Reproduce “Autoregressive Decoding as a Martingale: Characterizing the Stochastic Dynamics of LLM Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiahui Zhu, Zhenyi Qi, Wanwei Liu, Haoang Chi, Yang Tan, Xinyu Sun, Ji Wang
The authors characterize how prediction errors accumulate during text generation. At each step, they subtract the model’s next-token probability vector from the one-hot vector of the sampled token, forming cumulative errors with zero conditional mean. They show that the expected squared size at stopping equals an expected sum based on sampled-token probabilities, and experiments find close agreement under tested conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.