Darwin Family: Training-Free Model Merging Boosts LLM Reasoning to 86.9% on GPQA Diamond

Researchers have unveiled the Darwin Family, a novel framework that facilitates the evolutionary merging of large language models without the need for training, utilizing gradient-free weight-space recombination. This technique reorganizes latent capabilities from pre-existing checkpoints. It features three primary innovations: a 14-dimensional adaptive merge genome for precise recombination, MRI-Trust Fusion that harmonizes diagnostic layer signals with evolutionary search, and an Architecture Mapper that supports cross-architecture breeding. The standout model, Darwin-27B-Opus, achieved an impressive 86.9% on GPQA Diamond, securing the #6 position among 1,252 models assessed, and surpassing its fully trained foundation model without employing gradient-based training. This method is scalable from 4B parameters and beyond.