How can LLMs explain their reasoning to humans?
Implementation and Application of an Intelligibility Protocol for Interaction with an LLM
October 29, 2024
https://arxiv.org/pdf/2410.20600This research explores a communication protocol (PXP) for interactions between human experts and large language models (LLMs) in multi-agent systems.
The paper implements a simplified version of PXP, demonstrating its use in tasks like X-ray diagnosis and molecule synthesis. Experiments show promising results, highlighting PXP's ability to:
- Facilitate "intelligible" communication between humans and LLMs, enabling them to understand and refine each other's predictions and explanations.
- Improve LLM performance by incorporating human feedback through iterative message exchanges.
- Quantify the level of intelligibility in human-LLM interactions.
The research suggests PXP can be a valuable tool for building collaborative human-LLM systems for complex data analysis.