Research map / Vision-language and multimodal / Human and embodied modeling
Attention over Learned Object Embeddings Enables Complex Visual Reasoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use Aloe instead of DCL | CLEVRER · CLEVRER | accuracy across question types: 90.7 -> 94.0; 82.8 -> 96.0; 82.0 -> 87.5; 46.5 -> 75.6 | Helped |
| Use Aloe instead of NS-OPT | ACRE · ACRE compositional and systematic splits | All accuracy: 69.04 -> 91.76; 67.44 -> 93.90 | Helped |
| Use Aloe instead of NS-OPT | ACRE · ACRE systematic split, backward-blocked | accuracy: 16.06 -> 94.48 | Helped |
| Use Aloe instead of CNN-BERT | ACRE · ACRE systematic split, screen-off | accuracy: 0.00 -> 98.97 | Helped |
| Replace self-attention with an MLP (ablation) | Aloe · CLEVRER | accuracy: 94.0 -> 45.4; 96.0 -> 16.0; 87.5 -> 27.7; 75.6 -> 9.9 | Hurt |
| Replace MONet object representations with ResNet hyperpixels (ablation) | Aloe · CLEVRER | accuracy: 94.0 -> 74.9; 96.0 -> 66.1; 87.5 -> 58.3; 75.6 -> 32.4 | Hurt |
| Replace global cross-frame object attention with hierarchical attention (ablation) | Aloe · CLEVRER | accuracy: 94.0 -> 80.6; 96.0 -> 87.4; 87.5 -> 73.5; 75.6 -> 55.1 | Mixed |
| Add the auxiliary self-supervised loss (ablation) | Aloe · CLEVRER | accuracy: 91.0 -> 94.0; 92.8 -> 96.0; 82.8 -> 87.5; 68.7 -> 75.6 | Helped |
| Use L2 rather than contrastive auxiliary loss (ablation) | Aloe · CLEVRER | accuracy across question types | Helped |
| Mask one object per frame rather than other tested masking schemes (ablation) | Aloe · CLEVRER | accuracy across question types | Helped |
| Use non-local temporal prediction or infilling instead of local adjacent-target prediction (ablation) | Aloe · CLEVRER | accuracy | Hurt |
| Vary contrastive-loss negatives, temperature, or similarity function (ablation) | Aloe · CLEVRER | accuracy | No clear effect |
Closest papers by meaning
- Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.88)
- Train the Agent, Not the Expert: Learning to Harness Heterogeneous Experts for Multi-Turn Visual Reasoning (NeurIPS 2026 · similarity 0.87)
- Look, Remember and Reason: Grounded Reasoning in Videos with Language Models (ICLR 2024 · similarity 0.87)
- Latent Implicit Visual Reasoning (CVPR 2026 · similarity 0.87)
- VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.87)
- Towards flexible perception with visual memory (ICML 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.