Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning

Hugging Face Daily Papers Papers
evolutionary-merging model-merging language-models training-free reasoning

Summary

Darwin Family is a training-free framework for evolutionary merging of large language models via gradient-free weight-space recombination, achieving strong reasoning performance without additional training. The method introduces MRI-Trust Fusion and cross-architecture breeding to combine heterogeneous models.

We present Darwin Family, a framework for training-free evolutionary merging of large language models via gradient-free weight-space recombination. We ask whether frontier-level reasoning performance can be improved without additional training, by reorganizing latent capabilities already encoded in existing checkpoints. Darwin introduces three key ideas: (i) a 14-dimensional adaptive merge genome enabling fine-grained component- and block-level recombination; (ii) MRI-Trust Fusion, which adaptively balances diagnostic layer-importance signals with evolutionary search through a learnable trust parameter; and (iii) an Architecture Mapper that enables cross-architecture breeding between heterogeneous model families. Empirically, the flagship Darwin-27B-Opus achieves 86.9% on GPQA Diamond, ranking #6 among 1,252 evaluated models, and outperforming its fully trained foundation model without any gradient-based training. Across scales from 4B to 35B parameters, Darwin models consistently improve over their parents, support recursive multi-generation evolution, and enable a training-free evolutionary merge that combines Transformer- and Mamba-based components. Together, the Darwin Family demonstrates that diagnostic-guided evolutionary merging is a practical and reproducible alternative to costly post-training pipelines for reasoning-centric language models.
Original Article
View Cached Full Text

Cached at: 05/15/26, 04:23 AM

Paper page - Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning

Source: https://huggingface.co/papers/2605.14386

Abstract

The Darwin Family framework enables training-free evolutionary merging of large language models through gradient-free weight-space recombination, achieving superior reasoning performance without additional training.

We present Darwin Family, a framework for training-freeevolutionary mergingof large language models viagradient-free weight-space recombination. We ask whether frontier-levelreasoning performancecan be improved without additional training, by reorganizing latent capabilities already encoded in existing checkpoints. Darwin introduces three key ideas: (i) a 14-dimensional adaptivemerge genomeenabling fine-grained component- and block-level recombination; (ii)MRI-Trust Fusion, which adaptively balances diagnostic layer-importance signals with evolutionary search through a learnabletrust parameter; and (iii) anArchitecture Mapperthat enablescross-architecture breedingbetween heterogeneous model families. Empirically, the flagship Darwin-27B-Opus achieves 86.9% on GPQA Diamond, ranking #6 among 1,252 evaluated models, and outperforming its fully trained foundation model without any gradient-based training. Across scales from 4B to 35B parameters, Darwin models consistently improve over their parents, support recursive multi-generation evolution, and enable a training-free evolutionary merge that combines Transformer- andMamba-based components. Together, the Darwin Family demonstrates that diagnostic-guidedevolutionary mergingis a practical and reproducible alternative to costly post-training pipelines for reasoning-centric language models.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2605\.14386

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper17

#### FINAL-Bench/Darwin-35B-A3B-Opus Image-Text-to-Text• 36B• Updatedabout 2 hours ago • 2.25k • 79 #### FINAL-Bench/Darwin-36B-Opus Text Generation• 35B• Updated17 minutes ago • 2.7k • 58 #### FINAL-Bench/Darwin-31B-Opus Text Generation• 33B• Updatedabout 2 hours ago • 993 • 42 #### FINAL-Bench/Darwin-4B-David Text Generation• 8B• Updatedabout 2 hours ago • 324 • 39 Browse 17 models citing this paper## Datasets citing this paper1

#### FINAL-Bench/World-Model Viewer• Updatedabout 2 hours ago • 100 • 928 • 36

Spaces citing this paper12

Collections including this paper1

Similar Articles

Model Merging Scaling Laws in Large Language Models

Hugging Face Daily Papers

This paper establishes empirical scaling laws for language model merging, identifying power-law relationships between model size, expert count, and performance to enable predictive planning for optimal model composition.

Enhancing Multilingual Reasoning via Steerable Model Merging

arXiv cs.CL

This paper proposes ST-Merge, a steerable model merging framework that uses a gated cross-attention mechanism to adaptively modulate contributions of a multilingual model and a reasoning model, outperforming fixed merging approaches on multilingual reasoning benchmarks across 21 languages.

Enhanced and Efficient Reasoning in Large Learning Models

arXiv cs.AI

This paper proposes a method for improving reasoning in large language models by recoding data to explicitly represent relationships, enabling efficient principled reasoning with polynomial-time learnability for relational rules, which addresses hallucinations and supports sound reasoning across multiple calls.