MEA: A Reward-Driven Multi-Agent System for Faithful Model Explanations
MEA trains a multi-agent system to produce more faithful natural-language explanations across tabular, text, and vision models.
MEA is a multi-agent framework that turns post-hoc ML explanations into natural language without requiring explainability expertise. A Proposer selects and configures tools by question and modality, while an Actor is optimized end-to-end for faithfulness across tabular, text, and vision. Question types cover feature attribution, counterfactual reasoning, and spurious-feature detection, each with a perturbation-based metric. Reward training beats post-hoc, agentic, and closed-source baselines, with gains of 28% (tabular), 21% (text), and 34% (vision), and the authors find frontier LLMs systematically unfaithful.
- Proposer selects tools; an Actor is trained on faithfulness rewards.
- Questions cover attribution, counterfactuals, and spurious-feature detection.
- Faithfulness rises 28% tabular, 21% text, and 34% vision.
- Frontier LLMs systematically produce unfaithful explanations.
From the original article
1 min read at huggingface.co
Recent years have seen the employment of a plethora of machine learning (ML) models in high-stakes domains, but they remain largely opaque to the practitioners who act on their predictions.
By Yuyang Cheng, Raghav Kaushik Ravi, Srivarshinee Sridhar · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.