Attributions All the Way Down? The Metagame of Interpretability
About
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.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Interaction Recognition | ImageNet-1k 2 objects | Interaction Recognition88 | 26 | |
| Interaction Recognition | ImageNet 1k (3 objects) | Interaction Recognition Accuracy90 | 26 | |
| Interaction Recognition | ImageNet-1k 4 objects | Interaction Recognition Accuracy91 | 26 | |
| Interaction Recognition | ImageNet-1k 1 object | Interaction Recognition89 | 26 | |
| Image Segmentation | Pascal VOC | Accuracy90.3 | 4 | |
| Image Segmentation | MS-COCO | Accuracy88.9 | 4 |