Reinforcement Learning
RL Beats SFT on Rule Shifts, but Training Cost Remains a Hard Constraint
In a “24-point” experiment, RL improved accuracy by 3.5% and 17.6% after rule changes, while SFT declined, exposing the limits of imitation.
1 article tagged with this topic
In a “24-point” experiment, RL improved accuracy by 3.5% and 17.6% after rule changes, while SFT declined, exposing the limits of imitation.