Reproduce “Propose to Learn, Learn to Propose: Evaluability-Aware Assistance under Bounded Rationality” 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
Yifan Zhu, Sammie Katt, Samuel Kaski
PROSE-PLAN treats each proposal as both a possible task change and evidence about user preferences and evaluability. It models acceptance from value gain, distance-dependent evaluation burden, and response sharpness, then plans a second proposal under hypothetical response updates. In controlled Probe-Commit simulations, it reached preferred goals in 0.815 of episodes, versus 0.555 for belief-frozen depth-2 planning, not real-user evidence.
Key insights
- Proposals most informative about a user's evaluability often have a higher probability of being rejected, challenging the notion of always maximizing immediate acceptance.
- Strategic "probing" proposals, even if they have low immediate value or are expected to be rejected, can be crucial for learning about user preferences and evaluability.
- An AI assistant must plan beyond immediate acceptance, anticipating how user responses (including rejections) provide information that refines future proposal strategies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.