Did OpenAI’s “encrypted / hidden reasoning” just get silently dropped? In a simple number‑guess + hash‑commit game: model promised a hash(n:salt) commitment, then it admitted it does't remember (n,salt). #OpenAI

Watermarking LLMs is crucial for traceability, and speculative sampling accelerates generation.

Can we have the best of both worlds?🤔

Our research shows it's not possible😮—we prove a no-go theorem and propose methods to navigate this trade-off. #NeurIPS2024

nips.cc/virtual/2024...

#EMNLP2024 Accepted to EMNLP 2024 Main Conference 🎉

”A Bayesian Approach to Harnessing the Power of LLMs in Authorship Attribution“

It works so well that a 7B open-source LLM outperforms GPT-4 with prompting at identifying unique writing style. 🔍

👉 Learn more: arxiv.org/pdf/2410.217...