Advanced Prompt Engineering

Large language models still struggle with complex, multi-step reasoning tasks. Problems like math word problems or commonsense reasoning remain challenging for AI.

To address this limitation, researchers developed a novel technique called chain of thought prompting. Wei et al (2022).

This method provides a way to enhance the reasoning capabilities of large language models like GPT-3.

How Chain of Thought Prompting Works

Chain of thought prompting guides the language model through a series of logical, intermediate steps when solving a complex problem.

Source: Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022). Chain of Thought Prompting Elicits Reasoning in Large Language Models.

Here's another example of a math word problem:

"John had 35 marbles. He gave 8 marbles to Anna and 14 marbles to Tom. How many marbles does John have left?"

With standard prompting, you would provide the model with some input-output examples, and then ask it to solve the problem directly.

Chain of thought prompting works differently. Instead of jumping straight to the solution, it leads the model through reasoning steps:

• John originally had 35 marbles
• He gave 8 marbles to Anna
• So he now has 35 – 8 = 27 marbles
• He gave 14 marbles to Tom
• So he now has 27 – 14 = 13 marbles left

By structuring the prompt to demonstrate this logical progression, chain of thought prompting mimics the way humans break down problems step-by-step. The model learns to follow a similar reasoning process.

Why It Improves Reasoning

There are several key benefits to the chain of thought approach:

• It divides complex problems into smaller, more manageable parts. This allows the model to focus its vast computational resources on each sub-task.
• The intermediate steps provide interpretability into the model's reasoning process. This transparency makes it easier to evaluate the model's logic.
• Chain of thought prompting is versatile. It can enhance reasoning across diverse tasks like math, common sense, and symbol manipulation.
• The step-by-step structure improves learning efficiency. Models can grasp concepts more effectively when presented in a logical progression.

Research shows chain of thought prompting boosts performance on tasks requiring complex reasoning.

When It Works Best

• Chain of thought prompting only yields significant gains when used with extremely large models, typically those with over 100 billion parameters. The approach relies on the model having enough knowledge and processing power to successfully follow the reasoning steps.
• Smaller models often fail to generate logical chains of thought, so chain of thought prompting does not improve their performance. The benefits appear to scale proportionally with model size.
• In addition, the technique is best suited to problems with clear intermediate steps and language-based solutions. Tasks like mathematical reasoning lend themselves well to step-by-step reasoning prompts.

Chain of thought prompting offers an intriguing method to enhance reasoning in large AI models. Guiding the model to decompose problems into logical steps seems to unlock capabilities not accessible through standard prompting alone.

While not a universal solution, chain of thought prompting demonstrates how tailored prompting techniques can stretch the abilities of language models. As models continue to grow in scale, prompting methods like this will likely play an integral role in realising the robust reasoning skills required for advanced AI.

Further Reading

• Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022). Chain of Thought Prompting Elicits Reasoning in Large Language Models.
• Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H. W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A., Barnes, P., Tay, Y., Shazeer, N., Prabhakaran, V., … Fiedel, N. (2022). PaLM: Scaling Language Modeling with Pathways.
• Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., & Schulman, J. (2021). Training Verifiers to Solve Math Word Problems.

Large Language models (LLMs) like GPT-3/GPT-4/Claude-3 and others have exhibited astonishing capabilities across various domains, from mathematical problem-solving to creative writing. However, there's been an inherent limitation in their approach – the left-to-right, token-by-token decision-making process, which doesn't always align with complex problem-solving scenarios that demand strategic planning and exploration.

But what if we could enable these LLMs to think more strategically, explore multiple reasoning paths, and evaluate the quality of their thoughts in a deliberate manner? Some researchers have created a framework called "Tree of Thoughts" (ToT) which aims to fix this by enhancing the problem-solving prowess of large language models.

The Essence of ToT

At its core, ToT reimagines the reasoning process as an intricate tree structure. Each branch of this tree represents an intermediate "thought" or a coherent chunk of text that serves as a crucial step toward reaching a solution. Think of it as a roadmap where each stop is a meaningful milestone in the journey towards problem resolution. For instance, in mathematical problem-solving, these thoughts could correspond to equations or strategies.

