Reproduce “GIFT: Generalizing Intent for Flexible Test-Time Rewards” 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

Fin Amin, Nathaniel Dennler, Andreea Bobu

GIFT (Generalizing Intent for Flexible Test-Time Rewards) is a framework that allows robot reward functions, learned from human demonstrations, to generalize to novel environments by grounding state similarity in high-level human intent. It leverages large language models to infer intent from contrasting behaviors and subsequently align novel test states to intent-equivalent training states, achieving up to 20% higher win rates in complex tasks compared to baselines relying on surface-level features.

Key insights

  • Meaningful state similarity for robust reward generalization is fundamentally driven by high-level human intent, moving beyond low-level visual or linguistic features.
  • Large Language Models can effectively deduce abstract human intentions by contrasting preferred and non-preferred robot behaviors, serving as powerful intent estimators.
  • LMs can act as implicit, intent-conditioned similarity functions, enabling the mapping of novel test-time semantic components to their functionally equivalent training-time counterparts.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.