AI Papers Reader

Personalized digests of latest AI research

View on GitHub

Large Language Models Can Now Think Like Humans, Thanks to a New Framework

Large language models (LLMs) are getting increasingly better at performing complex tasks, but they still struggle with reasoning. Traditional approaches to reasoning in LLMs, such as chain-of-thought prompting, represent reasoning as a linear sequence of steps, which doesn’t capture the complexity of human reasoning.

A new framework called Diagram of Thought (DoT), proposed in a recent paper by researchers at Tsinghua University and Shanghai Artificial Intelligence Laboratory, models reasoning in LLMs as the construction of a directed acyclic graph (DAG). This allows the LLM to explore complex reasoning pathways while maintaining logical consistency.

The DoT framework is based on the idea that reasoning is an iterative process of proposing ideas, critiquing them, refining them, and verifying them. Each node in the DAG represents a proposition, critique, refinement, or verification. The model can then transition seamlessly between proposing ideas and critically evaluating them, using role-specific tokens that allow it to perform different reasoning roles.

For example, the authors demonstrate DoT on the task of answering the question “Which one is larger, 9.11 or 9.8?” In their example, the LLM first proposes that 9.8 is larger because the first digit after the decimal point in 9.8 is 8, which is greater than the first digit after the decimal point in 9.11, which is 1. The LLM then critiques its own reasoning, realizing that it needs to align the decimal points of the two numbers before comparing them. The LLM then refines its reasoning, stating that 9.8 is indeed larger than 9.11.

The authors further formalize the DoT framework using Topos Theory, a branch of category theory that provides a rigorous mathematical foundation for reasoning. This ensures that the reasoning process is both logically consistent and sound.

DoT is a significant step forward in the development of LLMs capable of human-like reasoning. It offers a conceptual framework for designing next-generation reasoning-specialized models, emphasizing training efficiency, robust reasoning capabilities, and theoretical grounding.

The paper is available on arXiv: https://arxiv.org/abs/2409.10038