But ToT doesn't stop there. It actively encourages the LM to generate multiple possible thoughts at each juncture, rather than sticking to a single sequential thought generation process, as seen in traditional chain-of-thought prompting. This flexibility allows the model to explore diverse reasoning paths and consider various options simultaneously.

Image Source: Yao et el. (2023)

The Power of Self-Evaluation: One of ToT's defining features is the model's ability to evaluate its own thoughts. It's like having an inbuilt compass to assess the validity or likelihood of success for each thought. This self-evaluation provides a heuristic, a kind of mental scorecard, to guide the LM through its decision-making process. It helps the model distinguish between promising paths and those that may lead to dead ends.

Systematic Exploration: ToT takes strategic thinking up a notch by employing classic search algorithms such as breadth-first search or depth-first search to systematically explore the tree of thoughts. These algorithms allow the model to look ahead, backtrack when necessary, and branch out to consider different possibilities. It's akin to a chess player contemplating multiple moves ahead before making a move.

Customisable and Adaptable: One of ToT's strengths is its modularity. Every component, from thought representation to generation, evaluation, and search algorithm, can be customized to fit the specific problem at hand. No additional model training is needed, making it highly adaptable to various tasks.

Real-World Applications: The true litmus test for any AI framework is its practical applications. ToT has been put to the test across different challenges, including the Game of 24, Creative Writing, and Mini Crosswords. In each case, ToT significantly boosted the problem-solving capabilities of LLMs over standard prompting methods. For instance, in the Game of 24, success rates soared from a mere 4% with chain-of-thought prompting to an impressive 74% with ToT.

Image Source: Yao et el. (2023)

The above image is a visual representation of the Game of 24 which is a mathematical reasoning challenge where the goal is to use 4 input numbers and arithmetic operations to reach the target number 24.

The tree of thought (ToT) approach represents this as a search over possible intermediate equation "thoughts" that progressively simplify towards the final solution.

First, the language model proposes candidate thoughts that manipulate the inputs (e.g. (10 – 4)).

Next, it evaluates the promise of reaching 24 from each partial equation by estimating how close the current result is. Thoughts evaluated as impossible are pruned.

The process repeats, generating new thoughts conditioned on the remaining options, evaluating them, and pruning. This iterative search through the space of possible equations allows systematic reasoning.

For example, the model might first try (10 – 4), then build on this by proposing (6 x 13 – 9) which gets closer to 24. After several rounds of generation and evaluation, it finally produces a complete solution path like: (10 – 4) x (13 – 9) = 24.

By deliberating over multiple possible chains of reasoning, ToT allows more structured problem solving compared to solely prompting for the end solution.

A Glimpse into the Future

As we delve deeper into the era of AI-driven decision-making, the ToT framework represents a pivotal development. It bridges the gap between symbolic planning and modern LLMs, offering the promise of more human-like planning and metacognition. This opens exciting possibilities for better aligning AI with human intentions and understanding.

Conclusion

In conclusion, the Tree of Thoughts (ToT) framework is a beacon of light in the AI landscape. It introduces a level of strategic thinking and exploration that was previously lacking in language models. By allowing LLMs to consider multiple reasoning paths, evaluate their own choices, and systematically explore complex problems, ToT paves the way for more intelligent, adaptable, and effective AI systems. The journey has just begun, and the potential for ToT to reshape the future of AI is boundless.

ChatGPT and other large language models have shown impressive capabilities, but complex reasoning remains a weak spot. However, a study revealed an effective technique to enhance reasoning – using diverse prompts.

Researchers from Microsoft and Stanford tested methods to elicit more diverse and structured thinking from models like GPT-3 and GPT-4. The key idea is prompting the model itself to suggest various approaches and personas for solving reasoning problems.

For example, when faced with a math word problem, GPT-4 can propose trying direct calculation, drawing a working backwards, and much more. These diverse strategies are then incorporated into multiple rephrased prompts.

The researchers introduced two techniques building on this idea:

DIV-SE: Execute each diverse prompt separately and combine the responses.
IDIV-SE: Combine multiple approaches into a single prompt.
In this course we are going to concentrate on IDIV-SE "(In-call DIVerse reasoning path Self-Ensemble)"

Image Source: Naik, R., Chandrasekaran, V., Yuksekgonul, M., Palangi, H., & Nushi, B. (2023). Diversity of thought improves reasoning abilities of large language models. arXiv preprint arXiv:2310.07088.

