FAX Framework Improves Faithfulness in Agentic XAI Systems
A newly developed framework known as Faithful Agentic XAI (FAX) seeks to improve the reliability of explanations produced by agentic explainable AI systems utilizing large language models (LLMs). These AI systems often generate seemingly credible yet untrustworthy explanations, particularly when LLMs amplify unreliable outputs from complex models. FAX tackles this issue by breaking down draft explanations into individual claims and validating them against inherently reliable tools, thereby filtering out unsupported or contradictory statements prior to final output. Additionally, the researchers have created CRAFTER-XAI-Bench, an open-world reinforcement learning benchmark featuring intricate policies, varied objectives, and demanding scenarios to evaluate model-specific faithfulness. Using this benchmark, FAX enhanced simulation faithfulness from 0.20 to a higher, unspecified value. This research is available on arXiv under ID 2605.27879.
Key facts
- FAX stands for Faithful Agentic XAI
- FAX decomposes draft explanations into claims
- Claims are cross-checked against inherently faithful tools
- Unsupported or contradictory claims are filtered out
- CRAFTER-XAI-Bench is an open-world reinforcement learning benchmark
- Benchmark includes complex policies, diverse goals, and challenging scenarios
- FAX improved simulation faithfulness from 0.20 on CRAFTER-XAI-Bench
- Paper is on arXiv with ID 2605.27879
Entities
Institutions
- arXiv