| name | darwin-family-evolutionary-merging |
| description | Training-free evolutionary merging framework for LLMs using MRI-Trust Fusion and adaptive merge genome. Use when: merging LLMs without training, evolutionary model combination, gradient-free weight-space recombination, cross-architecture model breeding, diagnostic-guided model merging, training-free reasoning improvement. Activation: darwin family, evolutionary merging, MRI-Trust Fusion, model merging, gradient-free merging, cross-architecture breeding, training-free LLM, weight-space recombination. |
Darwin Family: Evolutionary Merging for LLMs
Training-free evolutionary merging framework that achieves frontier-level reasoning performance (86.9% GPQA Diamond, #6/1252) through gradient-free weight-space recombination of existing model checkpoints.
Metadata
- Source: arXiv:2605.14386
- Authors: Taebong Kim, Youngsik Hong, Minsik Kim, Sunyoung Choi, Jaewon Jang, Junghoon Shin, Minseo Kim
- Published: 2026-05-14
Core Problem
Post-training LLMs is expensive. Can we improve reasoning performance by reorganizing latent capabilities already encoded in existing checkpoints, without any gradient-based training?
Three Key Innovations
1. 14-Dimensional Adaptive Merge Genome
- Fine-grained component-level recombination (attention layers, FFN, embeddings)
- Block-level recombination (layer-group mixing)
- Search space captures architectural flexibility beyond simple linear interpolation
2. MRI-Trust Fusion
- Layer-importance signals: Diagnostic metrics identify which layers contribute most to reasoning
- Learnable trust parameter: Adaptively balances diagnostic signals with evolutionary search
- Avoids naive averaging that dilutes specialized capabilities
3. Architecture Mapper
- Enables cross-architecture breeding between heterogeneous model families
- Combines Transformer and Mamba-based components
- Handles structural mismatches through alignment mapping
Implementation Guide
Prerequisites
- Multiple LLM checkpoints (same or different architectures)
- Evaluation benchmark for fitness scoring
- Compute for inference-based evaluation (no gradients needed)
Step-by-Step
-
Genome Definition
genome = {
'embed_ratio': 0.5,
'attn_q_ratio': 0.3,
'attn_kv_ratio': 0.7,
'ffn_up_ratio': 0.4,
'ffn_down_ratio': 0.6,
}
-
MRI-Trust Layer Analysis
layer_importance = compute_layer_importance(
model, benchmark, metric='reasoning_accuracy'
)
trust_param = optimize_trust(layer_importance)
-
Evolutionary Search
for generation in range(n_gens):
fitness = [evaluate(genome, benchmark) for genome in population]
population = select_and_mutate(population, fitness)
-
Cross-Architecture Mapping
mapping = architecture_mapper(transformer_model, mamba_model)
merged = apply_merge(mapping, best_genome)
Applications
- Training-free LLM capability improvement
- Multi-model ensemble without inference overhead
- Cross-architecture knowledge transfer
- Resource-efficient reasoning model scaling (4B to 35B)
- Recursive multi-generation model evolution
Pitfalls
- Evaluation cost: Fitness evaluation requires full benchmark runs — expensive for large populations
- Architecture compatibility: Cross-architecture merging requires careful component alignment
- Overfitting to benchmark: Evolutionary search may optimize for specific benchmark rather than general capability
- Trust parameter tuning: MRI-Trust requires domain-specific diagnostic metrics
Related Skills
- near-policy-distillation
- sd-search-on-policy-hindsight-distillation
- hybrid-quantum-classical-framework