Research map / Vision-language and multimodal / Human and embodied modeling
Visual Reasoning by Progressive Module Networks
Seung Wook Kim, Makarand Tapaswi, Sanja Fidler
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 |
|---|---|---|---|
| Compose VQA with object and attribute modules (ablation) | VQA baseline · VQA 2.0 val | accuracy: 62.05 -> 63.38 | Improved |
| Add relationship module to the VQA composition (ablation) | PMN M_vqa1 · VQA 2.0 val | accuracy: 63.38 -> 63.64 | Improved |
| Add counting module to the VQA composition (ablation) | PMN M_vqa1 · VQA 2.0 val | accuracy: 63.38 -> 64.06 | Improved |
| Add relationship and counting modules (ablation) | PMN M_vqa1 · VQA 2.0 val | accuracy: 63.38 -> 64.36 | Improved |
| Add captioning module to the full VQA composition (ablation) | PMN M_vqa4 · VQA 2.0 val | accuracy: 64.36 -> 64.68 | Improved |
| Replace learned soft gating with equal importance scores (ablation) | PMN M_vqa5 · VQA 2.0 val | accuracy: 64.68% -> 63.65% | Hurt |
| Increase capacity without composing modules | VQA baseline · VQA 2.0 val | accuracy: 62.05% -> 62.0% | No clear effect |
| Add relationship-derived question-answer training data to M_vqa1 | VQA M_vqa1 · VQA 2.0 val | accuracy: 63.38% -> 63.05% | Hurt |
| Compose relationship detection with object and attribute modules (ablation) | Relationship detection · Visual Genome relationship validation | Object accuracy (%): 51.0 -> 53.4 | Improved |
| Compose relationship detection with object and attribute modules (ablation) | Relationship detection · Visual Genome relationship validation | Subject accuracy (%): 55.9 -> 57.8 | Improved |
| Add object and attribute modules to counting (ablation) | Counting baseline · Object counting | accuracy (%): 45.4 -> 47.4 | Improved |
| Add the relationship module to counting with object and attribute modules (ablation) | Counting M_cnt1 · Object counting | accuracy (%): 47.4 -> 50.0 | Improved |
Closest papers by meaning
- GENOME: Generative Neuro-Symbolic Visual Reasoning by Growing and Reusing Modules (ICLR 2024 · similarity 0.89)
- DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning (ICLR 2026 · similarity 0.88)
- ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both (NeurIPS 2026 · similarity 0.88)
- From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.88)
- CogFlow: Bridging Perception and Reasoning through Knowledge Internalization for Visual Mathematical Problem Solving (ICLR 2026 · similarity 0.87)
- SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · 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.
