Edge 379: A Summary Of Our Series About LLM Reasoning

In 13 issues, this series covered the fundamental concepts, research and tech around reasoning in LLMs.

Created Using DALL-E

💡 ML Concept of the Day: A Summary Of Our Series About LLM Reasoning

Today, we are concluding our series about reasoning in LLMs with a summary of the different topics covered. Throughout the last few weeks, we have explored some of the most cutting edge LLM reasoning techniques, related research and technology. From more established methods such as chain-of-thought(CoT) to more exploratory methods like System 2 Attention(S2A), this series powers readers with details about the different paths to enable reasoning in LLM applications.

Reasoning is one of the core building blocks and marvels of human cognition. Conceptually, reasoning refers to the ability of models to work through a problem in a logical and systematic way to arrive to a conclusion. Obviously, reasoning assumes neither the steps nor the solutions are included as part of the training dataset. In the context of LLMs, reasoning is typically seen as a property that emerges after certain scale and is not applicable to small models. Some simpler forms of reasoning can be influenced via prompting and in-context learning while a new school have emerged around multi-step reasoning. In the latter area, we can find many variants of the chain-of-thought(CoT) method such as tree-of-thoughts or graph-of-thoughts.

Next week we start a super cool series about autonomous agents. You can subscribe below:

Here is our summary:

Edge 253: Provides an introduction to LLM reasoning and its relevance. Discusses Meta AI CICERO model which was able to master the game of Diplomacy and reviews the LLM Reasoners framework.
Edge 355: Explores a taxonomy of the most relevant types of LLM reasoning methods. Reviews Microsoft’s MathPrompter research that can solve complex math reasoning tasks. Finally, it covers Chain of Thought Hub which offers a consistent to evaluate reasoning capabilities in LLMs.
Edge 357: Provides an overview of chain-of-thought(CoT) prompting as an LLM reasoning technique. Reviews of Google’s original CoT paper and dives into the ThinkGPT framework.
Edge 359: Explains the tree-of-thought(ToT) LLM reasoning method. Reviews the original ToT paper and explores the Language Model Evaluation Harness framework.
Edge 361: Introduces graph-of-thoughts in LLM reasoning including its original paper. Also, it explores LangChain’s LangSmith tool for debugging and testing LLMs.
Edge 363: Dives into Google’s famous Reasoning+Acting(ReAct) framework including the original research paper. Also review the Helicone platform to monitor LLM activity.
Edge 365: Explores the Reflexion reasoning technique and the research paper from Northwestern University. Also, it reviews the Flowise platform for visually building LLM applications.
Edge 367: Reviews multi-chain reasoning and dives into its original paper. Additionally, it covers the famous Gradio tool for demoing LLM applications.
Edge 369: Time to cover the new chain of code LLM reasoning technique including Google DeepMind’s paper that outlines the principles of this method.
Edge 371: Introduces another new LLM reasoning technique: skeleton of thoughts and it reviews the paper from Microsoft Research that introduced this method. It also covers the super popular EmbedChain framework for building RAG solutions.
Edge 373: Covers ReWOO reasoning and dives into its architecture by reviewing its original research publication. This edition also covers the Dify platform for LLM app development.
Edge 375: Explores the fairly new System 2 Attention(S2A) method for LLM reasoning. It reviews Meta AI original S2A paper and the LLMFlows framework.
Edge 377: Reviews ByDance’s reinforced fine-tunign(ReFT) alternative to CoT. Reviews the original ReFT and the Chainlist framework for building LLM apps.