Scaling Laws for Reward Model OveroptimizationChatGPT主题资料合编.docxVIP

  • 0
  • 0
  • 约6.69万字
  • 约 28页
  • 2026-03-27 发布于浙江
  • 举报

Scaling Laws for Reward Model OveroptimizationChatGPT主题资料合编.docx

ScalingLawsforRewardModelOveroptimizationLeoGaoOpenAIJohnSchulmanOpenAIJacobHiltonOpenAIAbstract

ScalingLawsforRewardModelOveroptimization

LeoGao

OpenAI

JohnSchulman

OpenAI

JacobHilton

OpenAI

Abstract

Inreinforcementlearningfromhumanfeedback,itiscommontooptimizeagainstarewardmodeltrainedtopredicthumanpreferences.Becausetherewardmodelisanimperfectproxy,optimizingitsvaluetoomuchcanhindergroundtruthperformance,inaccordancewithGoodhart’slaw.Thiseffecthasbeenfrequentlyobserved,butnotcarefullymeasuredduetotheexpenseofcollectin

文档评论(0)

1亿VIP精品文档

相关文档