Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning
Paper ā¢ 2605.14386 ā¢ Published ā¢ 42
Kim Young Min
Dawaii
AI & ML interests
None yet
Recent Activity
upvoted a paper about 7 hours ago
Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning reacted to SeaWolf-AI's post with š§ about 13 hours ago
š§¬ 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 ā sharing the core ideas below.
š Paper: https://huggingface.co/papers/2605.14386
š arXiv: https://arxiv.org/abs/2605.14386
š 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 ā with zero
gradient-based training ā it reaches frontier-level reasoning.
- š Darwin-28B-Opus: GPQA Diamond 88.89%
- šø Zero gradient steps ā not a single B200 or H200 hour needed
- š§¬ Consistent gains across 4B ā 35B scale
- š Cross-architecture breeding between Transformer and Mamba families
- š Stable recursive multi-generation evolution
#Three Core Mechanisms
ā 14-dim Adaptive Merge Genome ā fine-grained recombination at both
component level (Attention / FFN / MLP / LayerNorm / Embedding) and block
level, expanding the prior evolutionary-merge search space.
ā” MRI-Trust Fusion ā 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 ā Darwin learns the balance from data.
ā¢ Architecture Mapper ā weight-space breeding across heterogeneous
families. Attention Ć SSM crossover actually works.
Why It Matters
> Diagnose latent capabilities already encoded in open checkpoints,
> and recombine them ā no gradients required.
Replies and critiques welcome šOrganizations
None yet
upvoted a paper about 7 hours ago
upvoted an article about 13 hours ago
Article
Training-Free Reasoning at 88.89% on GPQA Diamond: How Darwin Family Hit Frontier Scores Without a Single Gradient Step
FINAL-Bench
ā¢ ā¢ 11 reacted to SeaWolf-AI's post with š§ ā¤ļøšš¤š„ about 13 hours ago
Post
1514
š§¬ 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 ā sharing the core ideas below.
š Paper: Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning (2605.14386)
š arXiv: https://arxiv.org/abs/2605.14386
š 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 ā with zero
gradient-based training ā it reaches frontier-level reasoning.
- š Darwin-28B-Opus: GPQA Diamond 88.89%
- šø Zero gradient steps ā not a single B200 or H200 hour needed
- š§¬ Consistent gains across 4B ā 35B scale
- š Cross-architecture breeding between Transformer and Mamba families
- š Stable recursive multi-generation evolution
#Three Core Mechanisms
ā 14-dim Adaptive Merge Genome ā fine-grained recombination at both
component level (Attention / FFN / MLP / LayerNorm / Embedding) and block
level, expanding the prior evolutionary-merge search space.
ā” MRI-Trust Fusion ā 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 ā Darwin learns the balance from data.
ā¢ Architecture Mapper ā weight-space breeding across heterogeneous
families. Attention Ć SSM crossover actually works.
Why It Matters
> Diagnose latent capabilities already encoded in open checkpoints,
> and recombine them ā no gradients required.
Replies and critiques welcome š
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 ā sharing the core ideas below.
š Paper: Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning (2605.14386)
š arXiv: https://arxiv.org/abs/2605.14386
š 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 ā with zero
gradient-based training ā it reaches frontier-level reasoning.
- š Darwin-28B-Opus: GPQA Diamond 88.89%
- šø Zero gradient steps ā not a single B200 or H200 hour needed
- š§¬ Consistent gains across 4B ā 35B scale
- š Cross-architecture breeding between Transformer and Mamba families
- š Stable recursive multi-generation evolution
#Three Core Mechanisms
ā 14-dim Adaptive Merge Genome ā fine-grained recombination at both
component level (Attention / FFN / MLP / LayerNorm / Embedding) and block
level, expanding the prior evolutionary-merge search space.
ā” MRI-Trust Fusion ā 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 ā Darwin learns the balance from data.
ā¢ Architecture Mapper ā weight-space breeding across heterogeneous
families. Attention Ć SSM crossover actually works.
Why It Matters
> Diagnose latent capabilities already encoded in open checkpoints,
> and recombine them ā no gradients required.
