What is reward hacking in reasoning training?
Reward hacking occurs when a model earns a favorable training signal through a shortcut or evaluator weakness instead of accomplishing the intended task correctly.
Explore This Question
Search for additional explanations using these tools:
More FAQs in What role does reinforcement learning play in developing AI reasoning abilities?
No. Models can develop useful problem-solving behavior through several routes, including pre-training, supervised examples, and external tools. Reinforcement learning is one possible contribution.
Some problems in these domains have answers that can be checked automatically. Those checks provide useful feedback, although incomplete tests and flawed verifiers remain possible.
Still have questions?
Browse all our FAQs or reach out to our support team