Thinking in Chains: Designing Iterative Reasoning Prompts for Complex Tasks

Stepwise prompts boost LLM accuracy by over 30% on complex tasks, with research proving chain-of-thought outperforms direct queries. Practical insights into chaining, self-verification, and meta-prompts for program-aided reasoning. Join the prompt revolution and unlock AI’s true reasoning power!

Overview

This talk explores how to harness the power of stepwise prompting techniques, such as chain-of-thought (CoT), self-verification, and meta-prompting for program-aided reasoning, to significantly improve large language model (LLM) performance on complex, multi-step problems. Attendees will learn how to design iterative prompts that guide models through reasoning chains, boosting accuracy and explainability for challenging real-world tasks.

Why This Matters

Complex applications—ranging from mathematical problem solving to multi-hop question answering and code generation—require more than a one-shot response. Standard prompting often leads to incomplete or incorrect answers. Iterative reasoning techniques have been shown to improve accuracy by over 30%, making AI systems more reliable, transparent, and creative. Understanding these methods is crucial for practitioners pushing the boundaries of AI-driven decision-making.

Who Is the Talk For

• Prompt engineers looking to design more effective and reliable AI-driven systems.
• AI researchers interested in the latest techniques for boosting model reasoning and accuracy.
• Product managers and builders responsible for deploying language models in production applications.
• Data scientists and machine learning engineers applying LLMs to solve complex, real-world problems.
• Technical leaders aiming to enhance team workflow and best practices around prompt development.

What Will Participants Take Away?

• Practical frameworks for designing chain-of-thought and iterative prompts
• Insights into self-verification methods that increase output reliability
• How to leverage meta-prompts to enable program-aided reasoning and multi-step logic
• Real-world examples and case studies demonstrating improved LLM performance
• Best practices and common pitfalls when implementing reasoning chains

Talk Structure (30 minutes total)

1. Introduction (3 min)
   Brief context on the challenge of complex reasoning with LLMs and the limits of one-shot prompting.

2. Why Stepwise Reasoning Works (5 min)
   Overview of chain-of-thought prompting and related research showing significant accuracy gains. On the GSM8K math benchmark, Google’s PaLM 540B model improved from 55% accuracy with standard prompting to 74% with chain-of-thought prompting—a 19 percentage point jump.

3. Techniques Deep Dive (10 min)
   Exploration of chaining prompts, self-verification strategies, and meta-prompting for program-aided reasoning.

4. Applications & Case Studies (7 min)
   Real-world demos including math problem solving, multi-hop QA, and code debugging workflows.

5. Best Practices & Pitfalls (3 min)
   Tips for prompt design, tuning, and avoiding common errors when chaining reasoning steps.

6. Q&A (5 min)
   Open floor for audience questions and discussion.

This talk equips participants with cutting-edge prompt engineering methods to unlock more powerful and trustworthy AI reasoning in their projects.

Notes

• Technical Requirements: Please ensure access to high-speed internet and a projector suitable for live demo presentations.

Why I Am the Best Person to Give This Talk

I bring a unique blend of research, practical experience, and public speaking:

• Academic Credentials: I hold a Master’s in Data Science from King’s College London.
• Industry Expertise: As a Senior Data Scientist at Publicis Sapient, I lead NLP innovation within a specialist data science team, with deep experience across AI, quantum computing, brain-computer interfaces, and building large-scale products.
• Cross-Industry Experience: I’ve spearheaded data science projects at Ai Palette (AI for CPG innovation) and at Collinson, where I honed expertise in text mining, personalization, and scalable NLP.
• Accomplished Speaker: An experienced Mozilla Tech Speaker, I have delivered talks at global events like PyCon, MozFest, and CodeMash, and created a LinkedIn Learning Rust course with over 80,000 learners.
• Research & Development: I’ve published at IEEE and ACM conferences, worked as a developer and research fellow at SAP Labs, and have hands-on experience in web development, computer vision, and RPA.
• Real-World NLP Innovations: My track record includes deploying NLP systems—such as BERTopic for topic modeling—and my talk for Analytics Vidhya’s DataHour was attended by 4,200+ participants, earning a 4.6/5 rating.
• Clarity and Impact: My sessions are known for making technical topics accessible and actionable for diverse audiences.

With this background, I will deliver clear, practical, and research-informed guidance that empowers attendees to master advanced prompt engineering for real-world impact.