Research map / Vision-language and multimodal / Multimodal LLMs
Natural Image Pretraining Improves Abstract Reasoning
Xiaoman Ding, Keya Hu, Katelyn Gan, Victor Yin, Kaiming He
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| initialize encoder from ImageNet MAE instead of random init (ablation) | Nat-ARC (VARC pipeline) · ARC-1 eval (base, large, huge) | pass@2 | Improved |
| add ImageNet MAE pretraining during offline training (ablation) | Nat-ARC · ARC-1 train offline eval | convergence speed and accuracy | Improved |
| scale encoder from base to large to huge without pretraining (ablation) | Nat-ARC from scratch · ARC-1 eval | pass@2 | Mixed |
| scale encoder size with ImageNet pretraining (ablation) | Nat-ARC · ARC-1 eval | pass@2 | Improved |
| ensemble two non-pretrained checkpoints vs single (ablation) | Nat-ARC scratch · ARC-1 eval | pass@2 | Improved |
| include ImageNet-pretrained checkpoints in ensemble vs scratch-only (ablation) | Nat-ARC ensemble · ARC-1 eval | pass@2 | Improved |
| two-ImageNet vs mixed scratch-ImageNet ensemble (ablation) | Nat-ARC ensemble · ARC-1 eval | pass@2 | No clear effect |
| add ARC-style grid MAE pretraining (ablation) | Nat-ARC · ARC-1 eval | pass@2 | Improved |
| scale ARC-style pretrained model from large to huge (ablation) | Nat-ARC ARC-style pretrain · ARC-1 eval | pass@2 | No clear effect |
| pool predictions from no-pretrain, ImageNet and ARC-style regimes (ablation) | Nat-ARC ensemble · ARC-1 eval | pass@2: 63.4 -> 70.2 | Improved |
| enable RoPE with ImageNet MAE init (base scale) (ablation) | Nat-ARC base · ARC-1 offline eval | accuracy | No clear effect |
Closest papers by meaning
- Moving Beyond More Views: Redundancy-Aware Ego–Exo Fusion for Proficiency Estimation (ECCV 2026 · similarity 0.97)
- Learning Egocentric Cues from Exocentric Video using Privileged Egocentric Supervision (ECCV 2026 · similarity 0.90)
- EgoExo-Con: Exploring View-Invariant Video Temporal Understanding (ECCV 2026 · similarity 0.89)
- EgoExoBench: A Benchmark for First- and Third-person View Video Understanding in MLLMs (NeurIPS 2025 · similarity 0.88)
- Bootstrap Your Own Views: Masked Ego-Exo Modeling for Fine-grained View-invariant Video Representations (CVPR 2025 · similarity 0.88)
- Viewpoint Rosetta Stone: Unlocking Unpaired Ego-Exo Videos for View-invariant Representation Learning (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
