Google

The prediction of a Large Language Model, denoted as $$\hat{y}_{\sigma}$$, when using a soft prompt $$\sigma$$ (a compressed context) and an input $$z$$, is determined by selecting the output $$y$$ that maximizes the conditional probability. This is formally expressed as: $$\hat{y}_{\sigma} = \underset{y}{\arg\max}\, \text{Pr}(y|\sigma, z)$$ This prediction is compared against the prediction from the full context to optimize the soft prompt.

Formula for LLM Prediction with Compressed Context

The optimal soft prompt, denoted as $$\hat{\sigma}$$, is determined by minimizing a function $$s(\cdot, \cdot)$$ that compares the prediction from the full context, $$\hat{y}$$, with the prediction from the compressed context, $$\hat{y}_{\sigma}$$. This function typically represents a loss or dissimilarity measure. The optimization problem is formally expressed as: $$\hat{\sigma} = \underset{\sigma}{\arg\min}\, s(\hat{y}, \hat{y}_{\sigma})$$

Formula for Optimizing Soft Prompts via Context Compression

An alternative approach to optimizing soft prompts involves minimizing the Kullback-Leibler (KL) divergence between the output probability distribution from the full context, $$\text{Pr}(\cdot|\mathbf{c}, \mathbf{z})$$, and the distribution from the soft prompt, $$\text{Pr}(\cdot|\sigma, \mathbf{z})$$. The goal is to find the soft prompt $$\hat{\sigma}$$ that makes these two distributions as similar as possible. The optimization is expressed by the formula: $$\hat{\sigma} = \underset{\sigma}{\arg\min}\, \text{KL}(\text{Pr}(\cdot|\mathbf{c}, \mathbf{z}) \|\| \text{Pr}(\cdot|\sigma, \mathbf{z}))$$

Formula for Soft Prompt Optimization by Minimizing KL Divergence

A large language model uses the following formula to make a prediction when given a compressed summary of a document (σ) and a specific user query (z):

$$\hat{y}_{\sigma} = \underset{y}{\arg\max}\, \text{Pr}(y|\sigma, z)$$

Break down this formula by explaining what each of the following four components represents in this specific scenario:
1.  `σ`
2.  `z`
3.  `y`
4.  `ŷ_σ`

Deconstructing the LLM Prediction Formula

A large language model is given a compressed context representation, denoted as `σ`, and an input query, denoted as `z`. Based on the probability distribution calculated by the model (shown in the case study), identify the model's final prediction and explain why it produced this specific output, even if it is factually incorrect. Your explanation must be grounded in the mathematical principle the model uses for prediction.

Analyzing an LLM's Incorrect Prediction

An LLM is provided with a compressed representation of context, denoted as `σ`, and an input `z`. The model's goal is to predict the most likely output `y`. After processing `σ` and `z`, the model computes the following conditional probabilities for four possible outputs:

- Pr(y='mat' | σ, z) = 0.65
- Pr(y='roof' | σ, z) = 0.25
- Pr(y='sky' | σ, z) = 0.05
- Pr(y='idea' | σ, z) = 0.05

Based on the principle of selecting the output that maximizes the conditional probability, what will the model'

Learn Before

Related