Reproduce “Calibration, Entropy Rates, and Memory in Language Models” 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
Mark Braverman, Xinyi Chen, Sham M. Kakade, Karthik Narasimhan, Cyril Zhang, Yi Zhang
The authors measure how language models’ generated-text uncertainty changes during sampling and propose a local calibration step that reweights next-word choices using a one-step lookahead. For GPT-2 on WebText, Test perplexity is 23.7 and generated-text EntRate is 61.2; its entropy estimate uses New York Times articles as a proxy validation set. The reported quantitative calibration result is for an LSTM.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.