Chain of Thought: Verification and Density Lead to Smarter LLM response

LLMs are capable of generating human-quality text, code, and other creative content. However, they are also prone to hallucinations, or the generation of plausible but incorrect factual information. This is a major challenge for the widespread use of LLMs in many applications.

Two promising new techniques, Chain-of-Verification (CoVe) and Chain-of-Density (CoD), are poised to address some of the key challenges facing LLMs. CoVe is a structured process that helps LLMs to double-check their own work, ensuring that their responses are accurate and reliable. CoD, on the other hand, is a technique for controlling the density of information in generated text, ensuring that summaries are neither too sparse nor too dense.

Before delving into the solutions provided by CoVe and CoD, it's crucial to understand the challenges that LLMs currently face.

• Hallucinations: One of the foremost challenges is the generation of plausible yet incorrect factual information, often referred to as "hallucinations." LLMs have a tendency to generate information that sounds reasonable but is, in fact, inaccurate. This issue poses a significant obstacle in achieving reliable AI-generated text.
• Information Density: In tasks such as text summarization, determining the right balance of information density is a challenge. Summaries that are too sparse may lack essential details, while overly dense summaries can be overwhelming. Striking the right balance is vital for comprehension and usability.
• Task-Specific Quality: LLMs may excel in some tasks but fall short in others. Achieving consistent quality across diverse tasks remains a challenge. Task-specific nuances often require tailored approaches.

CoVe is a four-step process that helps LLMs to generate more reliable responses:

1. Drafting responses: The LLM generates an initial response to the prompt.
2. Planning verification questions: The LLM identifies key factual claims in its response and generates a set of verification questions.
3. Answering them independently: The LLM answers the verification questions independently, using its knowledge base and external sources.
4. Generating a final verified response: The LLM generates a final response based on the initial response and the answers to the verification questions.

The COD Process is a step-by-step approach to generating summaries with a desired level of information density:

1. Identify the Text for Summarization: Choose the document, article, or any piece of text that you wish to summarize.
2. Craft the Initial Prompt: Create an initial summarization prompt tailored to the selected text. The aim here is to guide the LLM towards generating a basic summary.
3. Analyze the Initial Summary: Review the summary generated from the initial prompt. Identify if the summary is too sparse (missing key details) or too dense (containing unnecessary details).
4. Design Chained Prompts: Based on the initial summary's density, construct additional prompts to adjust the level of detail in the summary. These are the "chained prompts" and are central to the Chain of Density technique.
5. Execute Chained Prompts: Feed these chained prompts back to the LLM. These prompts are designed to either increase the density by adding essential details or decrease it by removing non-essential information.
6. Review the Adjusted Summary: Examine the new summary generated by executing the chained prompts. Ensure that it captures all essential points while avoiding unnecessary details.
7. Iterate if Necessary: If the summary still doesn't meet the desired criteria for information density, return to step 4 and adjust the chained prompts accordingly.
8. Finalize the Summary: Once the summary meets the desired level of information density, it is considered finalised and ready for use.

Example

Let's start our exploration with a practical example from the field of banking. Imagine a customer walks into a bank or accesses their online services and poses a common query: "Can you tell me the process for applying for a credit card?" This seemingly straightforward question highlights some of the complexities and challenges that LLMs must overcome in delivering accurate, reliable, and appropriately dense responses.

Customer Query: "Can you tell me the process for applying for a credit card?"

Without Using CoVe or CoD:

Response: "To apply for a credit card, you can visit our website or one of our local branches. The eligibility criteria include a minimum age of 21, a stable source of income, and a good credit score. You will need to fill out an application form and provide necessary documents. After submission, our team will review your application, and if approved, you will receive your credit card in 7-10 business days."

Using CoVe:

CoVe Prompt:

"Explain the process for applying for a credit card."

• Initial Draft: "To apply for a credit card, you can visit our website or a local branch. You should be at least 21 years old, have a steady income, and a good credit score. Fill out an application and provide required documents. Our team reviews the application, and if approved, you'll receive the card in 7-10 business days."
• Verification Questions:> Was the minimum age requirement stated accurately?

>Is the requirement for a steady income mentioned?

>Is there information about required documents?

While accuracy is a crucial aspect, the challenge of information density cannot be overlooked. In tasks like text summarisation, striking the right balance between being too sparse and overly dense is essential for comprehensibility and usability. This is where CoD comes into play.

CoD Prompt:

You will generate increasingly concise, entity-dense summaries of the provided information.

Repeat the following 2 steps 5 times.

Step 1. Identify 1-3 informative Entities ("; " delimited) from the Initial Summary which are missing from the previously generated summary. Step 2. Write a new, denser summary of identical length which covers every entity and detail from the previous summary plus the Missing Entities.

A Missing Entity is:

• Relevant: to the credit card application process.
• Specific: descriptive yet concise (5 words or fewer).
• Novel: not in the previous summary.
• Faithful: present in the Initial Summary.
• Anywhere: located anywhere in the Initial Summary.

Guidelines:

• The first summary should be long (4-5 sentences, ~80 words) yet highly non-specific, containing little information beyond the entities marked as missing. Use overly verbose language and fillers (e.g., "the application process involves") to reach ~80 words.
• Make every word count: rewrite the previous summary to improve flow and make space for additional entities.
• Make space with fusion, compression, and removal of uninformative phrases like "you must be 21 or older".
• The summaries should become highly dense and concise yet self-contained, e.g., easily understood without the Initial Summary.
• Missing entities can appear anywhere in the new summary.
• Never drop entities from the previous summary. If space cannot be made, add fewer new entities.

Remember, use the exact same number of words for each summary.

By combining CoVe and CoD, we can create LLMs that are able to:

• Generate accurate and reliable responses, even to complex and challenging queries.
• Produce informative and concise summaries of long texts, tailored to the specific needs of the user.

This combination of techniques has the potential to transform the way we interact with computers and information.

For example, imagine a future where LLMs are used to generate personalized news summaries, translate languages in real time, and even write creative content such as poems and stories. With CoVe and CoD, these LLMs could be trusted to provide us with accurate and reliable information, tailored to our individual needs and interests.

The journey of LLMs is far from over, and the future holds exciting possibilities:

• Efficiency and Accessibility

Researchers are actively exploring ways to make LLMs more efficient and less computationally expensive. This effort aims to democratize access to LLM technology and reduce the barriers to entry for innovation.

• Ethical AI

Ethical considerations will continue to be a central focus. LLMs will undergo rigorous training and fine-tuning with a strong emphasis on ethical principles, reducing biases, promoting fairness, and ensuring responsible AI usage.

• Task-Specific Quality

Future advancements will lead to LLMs that excel in a wide range of tasks. They will adapt seamlessly to new domains and demonstrate consistent quality in diverse applications.

In the meantime, by using CoVe and CoD, we can ensure that LLMs generate outputs that are truthful, coherent, respectful, diverse, and novel.