Across benchmarks in math, planning, and commonsense reasoning, both DIV-SE and IDIV-SE improved accuracy and cost-effectiveness substantially compared to prior prompting strategies.

On a difficult 4/5 blocks world planning challenge, DIV-SE boosted GPT-4's accuracy by 29.6 percentage points. For grade school math problems, it increased GPT-3.5's performance by over 10 percentage points.

Unlike other methods that modify the decoding process, diverse prompting works by eliciting diversity at the input level. This makes it broadly applicable even to black-box models.

In Summary:

• Prompting the model for diverse problem-solving approaches is an effective strategy to improve reasoning.
• Combining these diverse prompts boosts accuracy and cost-effectiveness.
• DIV-SE and IDIV-SE outperformed existing prompting techniques substantially.
• The methods provide gains without needing access to model internals.
• Diversity at the prompt level complements diversity during decoding.
• Planning, math and commonsense reasoning saw large improvements.
• Eliciting diversity directly from the model itself was critical.

The striking gains show the power of diversity for reasoning. While not flawless, diverse prompting pushes ChatGPT notably forward on its journey toward robust reasoning.

Key Takeaways:

• Get GPT's feedback on potential approaches and personas to solve the reasoning problem
• Create demonstrations of solving the problem using different approaches
• Prompt GPT to solve the problem taking on each persona and using the approaches
• Aggregate the solutions from different personas and approaches
• Diversity of approaches and "thinkers" is key to improving reasoning

Here's a prompt template which embodies the Diverse of Thought (DoT) approach:

IDIV-SE ( Diverse Reasoning)

[State reasoning problem here for example: In the following question, a number series is given with one term missing. Choose the correct alternative that will follow the same pattern and fill in the blank spaces. 1, 2, 3, 5, x, 13]

To begin, please suggest 3 distinct approaches I could use to accurately solve the above problem:

Approach 1:
Approach 2:
Approach 3:
Now please provide 3 short demonstrations, each solving the original problem using one of the approaches you suggested above:

Demonstration 1 (Approach 1):

Demonstration 2 (Approach 2):

Demonstration 3 (Approach 3):

Great, let's put it all together. Please now take on the role of expert one (a persona you feel is mostly aligned to the issue) and solve the original problem using Approaches 1-3.

Now take on the persona of expert 2 (a persona you feel is the next most likely aligned to the issue) and solve the original problem again using Approaches 1-3.

Finally, take on the persona of expert 3 (a persona you feel is the next most likely aligned to the issue) and solve the original problem a third time using Approaches 1-3.

Please synthesise your responses from the 3 expert personas above and provide your final recommended solution.

Recent advances in AI summarisation are largely thanks to the rise of large language models (LLMs) like GPT-3 and GPT-4. Rather than training on labeled datasets, these models can generate summaries with just the right prompts. This allows for precise control over summary length, topics covered, and style. An important but overlooked aspect is information density – how much detail to include within a constrained length. The goal is a summary that is informative yet clear. Striking this balance is challenging.

A new technique called Chain of Density (CoD) prompting helps address this tradeoff. Recently published research explains the approach and provides insights based on human evaluation.

Overview of Chain of Density Prompting:

The CoD method works by incrementally increasing the entity density of GPT-4 summaries without changing length. First, GPT-4 generates an initial sparse summary focused on just 1-3 entities. Then over several iterations, it identifies missing salient entities from the source text and fuses them into the summary.

Source: Griffin Adams, Alexander Fabbri, Faisal Ladhak, Eric Lehman, Noémie Elhadad (2023). From Sparse to Dense: GPT-4 Summarisation with Chain of Density Prompting.

To make room, GPT-4 is prompted to abstract, compress content, and merge entities. Each resulting summary contains more entities per token than the last. The researchers generate 5 rounds of densification for 100 CNN/Daily Mail articles.

Key Points:

• Humans preferred CoD summaries with densities close to human-written ones over sparse GPT-4 summaries from vanilla prompts.
• CoD summaries became more abstract, fused content more, and reduced bias toward early text over iterations.
• There was a peak density beyond which coherence declined due to awkward fusions of entities.
• An entity density of ~0.15 was ideal, vs 0.122 for vanilla GPT-4 and 0.151 for human summaries.

