Mingkai Deng

Figure 1: Overview of RL Prompt for discrete prompt optimization. All language models (LMs) are frozen. We build our policy network by training a task-specific multi-layer perceptron (MLP) network inserted into a frozen pre-trained LM. The figure above illustrates 1) generation of a prompt (left), 2) example usages in a masked LM for classification (top right) and a left-to-right LM for generation (bottom right), and 3) update of the MLP using RL reward signals (red arrows).

Figure 1: Our framework classifies language generation tasks into compression, transduction, and creation (left), and unifies the evaluation (middle) of key quality aspects with the common operation of information alignment (right). By Mingkai Deng TL;DR: Evaluating natural language generation (NLG) is hard. Our general framework helps solve the difficulty by unifying the evaluation with a common central operation.