Our new X account is live! Follow @wizwand_team for updates
WorkDL logo mark

KTO: Model Alignment as Prospect Theoretic Optimization

About

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.

Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, Douwe Kiela• 2024

Related benchmarks

TaskDatasetResultRank
Multi-turn Dialogue EvaluationMT-Bench
Overall Score7
331
Instruction FollowingAlpacaEval 2.0
LC Win Rate33.1
281
Mathematical ReasoningMathQA
Accuracy75.8
95
LLM Alignment EvaluationAlpacaEval 2
LC Win Rate43.86
72
LLM Alignment EvaluationArena Hard
Win Rate26.8
67
Instruction Following and Helpfulness EvaluationAlpacaEval 2.0
Win Rate10
58
AlpacaEval 2.0UltraFeedback
LC18.8
42
MT-BenchUltraFeedback
MT-Bench Score8
42
Safety Alignment EvaluationLlama-Guard
Harmfulness (%)83.42
36
Language Model Alignment EvaluationArena Hard v0.1
Win Rate (%)30.5
36
Showing 10 of 31 rows

Other info

Follow for update