Gibbs Distribution

توزيع غيبس

توزيع احتمالي يتناسب مع exp(المكافأة/الحرارة)، مُستمد من الميكانيكا الإحصائية. السياسة المُثلى في RLHF المُقيَّد بتباعد KL تأخذ هذا الشكل بالضبط.

A probability distribution proportional to exp(reward/temperature), from statistical mechanics. The optimal policy in KL-constrained RLHF takes exactly this form.

Also translated asتوزيع بولتزمان، التوزيع الأسّي المُقيَّس

First appears in this corpus in: Stochastic Relaxation, Gibbs Distributions, and the Bayesian Restoration of Images (1984)

Appears in these papers

Direct Preference Optimization: Your Language Model Is Secretly a Reward Model2023 in the sky ✦
Direct Preference Optimization: Your Language Model Is Secretly a Reward Model2023 in the sky ✦
Stochastic Relaxation, Gibbs Distributions, and the Bayesian Restoration of Images1984 in the sky ✦