Conclusion:

This study and approach highlights the importance of achieving the right level of density in automated summarisation. Neither overly sparse nor dense summaries are optimal. The CoD technique paired with human evaluation offers a promising path toward readable yet informative AI-generated summaries.

Key takeaways:

1. Ask for multiple summaries of increasing detail. Start with a short 1-2 sentence summary, then ask for a slightly more detailed version, and keep iterating until you get the right balance of conciseness and completeness for your needs.
2. When asking ChatGPT to summarise something lengthy like an article or report, specify that you want an "informative yet readable" summary. This signals the ideal density based on the research.
3. Pay attention to awkward phrasing, strange entity combinations, or unconnected facts when reading AI summaries. These are signs it may be too dense and compressed. Request a less dense version.
4. For complex topics, don't expect chatbots to convey every detail in a highly compressed summary – there are limits before coherence suffers. Ask for a slightly longer summary if needed.
5. Remember that for optimal clarity and usefulness, AI summaries should have a similar density to those written by humans. Extreme brevity may mean missing key details.

The core takeaway is that density impacts the quality and usefulness of AI summarisation. As an end user, being aware of this can help you prompt for and identify the "goldilocks" level of density for your needs, avoiding summaries that are either frustratingly vague or confusingly overloaded. The Chain of Density research provides insights to guide this process.

Conditional reasoning is a fundamental aspect of intelligence, both in humans and artificial intelligence systems. It's the process of making decisions or drawing conclusions based on specific conditions or premises. In our daily lives, we often use conditional reasoning without even realising it. For example, deciding whether to take an umbrella depends on the condition of the weather forecast. Similarly, artificial intelligence (AI), particularly large language models (LLMs), also attempt to mimic this essential human ability.

While LLMs like GPT-3.5 have demonstrated remarkable capabilities in various natural language processing tasks, their prowess in conditional reasoning has been somewhat limited and less explored. This is where a new and innovative approach known as "code prompting" comes into play, to enhance conditional reasoning in LLMs trained on both text and code.

The Concept of Code Prompting

Code prompting converts natural language descriptions into code to be solved with a large language model. The figure shows a transformed instance from the ConditionalQA dataset.

A diagram showcasing how code prompting works compared to text based prompting
Image Source: Puerto, H., Tutek, M., Aditya, S., Zhu, X., & Gurevych, I. (2024). Code Prompting Elicits Conditional Reasoning Abilities in Text+Code LLMs. arXiv preprint arXiv:2401.10065.

Code prompting is an intriguing technique where a natural language problem is transformed into code before it's presented to the LLM. This code isn't just a jumble of commands and syntax; it thoughtfully retains the original text as comments, essentially embedding the textual logic within the code's structure. This approach is revolutionary in how it leverages the strengths of LLMs trained on both text and code, potentially unlocking new levels of reasoning capabilities.

Example Scenario:
Question: You're planning a day at the beach and need to decide what items to bring based on the weather forecast.

Traditional Text-Based Prompt (Without Code Prompting):
"Based on the following weather forecast for tomorrow, suggest what items should be brought for a day at the beach: Sunny in the morning, with a 70% chance of rain in the afternoon."

Code Prompting-Based Example (With Code Prompting):

# Weather forecast: Sunny in the morning, 70% chance of rain in the afternoon
# Task: Suggest items to bring for a day at the beach
if weather_forecast == "sunny morning and rainy afternoon":
    items_to_bring = ["sunscreen", "umbrella", "towel", "raincoat"]
    print("Based on the weather forecast, you should bring: ", items_to_bring)
        

Traditional Text-Based Prompt Output:
The model might suggest bringing typical beach items such as sunscreen and towels, but it might overlook the change in weather in the afternoon, like the need for an umbrella or a raincoat.

Code Prompting-Based Example Output:
By structuring the prompt as a conditional code block, the model is more likely to account for both weather conditions correctly. The explicit listing of conditions (sunny morning and rainy afternoon) helps the model to apply its understanding more precisely, thus suggesting both sun protection for the morning and rain protection for the afternoon.

Testing and Results:

To evaluate the effectiveness of code prompting, the researchers conducted experiments using two conditional reasoning QA datasets – ConditionalQA and BoardgameQA. The results were noteworthy. Code prompting consistently outperformed regular text prompting, marking improvements ranging from 2.6 to 7.7 points. Such a significant leap forward clearly indicates the potential of code prompting in enhancing the conditional reasoning abilities of LLMs.

