Atoms of Thought (github.com)

Anthropic has unveiled an innovative reasoning framework called "Atom of Thoughts," integrated into its Claude AI models. This framework is built on the concepts presented in the paper "Atom of Thoughts for Markov LLM Test-Time Scaling" by Teng et al., slated for NeurIPS 2025. Significantly, Claude can now utilize this decomposition-contraction methodology directly, eliminating the need for a separate MCP server. This streamlined approach enhances the model’s ability to perform complex reasoning tasks more effectively. The Atom of Thoughts framework includes two variants: the full version, which features a Directed Acyclic Graph (DAG) decomposition method with depth 5 and confidence scoring, and a lighter version that enables faster processing with a single cycle and simplified confidence metrics. These developments improve token efficiency and ensure structured data is parsed more cleanly than traditional prose markdown, facilitating better instruction following. This advancement marks a substantial shift in how LLMs can dynamically manage and improve their reasoning processes, likely paving the way for more sophisticated applications of AI in various domains.