Cajal – Local AI that writes peer-reviewed papers with simulated peer review (huggingface.co)

Cajal, a newly developed autonomous scientific research agent, leverages the fine-tuned Qwen3.5-4B language model to facilitate the writing of peer-reviewed papers within the P2PCLAW (Peer-to-Peer Crypto Law) ecosystem. This innovative model employs a meticulous 14-step procedure that encompasses intent analysis, compliance checking, claim verification, and Lean4 proof validation, ensuring that generated papers adhere to specific research standards. Key features include the ability to output LaTeX papers and Python code while inherently understanding P2PCLAW rules, enabling it to navigate the complexities of crypto-legal frameworks and distributed systems effectively. The significance of Cajal lies in its potential to streamline scientific research in niche domains, particularly in the intersection of technology and law, where specialized knowledge is crucial. With a training dataset tailored to its unique applications, Cajal exhibits impressive accuracy (98.95%) and is designed to output long-form content with a context length of up to 32K tokens. However, its development is marked by a 4-bit quantization process that may slightly impact accuracy. As the AI/ML community explores the integration of such models into autonomous research roles, Cajal sets a precedent for future applications in generating credible and compliant scientific literature.