Research map / Vision-language and multimodal / Visual question answering
Visual Programming: Compositional Visual Reasoning Without Training
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 |
|---|---|---|---|
| majority voting across 5 random prompt runs instead of a single random prompt run (ablation) | VISPROG GQA · GQA testdev subset | accuracy: 48.2 to 50.5 | Helped |
| curated 20-example prompt instead of random sampling (ablation) | VISPROG GQA · GQA testdev subset | accuracy: 48.2 to 50.0 | Helped |
| VISPROG with in-context LLM program generation vs. ViLT VQA alone | GQA · GQA testdev subset | accuracy: 47.8 to 50.5 | Helped |
| majority voting across 5 random prompt runs instead of a single random prompt run (ablation) | VISPROG NLVR · NLVR test | accuracy: 61.3 to 62.4 | Helped |
| curated 12-example prompt instead of random sampling (ablation) | VISPROG NLVR · NLVR test | accuracy: 61.3 to 61.8 | Helped |
| modifying instructions after inspecting visual rationales (instruction tuning) (ablation) | VISPROG knowledge tagging · Knowledge tagging | tagging F1: 63.7 to 75.7 | Helped |
| modifying instructions after inspecting visual rationales (instruction tuning) (ablation) | VISPROG knowledge tagging · Knowledge tagging | localization F1: 80.6 to 84.9 | Helped |
| modifying instructions after inspecting visual rationales (instruction tuning) (ablation) | VISPROG image editing · Image editing | accuracy: 59.8 to 66.4 | Helped |
| zero-shot VISPROG with single-image ViLT VQA and LLM reasoning vs. NLVR-finetuned ViLT upper bound | VISPROG NLVR · NLVR test | accuracy: 62.4 vs 76.3 | Mixed |
Closest papers by meaning
- NePTune: A Neuro-Pythonic Framework for Tunable Compositional Reasoning on Vision-Language (ICLR 2026 · similarity 0.92)
- Synthesizing Visual Concepts as Vision-Language Programs (CVPR 2026 · similarity 0.91)
- GENOME: Generative Neuro-Symbolic Visual Reasoning by Growing and Reusing Modules (ICLR 2024 · similarity 0.90)
- Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement (CVPR 2024 · similarity 0.89)
- VisRes Bench: On Evaluating the Visual Reasoning Capabilities of VLMs (CVPR 2026 · similarity 0.89)
- Visual Program Distillation: Distilling Tools and Programmatic Reasoning into Vision-Language Models (CVPR 2024 · similarity 0.88)
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.