Show HN: A repo to turn any model into a reasoning model without training (github.com)

A new GitHub repository introduces an innovative method to enhance large language model (LLM) reasoning without necessitating additional training. This approach utilizes evolutionary algorithms to finely tune models in latent space, producing responses that are more specific and actionable. Instead of generating generic outputs like "Define the goal" for planning, the model can deliver detailed technical instructions, such as how to implement rate limiting in an API, complete with concrete steps. This significant advancement is particularly impactful for the AI/ML community as it promises improvements in the quality and relevance of automated content generation. By encoding queries into an LLM's hidden states and subjecting them to a process of selection, mutation, and crossover, the method can effectively evolve these representations into higher-quality textual outputs. The system supports multiple transformer models and incorporates a trained neural network for evaluating response quality, making it a versatile tool for developers. Its user-friendly interface and configuration options ensure accessibility while delivering powerful capabilities for understanding complex queries.