LEMON: Local Explanations via Modality-aware OptimizatioN
Yu Qin, Phillip Sloan, Raul Santos-Rodriguez, Majid Mirmehdi, Telmo de Menezes e Silva Filho · 2026
Read LEMON: Local Explanations via Modality-aware OptimizatioN by Yu Qin, Phillip Sloan, Raul Santos-Rodriguez, Majid Mirmehdi, Telmo de Menezes e Silva Filho (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-02786.txt.
Description
Yu Qin and 4 co-authors explain multimodal model predictions with one surrogate that respects each modality's distinct contribution, in a 2026 paper. Multimodal models are everywhere, but explanation methods tend to be single-modal, tied to one design, or too expensive to run at scale.
LEMON fits a single modality-aware surrogate model with group-structured sparsity around each prediction. The result separates how much each modality contributed from which features inside it mattered. The predictor is treated purely as a black box. Few forward passes are needed, and explanations stay faithful under repeated nudges. Testing spans vision-language question answering and a clinical task mixing image, text, and tabular inputs. Against strong multimodal rivals, faithfulness stayed competitive while black-box tests dropped by tens of times and runtime fell several-fold. Local explanation at fleet scale becomes workable.
Catalog record
| CREATOR | Yu Qin, Phillip Sloan, Raul Santos-Rodriguez, Majid Mirmehdi, Telmo de Menezes e Silva Filho |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-02786