New framework InterPol de-anonymizes LM Arena responses

A new research paper introduces InterPol, a framework designed to de-anonymize language model responses on leaderboards like LM Arena. Unlike previous methods that relied on simple statistical features, InterPol synthesizes interpolated preference data to capture deeper stylistic patterns. Experiments show InterPol significantly outperforms existing baselines in identification accuracy and demonstrates a real-world threat through ranking manipulation simulations. AI

IMPACT Highlights potential vulnerabilities in LLM leaderboards and the need for more robust anonymity measures.

RANK_REASON Research paper detailing a new method for de-anonymizing language model responses. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• bag-of-words model
• DagsHub
• Hugging Face
• InterPol
• LM Arena
• Minsung Cho
• tf–idf

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →