Reproduce “CoFineLLM: Conformal Finetuning of LLMs for Language-Instructed Robot Planning” on OpenResearch

About this paper

Jun Wang, Yevgeniy Vorobeychik, Yiannis Kantaros

This paper introduces CoFineLLM, a framework that integrates Conformal Prediction into the fine-tuning process of Large Language Models for language-instructed robot planning. The method enables LLMs to produce more efficient and precise action prediction sets, substantially reducing the need for human intervention while maintaining probabilistic coverage guarantees for robot plans.

Key insights

• Integrating uncertainty awareness directly into the LLM's training process is crucial for optimizing prediction set efficiency and significantly reducing human intervention, beyond post-hoc application of Conformal Prediction.
• A conditional loss design, which prioritizes task correctness before optimizing for prediction set size, effectively overcomes the vanishing gradient problem encountered by prior conformal training methods when fine-tuning overconfident LLMs.
• Dynamically updating the Conformal Prediction threshold during training enables the LLM to become 'conformal-aware,' leading to more robust and calibrated uncertainty estimates that generalize across varying confidence levels and out-of-distribution scenarios.

Summary by alphaXiv.