Reproduce “Markovian Transformers for Informative Language Modeling” 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
Scott Viteri, Max Lamparth, Peter Chatain, Clark Barrett
The paper trains a Markovian model to generate a bounded-length chain of thought (CoT) from a question, then predict the answer from that CoT alone. This constraint makes the CoT the answer predictor’s only input, prioritizing answer-relevant information over a record of internal computation. On GSM8K, Llama 3.1 8B accuracy rises from 19.6% to 57.1%; this does not establish that CoTs mirror internal reasoning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.