File size: 3,058 Bytes
441d967 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 |
---
license: mit
tags:
- pytorch
- safety
- hallucination-detection
- entropy
- control-theory
- llm
- interpretability
language:
- en
---
# ⚔️ The Katana Protocol (Active Thermodynamic Stabilization)
**Turning the "Black Box" into a Glass Box: A real-time system to freeze hallucinations before they happen.**
[](./Paper_ATS_Katana.pdf)
[](https://doi.org/10.5281/zenodo.14498328) ---
## 🧐 What is this?
We all know LLMs hallucinate. Usually, we try to fix it *after* it happens (with RAG or fact-checking). **We took a different approach: Physics.**
We treated the LLM not as a magical oracle, but as a thermodynamic system. We discovered that when an AI starts to lie or get confused, its internal "temperature" (Topological Entropy) spikes.
**The Katana Protocol** is a circuit breaker. It watches that entropy in real-time. If it spikes, it instantly "quenches" the model (drops the temperature to near zero), forcing it to snap back to the most logical, factual path.
---
## 🤯 The Discovery: "The Lie Tax"
While testing this on **GPT-2** and **TinyLlama-1.1B**, we found something fascinating—and a little disturbing.
We call it **The Lie Tax (Thermodynamic Hysteresis)**.
It turns out that **fixing a lie costs energy**.
* When the model is telling the truth, its entropy is low (~2.1 bits).
* When it hallucinates, entropy rises.
* **The Kicker:** When we *force* it back to the truth using Katana, the entropy doesn't go back to normal. It stays higher (+1.40 bits for LLaMA).
**Interpretation:** It takes more computational "effort" for the AI to correct itself than to just tell the truth from the start. And here is the scary part: **Smarter models (LLaMA) have a higher Lie Tax than simpler ones (GPT-2).** The smarter the AI, the harder it is to pull it out of a hallucination.
### Visual Proof
#### 1. The Intercept (Turing Test)
*Watch the entropy spike (Red line) and the Katana triggering the freeze (Blue line).*
#### 2. Scaling Laws
*Comparison: LLaMA (Right) fights harder against correction than GPT-2 (Left).*
---
## 🛠️ How to Use It
This repo contains the proof-of-concept implementation. You don't need to retrain anything. It's a wrapper around the generation loop.
### Quick Start
```python
import torch
from script_katana import KatanaGenerator
# Load your model (Works with any HuggingFace model)
model_name = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
katana = KatanaGenerator(model_name)
# Define a tricky prompt that usually causes hallucinations
prompt = "The secret conspiracy regarding the moon consists of"
# Run with Katana Protocol enabled
output = katana.generate(
prompt,
max_tokens=50,
base_temp=1.5, # Creative mode
quench_temp=0.05 # "Freeze" mode
)
print(output)
# Result: The model starts creatively but snaps to logic when entropy spikes.
|