Reproduce “Active Preference Inference using Language Models and Probabilistic Reasoning” 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
Wasu Top Piriyakulkij, Volodymyr Kuleshov, Kevin Ellis
An algorithm developed by researchers at Cornell University augments Large Language Models with probabilistic reasoning to efficiently infer user preferences. This method achieves higher information gain per question and requires fewer user interactions to identify preferences compared to instruction-tuned or ReAct LLMs.
Key insights
- While LLMs can generate diverse questions and estimate answer plausibility, they are inherently inefficient at selecting the most informative questions for preference inference.
- Combining LLMs with external probabilistic reasoning enables the application of principled information-theoretic criteria to optimize question selection.
- In this specific probabilistic model, Expected Entropy Minimization and Expected Model Change Maximization are theoretically equivalent objectives for choosing informative questions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.