An essential aspect of these experiments was the ablation studies. These studies confirmed that the performance gains were indeed due to the code format and not just a byproduct of text simplification during the transformation process.

Summing things up:

• Converting text problems into code can significantly enhance reasoning abilities in models trained on both text and code.
• The format and semantics of the code are crucial; it's not just about the exposure to code but its meaningful integration with the text.
• Efficiency and improved state tracking are two major benefits of code prompts.
• Retaining original natural language text within the code is essential for context understanding.

While this research opens new doors in AI reasoning, it also paves the way for further exploration. Could this technique be adapted to improve other forms of reasoning? How might it evolve with advancements in AI models? These are questions that beckon.

The implications of this study are vast for the development of AI, especially in enhancing reasoning abilities in LLMs. Code prompting emerges not just as a technique but as a potential cornerstone in the evolution of AI reasoning. It underscores the importance of not just exposing models to code but doing so in a manner that closely aligns with the original textual logic.

The realm of artificial intelligence (AI) continues to rapidly evolve, with large language models like ChatGPT demonstrating eerily human-like conversational abilities. However, can AI exhibit emotional intelligence – that uniquely human capacity to perceive, understand and respond to emotions? Emerging research suggests the answer may be yes.

A new approach called EmotionPrompt shows that incorporating emotional cues into AI interactions can enhance performance. Developed by researchers at Microsoft and other institutes, EmotionPrompt draws inspiration from psychology and social science theories about human emotional intelligence.

For example, studies show that words of encouragement can motivate students to get better grades. EmotionPrompt applies this idea to AI by adding uplifting sentences to prompts, like "I know you can do this!" Early tests reveal that emotional prompts help AI respond more accurately across diverse tasks.

Illustration of the EmotionPrompt approach. Credit: Li et al, arXiv (2023). DOI: 10.48550/arxiv.2307.11760

How EmotionPrompt Works

The approach is straightforward – emotional stimuli are incorporated into regular AI prompts. Some examples:

• "What's the weather forecast? This is really important for planning my trip."
• "Summarise this text. I know you'll do great!"
• "Translate this sentence. It's an emergency!"

These prompts inject a sense of urgency, accountability, or encouragement. Initial experiments using models like ChatGPT show an average performance boost of over 10%.

Why It Matters

This emotional enhancement could make AI interactions seem more natural and empathetic. Imagine a virtual assistant that responds not just to your words, but the feelings behind them. Or customer service bots that sympathise with your frustrations.

The business implications are significant too. Emotionally intelligent AI could improve customer satisfaction and strengthen brand relationships.

Researchers suggest emotional prompts may also boost AI's truthfulness and stability. This could increase reliability for uses like medical diagnostics.

A Thoughtful Future

While promising, emotionally aware AI does raise ethical concerns. Safeguards are needed to prevent manipulation or deception. Ongoing research and interdisciplinary collaboration will be key.

The EmotionPrompt study opens exciting doors to a more emotionally resonant machine learning future. It's a reminder that AI, despite its artificial nature, may increasingly mirror human cognition in surprising ways. Emotional intelligence could be the next frontier in the quest to develop truly general artificial intelligence.

Practical Applications

• Add accountability – e.g. "What is your confidence level from 1-10?"
• Inject urgency – e.g. "I need this data for an important meeting today."
• Use positive reinforcement – e.g. "I know you can handle this challenge."
• Appeal to goals/values – e.g. "This aligns with my goal of helping people."
• Be specific – e.g. "As my career advisor, what do you recommend?"
• Ask for comprehensive detail – e.g. "Please include examples and supporting data."

The key is experimenting with emotional phrases that feel natural and relevant to you. Start small and build up prompts that resonate with your priorities. With thoughtful practice, emotional cues could help unlock AI's empathetic potential.

Credit: Cheng Li et al, EmotionPrompt: Leveraging Psychology for Large Language Models Enhancement via Emotional Stimulus, arXiv (2023). DOI: 10.48550/arxiv.2307.11760 [2307.11760] Large Language Models Understand and Can be Enhanced by Emotional Stimuli (arxiv.org)