deepdrift 1.1.3

Zero-Training Hidden-State Monitoring for Robustness in Vision, Language, and Generative Models

DeepDrift

---

DeepDrift is a runtime neural network monitoring library based on analyzing the rate of change of hidden states — Semantic Velocity.

It requires no additional training, calibrates only on normal data, and introduces <1.5% computational overhead.

✅ Zero‑training — no gradient updates, no labeled failures
✅ Architecture‑agnostic — works with PyTorch models (ViT, CNN, MLP, Transformers)
✅ Cross‑domain — validated on LLMs, RL, vision, and diffusion models
✅ Production‑ready — sparse sampling, IQR thresholding, <1.5% overhead
✅ Open source — MIT license, pip installable

---

📊 Graphical Abstract

Semantic Velocity monitors internal hidden‑state dynamics across four domains: LLM hallucination detection, RL failure prediction, Vision Transformer OOD detection, and diffusion memorization detection.

---

🚀 Installation

From PyPI (recommended)

pip install deepdrift

From source (latest development version)

git clone https://github.com/Eutonics/DeepDrift.git
cd DeepDrift
pip install -e .

For experiments and development (extra dependencies)

pip install -e ".[all]"

---

🔧 Quick Start

Vision Transformer OOD Detection

from deepdrift import DeepDriftMonitor
from torchvision.models import vit_b_16

model = vit_b_16(pretrained=True)
monitor = DeepDriftMonitor(
    model,
    layer_names=['encoder.layers.11'],  # monitor last layer
    pooling='cls'
)

x = torch.randn(1, 3, 224, 224)
_ = model(x)
vel = monitor.get_spatial_velocity()
print(f"Peak velocity: {max(vel):.4f}")

RL Agent Failure Prediction

from deepdrift import DeepDriftMonitor
from stable_baselines3 import PPO

model = PPO.load("ppo_lunarlander")
monitor = DeepDriftMonitor(
    model.policy,
    layer_names=None,  # auto-detect
    pooling='flatten'
)

obs = env.reset()
while True:
    vel = monitor.get_temporal_velocity()
    if vel > threshold:  # anomaly detected!
        print("⚠️ Agent panic detected!")
    # ... take action

---

📚 Documentation

• Paper: DeepDrift: Zero‑Training Hidden‑State Monitoring for Robustness in Vision, Language, and Generative Models (arXiv, 2026)
• Zenodo: 10.5281/zenodo.14763258
• GitHub: https://github.com/Eutonics/DeepDrift
• PyPI: https://pypi.org/project/deepdrift/
• Hugging Face Demo: DeepDrift Explorer

---

📈 Key Results

Domain	Model / Algo	AUC	Cohen's d	Lead Time
LLM Hallucination	Qwen-2.5-7B	0.891	3.12	7.2 tokens
RL (LunarLander)	PPO	0.968	2.47	12.3 steps
RL (CartPole)	DQN	0.985	2.79	168 steps
RL (CartPole+noise)	PPO	1.000	3.37	97 steps
ViT OOD	ViT-B/16	0.817	—	—
Diffusion Memorization	U-Net	3× earlier	—	—

---

🧪 Reproducibility

All experiments are fully reproducible:

git clone https://github.com/Eutonics/DeepDrift.git
cd DeepDrift
pip install -e ".[all]"
bash scripts/reproduce_all.sh

Results will be saved in results/ and paper/figures/.

---

📖 Citation

If you use DeepDrift in your research, please cite:

@article{evtushenko2026deepdrift,
  title={DeepDrift: Zero-Training Hidden-State Monitoring for Robustness in Vision, Language, and Generative Models},
  author={Alexey Evtushenko},
  year={2026},
  journal={arXiv preprint},
  url={https://github.com/Eutonics/DeepDrift}
}

---

📄 License

MIT License © 2026 Alexey Evtushenko

---

DeepDrift was developed as independent research. No affiliation with Yandex, K‑Dense, or any other organization.