READER: Robust Evidence-based Authorship Decoding via Extracted Representations

1. 1Build a dataset of LLM responses to query-varying prompts
2. 2Extract representations from the responses using techniques like embeddings
3. 3Configure a model to decode authorship based on the extracted representations
4. 4Test the model on unseen data to evaluate its robustness
5. 5Apply the technique to real-world applications to ensure provenance

AI engineers and researchers benefit from this technique to ensure transparency and accountability in LLM-driven systems, while product managers can apply this to improve model reliability

💡 Dynamic Black-Box LLM Provenance is crucial for transparency and accountability in agentic applications

🤖 Identify the source LLM of a black-box response with READER! 💡