NeuroScope

Functional MRI for Large Language Models

A unified toolkit for real-time visualization and analysis of LLM internal activations, functional connectivity, and feature extraction. Think of it as a brain scanner for transformer models.

Core Capabilities

• Feature Extraction -- SAE-based decomposition of activations into interpretable features (the "voxels")
• Functional Connectivity -- Graph-based mapping of co-activation patterns (the "connectome")
• Activation Narration -- Oracle models that describe what features represent (the "radiologist")
• Real-time Visualization -- 3D brain-scan-style dashboard (the "scanner")

Quick Start

pip install -e ".[all]"

from neuroscope import NeuroScanner
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("gpt2")
tokenizer = AutoTokenizer.from_pretrained("gpt2")

scanner = NeuroScanner(model, tokenizer, layers=[0, 6, 11])
result = scanner.scan("The capital of France is")

for sae_result in result.sae_results:
    print(f"Layer {sae_result.layer_idx}: {len(sae_result.features)} active features")
    for f in sae_result.features[:5]:
        print(f"  Feature {f.feature_idx}: strength={f.activation_strength:.3f}")

Architecture

neuroscope/
  extractors/      # Activation hooks + SAE feature extraction
    hooks.py       # Forward hook infrastructure
    sae.py         # Sparse autoencoder implementation
  connectivity/    # Functional connectivity analysis
    graph.py       # Co-activation graphs + network discovery
  visualization/   # 3D rendering + web dashboard
  scanner.py       # Main orchestrator (the "fMRI machine")

Key References

• Brain-Inspired Exploration of Functional Networks in LLMs (COLING 2025)
• Probing Neural Topology of LLMs (2025)
• Scaling Monosemanticity (Anthropic, 2024)
• Activation Oracles (2025)
• SAE Survey (2025)

License

MIT