Toby Ord: RL needs ~10,000x compute to go from 20% to 80% on reasoning benchmarks

tobyordoxford · x · 2026-11-14

Toby Ord shares preliminary fitting observations on RL compute scaling: across several reasoning benchmarks, it typically takes about 10,000x more RL compute to move from 20% to 80% performance, implying a power-law exponent of 0.3 (odds ∝ RLcompute^0.3).

The accompanying chart plots AIME (2022/2023) Pass@1 during training over 4 orders of magnitude of compute (10^-4 to 10^0)
OpenAI o1 and o3 trajectories are overlaid, with o3 clearly above the trend line
A dashed reference slope marks the 20%-to-80%-in-10,000x relationship, which the empirical curve tracks closely

Ord cautions these are early, rough observations, but they offer a concrete quantitative yardstick for how fast RL scaling returns diminish.

Original post →

Toby Ord: RL needs ~10,000x compute to go from 20% to 80% on reasoning benchmarks

More from Research