a
axiom
console
Overview
Demo
Leaderboard
GitHub ↗
Reward Robustness Leaderboard
How hard is each reward to cheat? Lower
hack_rate
(an agent fools it less often) and higher
honest_fidelity
(real work still gets paid) combine into
RRS
—
higher = harder to cheat
. ·
RRS = (1 − hack_rate) × honest_fidelity