Visual Representation of Iterative Soft Prompt Refinement
The iterative refinement of soft prompts can be visualized as a multi-step process. In each step, the soft prompts from the previous iteration, denoted as , are processed by the model's Transformer layers along with the standard input embeddings () and other vectors (). The resulting hidden states () from these layers are then used to compute an updated and more refined set of soft prompts, , for the subsequent step.
0
1
Tags
Ch.4 Alignment - Foundations of Large Language Models
Foundations of Large Language Models
Computing Sciences
Foundations of Large Language Models Course
Related
Visual Representation of Iterative Soft Prompt Refinement
Consider a system where a set of learnable, non-textual prompt vectors are processed by a model's main computational layers alongside the standard input. After this initial processing pass, the resulting internal representations are used to calculate an updated set of prompt vectors, which then replace the original ones for a subsequent processing pass. What is the primary function of the model's main computational layers within this cyclical process?
A specific prompting technique uses a model's own computational layers to progressively improve a set of learnable, non-textual prompt vectors in a feedback loop. Arrange the following events to correctly describe a single cycle of this refinement process.
Analyzing a Flawed Prompt Optimization Implementation
Learn After
A technique for improving a model's performance involves iteratively refining a set of special input vectors. The following items describe the stages within a single step of this refinement process. Arrange these stages in the correct logical order, from the initial inputs to the final output of one complete step.
A diagram illustrates a single step in an iterative process for enhancing a model's input vectors. In this step, the vectors from the previous iteration (denoted as σᵢ) are combined with standard input embeddings (eᵢ) and processed by the model's main layers to produce hidden states (hᵢ). These hidden states are then used to calculate the enhanced vectors for the next iteration (σᵢ₊₁). Which part of this flow most directly embodies the 'refinement' that occurs in each step?
Component Relationships in an Iterative Refinement Diagram