Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning
Paper β’ 2605.14386 β’ Published β’ 50
JangJaewon
Be2Jay
AI & ML interests
None yet
Recent Activity
reacted to SeaWolf-AI's post with π₯ 1 day ago
𧬠Darwin Family: Zero Gradient Steps, GPQA Diamond 88.89%
How far can we push LLM reasoning *without* training?
Our team at VIDRAFT submitted this paper to Daily Papers yesterday, and it's
currently #3. Huge thanks to everyone who upvoted β sharing the core ideas below.
π Paper: https://huggingface.co/papers/2605.14386
π arXiv: https://arxiv.org/abs/2605.14386
π Model: https://huggingface.co/FINAL-Bench/Darwin-28B-Opus
---
TL;DR
Darwin Family is a training-free evolutionary merging framework.
By recombining the weight spaces of existing LLM checkpoints β with zero
gradient-based training β it reaches frontier-level reasoning.
- π Darwin-28B-Opus: GPQA Diamond 88.89%
- πΈ Zero gradient steps β not a single B200 or H200 hour needed
- 𧬠Consistent gains across 4B β 35B scale
- π Cross-architecture breeding between Transformer and Mamba families
- π Stable recursive multi-generation evolution
#Three Core Mechanisms
β 14-dim Adaptive Merge Genome β fine-grained recombination at both
component level (Attention / FFN / MLP / LayerNorm / Embedding) and block
level, expanding the prior evolutionary-merge search space.
β‘ MRI-Trust Fusion β we diagnose each layer's reasoning contribution
via an **MRI (Model Reasoning Importance)** signal and fuse it with
evolutionary search through a **learnable trust parameter**. Trust the
diagnostic too much and search collapses; ignore it and search becomes
inefficient β Darwin learns the balance from data.
β’ Architecture Mapper β weight-space breeding across heterogeneous
families. Attention Γ SSM crossover actually works.
Why It Matters
> Diagnose latent capabilities already encoded in open checkpoints,
> and recombine them β no gradients required.
Replies and critiques welcome π liked a dataset 4 days ago
Idavidrein/gpqaOrganizations
upvoted a paper 1 day ago
reacted to SeaWolf-AI's post with π₯ 1 day ago
Post
3338
𧬠Darwin Family: Zero Gradient Steps, GPQA Diamond 88.89%
How far can we push LLM reasoning *without* training?
Our team at VIDRAFT submitted this paper to Daily Papers yesterday, and it's
currently #3. Huge thanks to everyone who upvoted β sharing the core ideas below.
π Paper: Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning (2605.14386)
π arXiv: https://arxiv.org/abs/2605.14386
π Model: FINAL-Bench/Darwin-28B-Opus
---
TL;DR
Darwin Family is a training-free evolutionary merging framework.
By recombining the weight spaces of existing LLM checkpoints β with zero
gradient-based training β it reaches frontier-level reasoning.
- π Darwin-28B-Opus: GPQA Diamond 88.89%
- πΈ Zero gradient steps β not a single B200 or H200 hour needed
- 𧬠Consistent gains across 4B β 35B scale
- π Cross-architecture breeding between Transformer and Mamba families
- π Stable recursive multi-generation evolution
#Three Core Mechanisms
β 14-dim Adaptive Merge Genome β fine-grained recombination at both
component level (Attention / FFN / MLP / LayerNorm / Embedding) and block
level, expanding the prior evolutionary-merge search space.
β‘ MRI-Trust Fusion β we diagnose each layer's reasoning contribution
via an **MRI (Model Reasoning Importance)** signal and fuse it with
evolutionary search through a **learnable trust parameter**. Trust the
diagnostic too much and search collapses; ignore it and search becomes
inefficient β Darwin learns the balance from data.
β’ Architecture Mapper β weight-space breeding across heterogeneous
families. Attention Γ SSM crossover actually works.
Why It Matters
> Diagnose latent capabilities already encoded in open checkpoints,
> and recombine them β no gradients required.
Replies and critiques welcome π
How far can we push LLM reasoning *without* training?
Our team at VIDRAFT submitted this paper to Daily Papers yesterday, and it's
currently #3. Huge thanks to everyone who upvoted β sharing the core ideas below.
π Paper: Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning (2605.14386)
π arXiv: https://arxiv.org/abs/2605.14386
π Model: FINAL-Bench/Darwin-28B-Opus
---
TL;DR
Darwin Family is a training-free evolutionary merging framework.
By recombining the weight spaces of existing LLM checkpoints β with zero
gradient-based training β it reaches frontier-level reasoning.
- π Darwin-28B-Opus: GPQA Diamond 88.89%
- πΈ Zero gradient steps β not a single B200 or H200 hour needed
- 𧬠Consistent gains across 4B β 35B scale
- π Cross-architecture breeding between Transformer and Mamba families
- π Stable recursive multi-generation evolution
#Three Core Mechanisms
β 14-dim Adaptive Merge Genome β fine-grained recombination at both
component level (Attention / FFN / MLP / LayerNorm / Embedding) and block
level, expanding the prior evolutionary-merge search space.
β‘ MRI-Trust Fusion β we diagnose each layer's reasoning contribution
via an **MRI (Model Reasoning Importance)** signal and fuse it with
evolutionary search through a **learnable trust parameter**. Trust the
diagnostic too much and search collapses; ignore it and search becomes
inefficient β Darwin learns the balance from data.
β’ Architecture Mapper β weight-space breeding across heterogeneous
families. Attention Γ SSM crossover actually works.
Why It Matters
> Diagnose latent capabilities already encoded in open checkpoints,
> and recombine them β no gradients required.
Replies and critiques welcome π
- 3 replies
liked a
dataset 4 days ago
updated a
Space 6 days ago
liked a
Space 12 days ago
updated a
Space 17 days ago
liked a
model 19 days ago
liked a
Space 22 days ago
published a
Space 22 days ago
liked a
Space about 1 month ago
upvoted an article about 1 month ago
Article
"Darwin-27B-Opus: Surpassing the Foundation Model Without Training"
FINAL-Bench
β’ β’ 13 upvoted a collection about 1 month ago
liked 2
models about 1 month ago
updated a
model about 1 month ago
published a
model about 1 month ago
liked a
Space about 1 month ago
liked a
model about 1 month ago
liked a
Space about 1 month ago
liked a
model about 2 months ago