# Attributions All the Way Down? The Metagame of Interpretability

Source: [Attributions All the Way Down? The Metagame of Interpretability](<https://www.sota2.com/research/paper/69fd6e215497b3b670c894d8>)

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

Data updated: 2026-05-09T05:00:12.752Z

Authors: Hubert Baniecki, Przemyslaw Biecek, Fabian Fumagalli

Published: 2026-05-07T13:59:26.000Z

## Abstract

We introduce the metagame, a conceptual framework for quantifying second-order interaction effects of model explanations. For any first-order attribution $\\phi(f)$ explaining a model $f$, we measure the directional influence of feature $j$ on the attribution of feature $i$, denoted as meta-attribution $\\varphi\_{j \\to i}(f)$, by treating the attribution method itself as a cooperative game and computing its Shapley value. Theoretically, we prove that attributions hierarchically decompose into meta-attributions, and establish these as directional extensions of existing interaction indices. Empirically, we demonstrate that the metagame delivers insights across diverse interpretability applications: (i) quantifying token interactions in instruction-tuned language models, (ii) explaining cross-modal similarity in vision-language encoders, and (iii) interpreting text-to-image concepts in multimodal diffusion transformers.

## 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.

- [Interaction Recognition on ImageNet-1k 2 objects](<https://www.sota2.com/research/sota/interaction-recognition-on-imagenet-1k-2-objects.md>): Interaction Recognition; ImageNet-1k 2 objects; Interaction Recognition: 88.

- [Interaction Recognition on ImageNet 1k (3 objects)](<https://www.sota2.com/research/sota/interaction-recognition-on-imagenet-1k-3-objects.md>): Interaction Recognition; ImageNet 1k (3 objects); Interaction Recognition Accuracy: 90.

- [Interaction Recognition on ImageNet-1k 4 objects](<https://www.sota2.com/research/sota/interaction-recognition-on-imagenet-1k-4-objects.md>): Interaction Recognition; ImageNet-1k 4 objects; Interaction Recognition Accuracy: 91.

- [Image Segmentation on Pascal VOC](<https://www.sota2.com/research/sota/image-segmentation-on-pascal-voc.md>): Image Segmentation; Pascal VOC; Accuracy: 90.3.

- [Image Segmentation on MS COCO (Acc, mIoU, mAP)](<https://www.sota2.com/research/sota/image-segmentation-on-ms-coco-acc-miou-map.md>): Image Segmentation; MS-COCO; Accuracy: 88.9.

- [Interaction Recognition on ImageNet-1k 1 object](<https://www.sota2.com/research/sota/interaction-recognition-on-imagenet-1k-1-object.md>): Interaction Recognition; ImageNet-1k 1 object; Interaction Recognition: 89.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2605.06295>)

[PDF](<https://arxiv.org/pdf/2605.06295.pdf>)