MEA: multi-agent reward optimization lifts ML explanation faithfulness by up to 34%
UVABiology · hf · 2026-10-06
Researchers propose MEA, a multi-agent framework that removes the expertise barrier to ML explainability. A Proposer agent selects and configures explanation tools based on question and modality, while an Actor agent is optimized end-to-end against faithfulness to produce natural-language explanations grounded in model behavior across tabular, text, and vision data. The work introduces question types spanning feature attribution, counterfactual reasoning, and spurious feature detection, each with perturbation-based faithfulness metrics. Notably, frontier LLMs are found to systematically produce unfaithful explanations. With reward-driven optimization plus a modality-adaptive penalty, MEA beats post-hoc, agentic, and closed-source baselines across six datasets, gaining +28% (tabular), +21% (text), and +34% (vision) faithfulness over the untrained backbone.
More from coding & agent
- Brainbase partners with Stripe on Agentic Payments, letting agents buy things for you — garrytan · 2026-10-07
- Every launches Slack-native AI coworker that files PRs, reports and auto-organizes tickets — danshipper · 2026-10-07
- DeskForge generates 1.2M annotated desktop observations, lifting GUI grounding accuracy by 11.5 points — docling-project · 2026-10-07
- 10 lessons from 15 Grokbot guides: one chief-of-staff bot beats twenty chats — leveragedupside · 2026-10-07
- Cortex Harness lets frontier models run real robots, with human takeovers feeding training data — kimmonismus · 2026-10-07
- Blog: Multi-agent systems aren't a new scaling axis, just structured inference scaling — willcb · 2026-10-07