EasyDeepLearn

How does the KL penalty interact with reward hacking?

hard

Answer

  • KL to πref{\pi}_{\mathrm{ref}} bounds how far model can drift → limits reward hacking of imperfect r_φ.
  • But: (1) large model can find hacks within KL budget (spurious  features  that  both  rφ  likes  AND  πref  probably  would  generate)(\mathrm{spurious}\;\mathrm{features}\;\mathrm{that}\;\mathrm{both}\;r{\varphi}\;\mathrm{likes}\;\mathrm{AND}\;{\pi}_{\mathrm{ref}}\;\mathrm{probably}\;\mathrm{would}\;\mathrm{generate}).
  • (2) β must be tuned per task — too high stalls learning.
  • (3) Better: iterative RM updates (relabel new-policy responses), adversarial reward shaping.
  • (4) Modern practice: KL as safety floor, not primary defense against hacking.
Check yourself — multiple choice
  • Fixes hacking
  • KL bounds drift → limits hacking but doesn't prevent hacks within KL budget; β must be tuned; iterative RM + adversarial + KL together — KL is safety floor, not primary defense
  • Random
  • Not real

KL: bounds drift, limits but doesn't prevent hacking within budget.

#llm#alignment#safety

Practise Reinforcement Learning

214 interview questions in this topic.

Related questions