Traditional methods like DPO require costly paired preference data, whereas KTO (Kahneman-Tversky Optimization) needs only a binary signal — whether an output is "desirable" — to optimize the model. The paper first introduces the concept of Human-Aware Losses (HALOs), showing that successful methods like DPO work because their loss functions implicitly encode cognitive biases from human decision-making (such as loss aversion). KTO goes further by directly designing a loss function that maximizes human-perceived utility, grounded in the Kahneman-Tversky model of human utility. Experiments show that at scales from 1B to 30B parameters, KTO matches or even surpasses DPO's performance using only binary feedback, handles extreme data imbalance, and can even skip the supervised fine-tuning step for sufficiently strong pretrained models. The paper's ultimate conclusion is that no universally superior alignment loss function exists — the best choice depends on the inductive biases suited to the specific task and data environment — and KTO offers an efficient and powerful alignment scheme for feedback data that is cheaper and simpler to obtain.