Cognexis is an open‑source research initiative exploring compute‑adaptive recurrent‑depth language models. We develop and maintain a modular, Rust‑based implementation skeleton, a comprehensive specification suite, and supporting research artifacts that aim to advance dynamic reasoning in large language models.

Our mission is to decouple reasoning depth from model size by enabling dynamic compute allocation through recurrent loops. By making the depth of computation a runtime variable rather than a fixed architectural constant, we hope to provide a more efficient and flexible path toward powerful language models.

• Cognexis Rust crate: A skeleton implementation of the Cognexis architecture in Rust, showcasing module boundaries and data structures for future development.
• Specifications: A set of 28 detailed specification documents covering every component of the architecture—from tokenization and embedding through recurrence, scheduling, and evaluation.
• White paper: A research paper outlining the theory, design, training protocols, and evaluation metrics of compute‑adaptive recurrent‑depth transformers.

We welcome contributions and collaboration from the research and open‑source communities. If you’d like to help develop the Cognexis implementation or explore compute‑adaptive LLMs, please check our issues, open discussions, or reach out via GitHub. For guidelines on contributing, refer to the documentation in this repository.