ROHIT SURYADEVARA

MAP — Model-Agnostic Probabilistic Attribution for Prompt Reversal

A statistical framework that identifies which language model produced a given output, classifying source models across 13 model families through black-box probing and probabilistic attribution. Presented at the Canadian Statistics Student Conference (Statistical Society of Canada).

• —Probabilistic attribution method to identify language-model sources using black-box probing and statistical classification across 13 model families.
• —90%+ attribution accuracy on held-out data, outperforming baseline heuristics and providing uncertainty estimates for ambiguous cases.
• —Generated attribution scores for 10K+ prompts across GPT, Claude, Llama, Gemini, Mistral and other families, supporting prompt-injection analysis and model fingerprinting.