Persona-Conditioned RL Adapts AI Explanations to Expert Perspectives
A new framework has been developed by researchers to create explainable AI outputs that cater to the unique viewpoints of individual experts, tackling the issue that the quality of explanations can vary. This approach, known as perspective-conditioned explanations, employs knowledge graph reasoning paths in drug discovery to illustrate how expert preferences align into distinct epistemic perspectives. These perspectives are represented by agentic personas, reflecting how experts assess explanations. Rewards aligned with these personas steer the generation of explanations through reinforcement learning, minimizing the necessity for extensive expert oversight. Studies involving expert users show a preference for perspective-conditioned explanations over generic ones, enhancing perceived relevance and validity, and achieving or surpassing state-of-the-art results. This research is detailed in a paper available on arXiv (2603.21846v2), which has been announced as a replacement.
Key facts
- Framework adapts explanation generation to epistemic variation in expert judgment.
- Uses knowledge graph reasoning paths in drug discovery.
- Preferences organize into coherent epistemic perspectives captured by agentic personas.
- Persona-aligned rewards guide reinforcement learning-based explanation generation.
- Reduces need for large-scale expert supervision.
- Expert user studies show preference for perspective-conditioned explanations.
- Improves perceived relevance and validity.
- Matches or exceeds state-of-the-art performance.
- Paper available on arXiv (2603.21846v2).
Entities
Institutions
- arXiv