🤖 Reward Hacking in 3D: A Robot Climbs the Proxy, Not the Goal
3D reward hacking demonstration built with Blender and Three.js: a robot and 139 other agents climb a proxy reward surface with a decoy peak above the true reward terrain, with presets, adjustable decoy strength, learning rate and noise, and four numerical checks.
⚙ Under the hood
3D reward hacking demonstration built with Blender and Three.js: a robot and 139 other agents climb a proxy reward surface with a decoy peak above the true reward terrain, with presets, adjustable decoy strength, learning rate and noise, and four numerical checks.
Reward hackingAI safetyGoodhart's lawReinforcement learningGradient ascent
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install