Reproduce “Prune 'n Predict: Optimizing LLM Decision-making with Conformal Prediction” on OpenResearch

About this paper

Harit Vishwakarma, Alan Mishler, Thomas Cook, Niccolò Dalmasso, Natraj Raman, Sumitra Ganesh

Researchers at JPMorganChase AI Research developed "Prune 'n Predict," a framework combining Conformal Revision of Questions (CROQ) with Conformal Prediction Optimization (CP-OPT), to improve Large Language Model accuracy by pruning multiple-choice options. The approach significantly reduces prediction set sizes while maintaining statistical coverage, leading to enhanced LLM performance and cost efficiency in high-stakes decision-making tasks.

Key insights

• Optimizing the score function within Conformal Prediction (CP-OPT) significantly reduces prediction set sizes while maintaining robust coverage guarantees, providing more precise and actionable uncertainty quantification for LLMs.
• Applying a 'process of elimination' strategy to LLMs through Conformal Revision of Questions (CROQ) directly enhances their decision-making accuracy by leveraging uncertainty estimates to prune incorrect choices.
• The integration of optimized uncertainty quantification (CP-OPT) with an iterative question revision strategy (CROQ) offers a principled and effective method to enhance LLM trustworthiness and operational efficiency in high-stakes scenarios.

Summary by alphaXiv.