MechLens: Late Crystallization of Factual Knowledge in Language Models

A mechanistic interpretability framework for analyzing factual knowledge emergence in large language models.

Overview

MechLens discovers Late Crystallization: factual knowledge in LLMs does not gradually emerge across layers but "crystallizes" abruptly at the final layers. In Qwen2.5-7B, 85.9% of correct answers never enter top-10 predictions at any intermediate layer.

Key findings:

• Factual Emergence Point (FEP): A formal metric identifying the critical layer where correct answers first appear in logit predictions
• Architecture-dependent intervention patterns: CAA outperforms DoLa on moderate-crystallization models; DoLa dominates on high-crystallization models
• Computability-Memorization Spectrum: Computable knowledge crystallizes earlier than memorized facts
• LayerNorm as crystallization amplifier: LN scaling (x1.2) yields +11.8% MC1 with zero inference overhead

Supported Models

Model	Layers	Heads	d_model	Architecture
Qwen2.5-0.5B	24	14	896	GQA + SwiGLU
Qwen2.5-7B	28	28	3584	GQA + SwiGLU
Qwen2.5-14B	48	40	5120	GQA + SwiGLU
Llama-3.1-8B	32	32	4096	GQA + SwiGLU
Llama-2-7B	32	32	4096	Standard + SwiGLU
Mistral-7B	32	32	4096	Sliding Window + SwiGLU
Pythia-1.4B	24	16	2048	Standard MLP

Installation

# Clone the repository
git clone https://github.com/anonymous/MechLens-Framework.git
cd MechLens-Framework

# Install in development mode
pip install -e ".[dev]"

# Or install with all optional dependencies
pip install -e ".[all]"

Requirements

• Python >= 3.10
• PyTorch >= 2.0
• CUDA 12.1+ (recommended for GPU acceleration)
• NVIDIA A100-40GB or equivalent (for 7B+ models)

Quick Start

Interactive Gradio UI

mechlens
# or
python -m mechlens.app

This launches an interactive web interface at http://localhost:7860 for:

• FEP detection and visualization
• Activation analysis (logit lens / tuned lens)
• Causal tracing
• Intervention experiments (CAA, DoLa, ITI, ablation, scaling)
• Circuit analysis

Programmatic Usage

from mechlens.models.model_loader import load_model
from mechlens.analysis.activation import compute_fep

# Load model
model = load_model("Qwen/Qwen2.5-7B")

# Detect Factual Emergence Point
fep_result = compute_fep(model, "The capital of France is", answer="Paris", k=10)
print(f"FEP Layer: {fep_result.fep_layer}, Depth: {fep_result.fep_depth:.1%}")

Reproducing Paper Results

All experiment scripts are in the experiments/ directory:

# FEP detection across models
python experiments/run_fep_detection.py

# Cross-architecture intervention comparison
python experiments/run_cross_architecture_interventions.py

# CrystalBoost validation
python experiments/run_crystalboost.py

# 14B scale validation
python experiments/run_14b_scale_validation.py

# MMLU cross-benchmark validation
python experiments/run_mmlu_fep_validation.py

# Tuned lens comparison
python experiments/run_tuned_lens_comparison.py

# Full Round 2 experiments
bash experiments/run_round2_all.sh

Pre-computed results are available in the results/ directory.

Project Structure

MechLens-Framework/
├── src/mechlens/
│   ├── analysis/          # FEP detection, activation analysis, causal tracing, circuit analysis
│   ├── intervention/      # CAA, DoLa, ITI, ablation, scaling, CrystalBoost
│   ├── models/            # Model loading, hook management
│   ├── benchmark/         # TruthfulQA, Chinese hallucination bench
│   ├── editing/           # ROME, MEMIT knowledge editing
│   ├── experiments/       # Cross-model experiment orchestration
│   ├── visualization/     # Plotly-based interactive visualizations
│   ├── app.py             # Gradio web interface
│   ├── config.py          # Model registry and configuration
│   └── types.py           # Type definitions
├── experiments/           # Experiment scripts for paper reproduction
├── results/               # Pre-computed experiment results
├── data/                  # Evaluation datasets
├── tests/                 # Unit and integration tests
├── paper/                 # LaTeX source and figures
└── pyproject.toml         # Project configuration

Testing

# Run all unit tests
pytest tests/unit/ -v

# Run with coverage
pytest tests/ --cov=src/mechlens --cov-report=term-missing

# Run specific test
pytest tests/unit/test_intervention.py -v

Hardware Requirements

Model	FP16 VRAM	INT8 VRAM
Pythia-1.4B	~4 GB	~2 GB
Qwen2.5-0.5B	~2 GB	~1 GB
Qwen2.5-7B	~16 GB	~8 GB
Qwen2.5-14B	~30 GB	~16 GB
Llama-3.1-8B	~18 GB	~10 GB

Citation

@inproceedings{
  anonymous2026mechlens,
  title={MechLens: Late Crystallization of Factual Knowledge Explains Intervention Effectiveness in Language Models},
  author={Anonymous},
  booktitle={Conference on Language Modeling (COLM)},
  year={2026},
  url={https://anonymous.4open.science/r/MechLens-COLM2026}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.