A GPT-2 Small (124M params) model where a 354-parameter genome controls 35.4 million connections. The genome decides which attention output projections and feed-forward layers are connected, permanently disabling one-third of masked connections.

99,970:1 compression ratio. 354 learned parameters specify the wiring for 35.4M connections.

How It Works

Neural DNA (NDNA) encodes network connectivity in a compact genome. Instead of storing one mask bit per connection, the genome stores a developmental program: 8 cell types with 8-dimensional affinity vectors and a compatibility matrix. Source and target type embeddings are compared to produce connection probabilities, which are thresholded to binary masks using a straight-through estimator. A metabolic cost term encourages sparsity, forcing the genome to be selective about which connections survive.

The genome and model weights are trained jointly from scratch on OpenWebText. Temperature annealing (1.0 to 10.0) smoothly transitions from soft to hard binary masks, locking the topology by the end of training.

What the genome controls:

Masked: W_O (768x768, bias=False) and FF1 (3072x768) per layer = 2.95M connections/layer, 35.4M total across 12 layers
Not masked: Q, K, V projections, FF2, LayerNorm, residual connections, LM head

The genome controls 28.4% of the model's total parameters. The remaining 71.6% train freely.

Results

Validation loss over 52,700 iterations. Best checkpoint: 3.04 at iteration 49,500.

Benchmark	NDNA GPT-2	GPT-2 Small	Result
WikiText-103 PPL	36.0	37.5	Beats GPT-2
Penn Treebank PPL	59.4	65.9	Beats GPT-2
LAMBADA PPL	22.2	35.1	Beats GPT-2
LAMBADA ACC	30.8%	46.0%	67% of GPT-2
HellaSwag ACC	28.7%	31.2%	92% of GPT-2
CBT-CN ACC	82.7%	87.7%	94% of GPT-2
CBT-NE ACC	74.5%	83.4%	89% of GPT-2
enwiki8 BPB	1.39	1.16
text8 BPC	1.27	1.17

Beats GPT-2 on 3 of 9 benchmarks (all perplexity metrics). The model produces better-calibrated probability distributions but less peaked predictions, leading to lower perplexity but lower exact-match accuracy.

Supplementary Results (Language Model Evaluation Harness)

Benchmark	Metric	NDNA GPT-2
ARC-Easy	ACC	41.8%
PIQA	ACC	60.1%
Winogrande	ACC	52.8%
HellaSwag	ACC (norm)	28.8%

Learned Topology

Final per-layer hard density at temperature 10.0. Sharp boundary between sparse (L1-L4) and dense (L5-L12) zones.

Layer	1	2	3	4	5-12
Hard Density	98.2%	45.2%	20.4%	7.7%	100%

Layers 5-12 are fully connected. Layers 1-4 are progressively pruned in a strict monotonic gradient. The genome treats the network as having two distinct zones. One-third of all masked connections (11.8M out of 35.4M) are permanently disabled.

Training Dynamics

Per-layer hard density over training. Layer 1 is pruned at iter 800, then re-activates at iter 26,000.

The genome goes through four distinct phases:

Over-activation (iter 0-200): Genome activates 83% of connections. Everything is on.
Aggressive pruning (iter 200-800): Layers 1-4 are pruned to 0% hard density. The genome discovers the model can function with only 8 of 12 layers carrying signal through masked projections.
Stable learning (iter 800-25K): Topology is fixed. Validation loss drops from 7.99 to 3.18. The network learns language with a sparse topology.
Layer 1 resurrection (iter 25K-50K): Layer 1 re-activates from 0% to 98.2% hard density. The genome changed its mind. Layers 2-4 remain pruned.

The layer 1 resurrection is the most unexpected finding. After 25,000 iterations of being completely disabled, the genome reverses its decision and reconnects layer 1. This is a discrete topological event, not a smooth interpolation.

Usage

python

import torch
from genome.model import Genome, GrownGPT2

# Load checkpoint
ckpt = torch.load("genome_ckpt.pt", map_location="cpu")

# Initialize
genome = Genome(n_types=8, type_dim=8, n_bands=14)
model = GrownGPT2(genome)

# Load weights (strip torch.compile prefix)
model_state = {k.replace("_orig_mod.", ""): v for k, v in ckpt["model"].items()}
genome_state = {k.replace("_orig_mod.", ""): v for k, v in ckpt["genome"].items()}
genome.load_state_dict(genome_state)
model.load_state_dict(model_state, strict=False)

# Enable hard binary masks for evaluation
model.hard_masks = True
model.eval()

Training Details

Dataset: OpenWebText (open reproduction of WebText)
Hardware: Single NVIDIA A100 80GB GPU
Iterations: 52,700 (best checkpoint at 49,500)
Tokens: 25.9 billion (491,520 tokens/iter = 12 seqs x 1024 tokens x 40 grad accum)
Weight optimizer: AdamW (lr=6e-4, beta1=0.9, beta2=0.95, cosine decay, warmup 2000, weight decay 0.1)
Genome optimizer: Adam (lr=0.01, constant, no decay)
Temperature: 1.0 to 10.0 over 50K iterations (linear anneal)
Sparsity weight: 0.005
Precision: bfloat16 with torch.compile
Gradient clipping: 1.0
Random seed: 42

Genome Configuration

Component	Shape	Parameters
Affinity matrix A	8x8	64
Compatibility matrix C	8x8	64
Connection scale	scalar	1
Depth penalty	scalar	1
Band type base	14x8	112
Band type gradient	14x8	112
Total		354

8 cell types, 8-dimensional affinity vectors, 14 bands (1 embedding + 12 transformer layers + 1 output).

Compression Ratio Progression

Experiment	Genome Params	Connections	Compression
MNIST MLP	226	174K	770:1
CIFAR-10 MLP	226	1.7M	7,553:1
IMDB Transformer	258	2.2M	8,384:1
GPT-2 Small	354	35.4M	99,970:1

Author

Tejas Parthasarathi Sudarshan Independent Researcher, Chennai, India tejas@fandesk.ai | tejassuds.com | LinkedIn

Citation

bibtex

@article{sudarshan2026ndna,
  title={Neural DNA: A Compact Genome for Growing Network Architecture},
  author={Sudarshan, Tejas Parthasarathi},
  year={2026},
  doi={10.5281/zenodo.19248389}
}

@article{sudarshan2026ndna_gpt2,
  title={Scaling Neural DNA to GPT-2: 354 Parameters Wire a Language Model},
  author={Sudarshan, Tejas Parthasarathi},
  year={2026},
  doi={10.5281/zenodo.19390927}
}

tejassuds/ndna-gpt2-small

NDNA GPT-2 Small