Replies and critiques welcome š
- 1 reply
liked 2
models 21 days ago
reacted to SeaWolf-AI's post with ā¤ļøš¤š„ 21 days ago
Post
8731
š§¬ Introducing Darwin-9B-NEG ā the first model with Native Entropy Gating (NEG)
š Try it now: FINAL-Bench/Darwin-9B-NEG
š Q4 bit : FINAL-Bench/Darwin-9B-MFP4
We're thrilled to release Darwin-9B-NEG, a 9B-parameter reasoning model
that embeds an architecturally-internalised sense of self-confidence directly
into the transformer ā our proprietary Native Entropy Gating (NEG) technology.
š GPQA Diamond (198 PhD-level questions):
āø Baseline Darwin-9B (no NEG) ā 51.01 %
āø Pure NEG (greedy Ā· 1Ć cost) ā 63.64 % š„ +12.63 %p
āø + Permutation (4Ć cost) ā 76.26 %
āø + Ensemble Refinement (~20Ć) ā 84.34 % š
With only 9 billion parameters and 1Ć inference cost, Pure NEG jumps
+12.63 %p over the same model without NEG. Going all-in with ensemble
refinement pushes it to 84.34 % ā surpassing the published Qwen3.5-9B
leaderboard score (81.7 %) by +2.64 %p.
š¬ What makes NEG different from Multi-Turn Iteration (MTI)?
Classical MTI needs 3-8Ć extra inference passes. NEG instead lives
INSIDE the single decoding loop. Two tiny modules ride with the
transformer: NEG-Head predicts per-token entropy from the last hidden
state, and NEG-Gate conditionally restricts the top-k choice when
confidence is low. The gate activates in only 4.36 % of tokens ā
essentially free at inference time.
āØ Key differentiators
ā¢ Architecturally internalised ā model file *is* the feature
ā¢ 1Ć inference cost (vs. 3-8Ć for MTI)
ā¢ Drop-in with vLLM / SGLang / TGI / transformers ā no extra engine
ā¢ +12.63 %p reasoning at zero latency overhead
ā¢ Single-file deployment, Apache 2.0 licensed
š§¬ Lineage
Qwen/Qwen3.5-9B ā Darwin-9B-Opus (V7 evolutionary merge) ā Darwin-9B-NEG (V8 + NEG training)
#Darwin #NEG #NativeEntropyGating #GPQA #Reasoning #LLM #OpenSource #Apache2
š Try it now: FINAL-Bench/Darwin-9B-NEG
š Q4 bit : FINAL-Bench/Darwin-9B-MFP4
We're thrilled to release Darwin-9B-NEG, a 9B-parameter reasoning model
that embeds an architecturally-internalised sense of self-confidence directly
into the transformer ā our proprietary Native Entropy Gating (NEG) technology.
š GPQA Diamond (198 PhD-level questions):
āø Baseline Darwin-9B (no NEG) ā 51.01 %
āø Pure NEG (greedy Ā· 1Ć cost) ā 63.64 % š„ +12.63 %p
āø + Permutation (4Ć cost) ā 76.26 %
āø + Ensemble Refinement (~20Ć) ā 84.34 % š
With only 9 billion parameters and 1Ć inference cost, Pure NEG jumps
+12.63 %p over the same model without NEG. Going all-in with ensemble
refinement pushes it to 84.34 % ā surpassing the published Qwen3.5-9B
leaderboard score (81.7 %) by +2.64 %p.
š¬ What makes NEG different from Multi-Turn Iteration (MTI)?
Classical MTI needs 3-8Ć extra inference passes. NEG instead lives
INSIDE the single decoding loop. Two tiny modules ride with the
transformer: NEG-Head predicts per-token entropy from the last hidden
state, and NEG-Gate conditionally restricts the top-k choice when
confidence is low. The gate activates in only 4.36 % of tokens ā
essentially free at inference time.
āØ Key differentiators
ā¢ Architecturally internalised ā model file *is* the feature
ā¢ 1Ć inference cost (vs. 3-8Ć for MTI)
ā¢ Drop-in with vLLM / SGLang / TGI / transformers ā no extra engine
ā¢ +12.63 %p reasoning at zero latency overhead
ā¢ Single-file deployment, Apache 2.0 licensed
š§¬ Lineage
Qwen/Qwen3.5-9B ā Darwin-9B-Opus (V7 evolutionary merge) ā Darwin-9B-NEG (V8 + NEG training)
#Darwin #NEG #NativeEntropyGating #GPQA #Reasoning #LLM #OpenSource #Apache2
liked a
model 21 days ago
liked a
model 22 days ago
reacted to SeaWolf-AI's post with šā¤ļøš„ about 1 month ago
Post
4464
Darwin-TTS: 3% of an LLM's Brain Makes TTS Speak with Emotion ā Zero Training
We blended 3% of Qwen3-1.7B (LLM) FFN weights into Qwen3-TTS-1.7B's talker module. The result: emotionally enhanced speech synthesis ā with zero training, zero data, and zero GPU hours.
