fresh attempts, revisions of one route, or verifier-guided search. How the paper's routed revision policy beats 256 parallel samples with 64 samples in its displayed MATH comparison. Why the comparison with a roughly fourteen-times-larger model depends on problem difficulty, workload volume, proprietary checkpoints, and an uncharged routing cost.
ICLR 2025. Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 9 min read, medium difficulty.