Braid: Bounded reasoning for LLMs using symbolic Mermaid graphs (arxiv.org)

Researchers have introduced BRAID (Bounded Reasoning for Autonomous Inference and Decisions), a novel framework designed to enhance the efficiency of Large Language Models (LLMs). By employing structured prompting through Mermaid-based instruction graphs, BRAID allows LLMs to reason in a more bounded and structured manner, rather than relying on unbounded natural-language token expansion. The framework was tested across various GPT models and benchmark datasets, demonstrating substantial improvements in reasoning accuracy and cost efficiency, especially for autonomous agent systems in production. The significance of BRAID lies in its potential to optimize inference processes for LLMs, addressing the nonlinear relationship between performance, cost, and token usage that often complicates deployments. By introducing structured, machine-readable prompts, BRAID paves the way for LLMs to achieve better performance with lower operational costs, making it a promising approach for developers and researchers focused on scalability and efficiency in AI applications. Detailed datasets and results supporting the findings are publicly accessible, encouraging further exploration and adaptation within the AI/ML community.