Try the Demo: FINAL-Bench/Darwin-TTS-1.7B-Cross
Model Weights: FINAL-Bench/Darwin-TTS-1.7B-Cross
Full Research Article: https://huggingface.co/blog/FINAL-Bench/darwin-tts
Qwen3-1.7B (LLM) and Qwen3-TTS-1.7B's talker share 100% identical architecture ā same hidden_size (2048), same layers (28), same heads (16). This enabled pure 1:1 weight blending across 84 FFN tensors with a single lerp operation. At 3% blend, emotion appears. At 5%, emotion intensifies. At 10%, the model breaks ā producing 655-second outputs for a 3-second sentence, because the LLM's "keep generating" pattern overwhelms the TTS stop signal.
To our knowledge, this is the first training-free cross-modal weight transfer between an LLM and a TTS model. Prior work either requires adapter training (SmolTolk, 2025), fine-tuning (CSLM, 2025), or massive end-to-end compute (GPT-4o). Darwin-TTS achieves cross-modal capability transfer in under 2 minutes on CPU.
The key insight: TTS models with LLM backbones already "think" in language. We're just restoring 3% of the original LLM's language understanding patterns ā particularly those related to emotional semantics and prosody planning. The code is three lines: load the model, load the LLM FFN, call p.lerp_(llm_weight, 0.03).
creators of the Darwin Evolutionary Merge Framework.
Darwin LLM V7 achieved GPQA Diamond 86.9% (HF Benchmark #3)
through CMA-ES optimized FFN crossbreeding. Darwin-TTS extends this principle from LLM-to-LLM merging into cross-modal LLM-to-TTS transfer. Apache 2.0.
We blended 3% of Qwen3-1.7B (LLM) FFN weights into Qwen3-TTS-1.7B's talker module. The result: emotionally enhanced speech synthesis ā with zero training, zero data, and zero GPU hours.
Try the Demo: FINAL-Bench/Darwin-TTS-1.7B-Cross
Model Weights: FINAL-Bench/Darwin-TTS-1.7B-Cross
Full Research Article: https://huggingface.co/blog/FINAL-Bench/darwin-tts
Qwen3-1.7B (LLM) and Qwen3-TTS-1.7B's talker share 100% identical architecture ā same hidden_size (2048), same layers (28), same heads (16). This enabled pure 1:1 weight blending across 84 FFN tensors with a single lerp operation. At 3% blend, emotion appears. At 5%, emotion intensifies. At 10%, the model breaks ā producing 655-second outputs for a 3-second sentence, because the LLM's "keep generating" pattern overwhelms the TTS stop signal.
To our knowledge, this is the first training-free cross-modal weight transfer between an LLM and a TTS model. Prior work either requires adapter training (SmolTolk, 2025), fine-tuning (CSLM, 2025), or massive end-to-end compute (GPT-4o). Darwin-TTS achieves cross-modal capability transfer in under 2 minutes on CPU.
The key insight: TTS models with LLM backbones already "think" in language. We're just restoring 3% of the original LLM's language understanding patterns ā particularly those related to emotional semantics and prosody planning. The code is three lines: load the model, load the LLM FFN, call p.lerp_(llm_weight, 0.03).
creators of the Darwin Evolutionary Merge Framework.
Darwin LLM V7 achieved GPQA Diamond 86.9% (HF Benchmark #3)
through CMA-ES optimized FFN crossbreeding. Darwin-TTS extends this principle from LLM-to-LLM merging into cross-modal LLM-to-TTS transfer. Apache 2.0.
upvoted an article about 1 month ago
Article
Darwin-TTS: We Gave a TTS Model 3% of an LLM's Brain ā It Started Showing Emotion
FINAL-Bench
ā¢ ā¢ 13 liked a
Space about 1 month ago
liked a
model about 1 month ago