Reproduce “Generate-then-Rank: Learning Symbolic Goal Interpretation for Embodied Agents” 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
Jisen Lei, Huizhi Liang
The authors train a language model to translate instructions into symbolic goals describing terminal object states, directed relations, and actions, then rank generated alternatives. On a repeatedly inspected 68-item diagnostic, SFT+CORE scores 0.736 Combined F1 and 0.642 Edge F1, compared with greedy SFT’s 0.716 and 0.606. These results measure goal matching, not downstream execution, and do not establish CORE’s causal effect.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.