[2402.01306] KTO: Model Alignment as Prospect Theoretic Optimization

arxiv.org · 7,025 words · saved by 1 readers

Abstract:Kahneman & Tversky's $\textit{prospect theory}$ tells us that humans perceive random variables in a biased but well-defined manner (1992); for example, humans are famously loss-averse. We show that objectives for aligning LLMs with human feedback implicitly incorporate many of these biases -- the success of these objectives (e.g., DPO) over cross-entropy minimization can partly be ascribed to them belonging to a family of loss functions that we call $\textit{human-aware losses}$ (HALOs). However, the utility functions these methods attribute to humans still differ from those in the prospect theory literature. Using a Kahneman-Tversky model of human utility, we propose a HALO that directly maximizes the utility of generations instead of maximizing the log-likelihood of preferences, as current methods do. We call this approach KTO, and it matches or exceeds the performance of preference-based methods at scales from 1B to 30B, despite only learning from a binary signal of whether an output is desirable. More broadly, our work suggests that there is no one HALO that is universally superior; the best loss depends on the inductive biases most appropriate for a given setting, an oft-overlooked consideration.

KTO: Model Alignment as Prospect Theoretic Optimization Kawin Ethayarajh 1 Winnie Xu 2 Niklas Muennighoff 2 Dan Jurafsky 1 Douwe Kiela 1 2 Abstract Kahneman & Tversky’s prospect theory tells us that humans perceive random variables in a biased but well-defined manner (1992); for example, hu- arXiv:2402.01306v5 [cs.LG] 8 Sep 2026…

[2402.01306] KTO: Model Alignment as Prospect Theoretic Optimization

related reading