Reproduce “Policy Learning with a Natural Language Action Space: A Causal Approach” 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
Bohan Zhang, Yixin Wang, Paramveer S. Dhillon
The authors present a Q-learning method for sequential text rewriting. It optimizes sentence embeddings using task-specific classifiers as Q-functions, then decodes them into text. On constructed two-stage Yelp tasks, TTS reaches 90.5% transfer strength, versus 88.0% for DeleteAndRetrieve. This classifier-based score is not evidence of real-world sentiment change or causal impact.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.