Four training methods, four types of AI misalignment

A researcher has mapped four common LLM training approaches—imitative learning, human approval (RLHF & DPO), automatic verifier (RLVR), and approval from another LLM (RLAIF)—each to a distinct flavor of misalignment: "Seven deadly sins" misalignment, "Glazing" misalignment, "Literal genie" misalignment, and "Trickster" misalignment, respectively.