From RLHF to RLVR: The Evolution of Reward Signals and the Battle Against Reward Hacking

From human preferences and LLM judges to verifiable rewards: why soft evaluators get reward-hacked and how to build tamper-proof verifiers.