🟦 Active Prompting

Last updated on October 3, 2024 by Valeriia Kuka

Information and Links

Technique	Institution	Date of Publication	Paper	Code
Active Prompting	The Hong Kong University of Science and Technology, University of Toronto, The University of Hong Kong, University of Illinois Urbana-Champaign	Feb 2023	Active Prompting with Chain-of-Thought for Large Language Models	Code

What is Active Prompting?

Active Prompting (or Active-Prompt)¹ is a technique for improving Chain-of-Thought (CoT) prompting performance by selectively human-annotating exemplars where the model shows the most uncertainty. This approach helps maximize the efficiency of human annotation efforts by focusing only on the most challenging questions for the model.

Active Prompting consists of four main steps:

1. Uncertainty Estimation:

• It prompts the model several times ($k$) with unlabeled questions using Chain-of-Thought Prompting with a few human-written chain-of-thoughts or Zero-Shot Chain-of-Thought (CoT) prompting without human-written chain-of-thoughts to generate possible answers with intermediate steps for a set of unlabeled questions.
• It calculates the uncertainty $u$ based on $k$ answers via a selected uncertainty metric.

2. Selection: The questions with the highest uncertainty are selected for human annotation.
3. Annotation: The selected questions from step 2 are manually annotated by humans to provide more accurate answers.
4. Inference: The newly annotated exemplars are used to improve the model’s performance in answering future questions.

Active Prompting saves significant human resources by reducing the need to annotate all training data. It outperforms other techniques such as Automatic Chain-of-Thought prompting, Random Chain-of-Thought prompting, and Self-Consistency on a range of reasoning tasks. Active Prompting research¹ is the first to show the benefits of selective question annotation in CoT prompting for solving complex reasoning tasks.

How to Use Active Prompting?

Let’s break down the Active Prompting process with an example. Assume you have a pool of $n$ unlabeled questions.

Step 1. Uncertainty Estimation

First, you prompt the model multiple times ($k$) for each unlabeled question using:

• A number of annotated examplars if you want to use CoT option
• "Think step-by-step" if you want to use Zero-Shot CoT option

Let’s say you choose the CoT option. You provide exemplars (Q1 and Q2), then ask your pool question (Q3). Repeat this process $k$ times for each question.

As a result you will get $k$ answers for each of your $n$ questions.

Next, you need to measure the uncertainty of the model for each question based on the $k$ answers it generates for a given question.

To do that, you select the uncertainty metric. An example metric could be disagreement:

You use the disagreement among $k$ generated answers for a given question from the pool. The disagreement calculates the unique answers in the predictions.

• You count the number of unique answers generated for a question, $h$
• Calculate the disagreement by dividing $uncertainty = h/k$

Step 2. Selection

Then, you select the questions with the highest uncertainty based on the metric. For simplicity, let's review just one example question from the set of the most uncertain questions.

Imagine you take disagreement as an useartainty metric. You prompt the model $k$ times and find that the below question consistently yields different LLM's outputs meaning the model is uncertain in the answer:

This is just one example while in reality there can be many of them.

Step 3. Annotation

You manually annotate the selected question to provide a clear, correct answer:

Step 4. Inference

This annotated question becomes an example for the model.

What are the Experimental Results for Active Prompting?

Active Prompting has demonstrated superior performance across several benchmarks, including arithmetic, commonsense, and symbolic reasoning tasks. It consistently outperforms traditional CoT and other baseline techniques, highlighting its effectiveness in enhancing LLM capabilities.

• Self-Consistency: Active-Prompt outperforms self-consistency across most tasks. Active-Prompt shows a 7.2% improvement over SC on arithmetic reasoning tasks. For commonsense and symbolic reasoning, Active-Prompt consistently outperforms SC across all tasks.
• Chain-of-Thought: Compared to CoT, Active-Prompt demonstrates significant improvements. For instance, Active-Prompt scores 83.4% on GSM8K, compared to 63.1% by CoT. Across other datasets (e.g., ASDiv, SVAMP, AQUA), Active-Prompt outperforms CoT by margins ranging from 1.0% to 15.4%, showing the robustness of the active selection method.
• Random Chain-of-Thought: Active-Prompt also shows consistent improvements over Random-CoT, with higher average performance across the datasets.
• Automatic Chain-of-Thought: Though Auto-CoT produces decent results, particularly with code-davinci-002, Active-Prompt surpasses it in every task.

Limitations of Active Prompting

Despite its advantages, Active Prompting has some limitations:

• Human Annotation Required: Some level of human involvement is needed to annotate the most uncertain questions.

• Choosing the right uncertainty metric matters: The way we measure uncertainty can impact performance, so we need to pick the right one based on the task at hand.

Conclusion

Active prompting really enhances how well large language models solve complex reasoning problems. By focusing on the questions the model is most uncertain about, we make the annotation process efficient and tailor it to boost the model's learning.

Footnotes

1. Diao, S., Wang, P., Lin, Y., Pan, R., Liu, X., & Zhang, T. (2024). Active Prompting with Chain-of-Thought for Large Language Models. https://arxiv.org/abs/2302.12246 ↩ ↩²