The pass@k price

RLVR raises pass@1 while the base model overtakes it at large k. Paste your own base and RL numbers and read off the number of crossings, the Karlin sign-variation bound, the regime, and the price (the worst pass@k loss). Runs entirely in your browser.

Paper (arXiv:2610.39295)Code

k values base pass@k RL pass@k

Same number of values in each box. Commas or spaces.

base per-prompt success RL per-prompt success

One probability per prompt, same order in both boxes. Shorthand 0.4x700 repeats a value 700 times.

gain at smallest k (RL minus base)

price: worst pass@k loss

pass@k crossings

Karlin bound S^-(G)