cellular-reasoning-fabric

Cellular Reasoning Fabric (CRF)

A bio-inspired alternative to the Transformer architecture for language modeling.

What is CRF?

The Cellular Reasoning Fabric replaces self-attention with a population of autonomous, communicating cells that:

• Communicate via k-nearest-neighbor message passing (not global attention)
• Split when energy is high (allocate more compute)
• Die when energy is low (free resources)
• Merge when states are similar (consolidate knowledge)
• Uses O(N*k) routing instead of O(N^2) attention

Benchmark Results

Model Details

Dynamic Cell Population

The CRF model features a fully operational cell lifecycle:

• Splits: 12-16 per forward pass (cells divide to add compute)
• Merges: 2+ per forward pass (redundant cells consolidate)
• Deaths: Cells with depleted energy are removed
• Energy tracking: Diagnostic metrics for monitoring cell health

Usage

import torch
from src.crf_reasoning.crf_vectorized import CRFLanguageModel, AblationConfig

cfg = AblationConfig()
model = CRFLanguageModel(
    vocab_size=256,
    d_model=64,
    d_hidden=32,
    n_init_cells=16,
    max_cells=64,
    n_crf_steps=4,
    k_neighbors=3,
    cfg=cfg,
)

# Load weights
state_dict = torch.load("crf_model.pt", map_location="cpu")
model.load_state_dict(state_dict)

# Forward pass with metrics
x = torch.randint(0, 256, (1, 32))
logits, loss, metrics = model(x, targets=x, collect_metrics=True)
print(f"Splits: {metrics.n_splits}, Merges: {metrics.n_merges}")

Limitations

• Small-scale experiments (d_model=64, seq_len=32, 400 training samples)
• Single seed - multi-seed validation needed
• Inference latency is 4-8x higher than Transformer
• Not yet tested on large-scale benchmarks (MMLU, HumanEval)

Citation

If you use this work, please cite:

@misc{usman2026crf,
  title={Cellular Reasoning Fabric: A Bio-Inspired Alternative to Transformers},
  author={Yasir Usman},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/YasirUsman/cellular-reasoning-fabric}
}

License

MIT