Aphanta: Diagnosing Task-Aligned Image-Edited Intermediates for Multimodal Reasoning
Aphanta tests when edited images actually help multimodal models reason, and finds the gains are narrow.
The framework compares direct reasoning, editor-made visual intermediates, and ideal reference intermediates in the same pipeline. Across 20 candidate tasks, benefits cluster in visual cue injection, grounding, and counterfactual state changes. Tasks needing symbol-sensitive construction or structural extrapolation are much less reliable. In the selected positive subset, the consolidated Qwen pipeline raises mean score from 0.343 to 0.445. HF Daily Papers' note
The framework compares direct reasoning, editor-made visual intermediates, and ideal reference intermediates in the same pipeline. Across 20 candidate tasks, benefits cluster in visual cue injection, grounding, and counterfactual state changes. Tasks needing symbol-sensitive construction or structural extrapolation are much less reliable. In the selected positive subset, the consolidated Qwen pipeline raises mean score from 0.343 to 0.445. HF Daily Papers' note
score 4