# Interpretability Guarantees with Merlin-Arthur Classifiers

Source: [Interpretability Guarantees with Merlin-Arthur Classifiers](<https://www.sota2.com/research/paper/68fa729e1e209131cea57953>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-26T02:30:51.475Z

Authors: Stephan W\\"aldchen, Kartikey Sharma, Berkant Turan, Max Zimmer, Sebastian Pokutta

Published: 2022-06-01T20:48:24.000Z

## Abstract

We propose an interactive multi-agent classifier that provides provable interpretability guarantees even for complex agents such as neural networks. These guarantees consist of lower bounds on the mutual information between selected features and the classification decision. Our results are inspired by the Merlin-Arthur protocol from Interactive Proof Systems and express these bounds in terms of measurable metrics such as soundness and completeness. Compared to existing interactive setups, we rely neither on optimal agents nor on the assumption that features are distributed independently. Instead, we use the relative strength of the agents as well as the new concept of Asymmetric Feature Correlation which captures the precise kind of correlations that make interpretability guarantees difficult. We evaluate our results on two small-scale datasets where high mutual information can be verified explicitly.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Image Classification on CLEVR-Hans3](<https://www.sota2.com/research/sota/image-classification-on-clevr-hans3.md>): Image Classification; CLEVR-Hans3; Completeness (Accuracy): 96.59.

- [Image Classification on CLEVR-Hans7](<https://www.sota2.com/research/sota/image-classification-on-clevr-hans7.md>): Image Classification; CLEVR-Hans7; Completeness Accuracy: 97.61.

- [Image Classification on CIFAR-100 (Completeness and Soundness)](<https://www.sota2.com/research/sota/image-classification-on-cifar-100-completeness-and-soundness.md>): Image Classification; CIFAR-100; Completeness Accuracy: 15.27.

- [Image Classification on ImageNet-1k (Accuracy and Robustness Metrics)](<https://www.sota2.com/research/sota/image-classification-on-imagenet-1k-accuracy-and-robustness-metrics.md>): Image Classification; ImageNet-1K; Completeness (Accuracy): 35.06.

- [Image Classification on COCOLogic](<https://www.sota2.com/research/sota/image-classification-on-cocologic.md>): Image Classification; COCOLogic; Completeness (Accuracy): 42.57.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2206.00759>)

[PDF](<https://arxiv.org/pdf/2206.00759.pdf>)