What's this about?
RLVR gives one sparse reward per response; OPD gives dense per-token feedback but caps you at teacher quality. SAF-OPD fuses both signals stably — dense guidance without the ceiling (24↑).
What exactly is RLVR?
Reinforcement learning with verifiable rewards: train on tasks where correctness is checkable (math answers, passing tests), so the reward can't be gamed. It's the engine of the current reasoning-model wave — this week's OpenAI math results are its poster child.
Why does it matter?
Three of the week's top papers are OPD variants — post-training recipes, not architectures, are where the competitive iteration is happening right now.