Exploring the Internal Representations of Pangram 3.3.2

Pangram Labs, a company specializing in AI detection models, released research analyzing the internal representations of its latest model, Pangram 3.3.2. This research focuses on how the model separates AI-generated text from human text across its layers, going beyond mere statistical features.

Key research content is as follows:

• Dataset composition: Based on 5,000 documents, human and AI samples were balanced, and the experiment was conducted including the latest family of LLM models such as Claude 3.7, GPT-4o, Gemini 2.5, and DeepSeek R1.
• Analysis methodology: Instead of the model's final output values, the EditLens architecture is used to collect Activations at specific layers and perform document-level analysis.
• Research purpose: The goal is to prevent shortcutting, where the model relies on specific words (e.g., 'delve') or sentence structures, to correct unintended model behavior, and to deeply understand the AI detection mechanism.

Rather than using simple metrics like Perplexity or Burstiness, Pangram implements advanced detection performance through internal features that the model itself has grasped during the training process.