Research map / Vision-language and multimodal / Visual question answering
Recursive Visual Programming
Jiaxin Ge, Sanjay Subramanian, Baifeng Shi, Roei Herzig, Trevor Darrell
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 |
|---|---|---|---|
| add recursive_query (recursive code generation) to ViperGPT | ViperGPT · VSR zero-shot split | accuracy: 61.59 -> 66.09 | Improved |
| add recursive_query to ViperGPT | ViperGPT · GQA test-dev | accuracy: 44.63 -> 45.62 | Improved |
| specify return type explicitly in recursive_query (dynamic-explicit vs non-recursive) (ablation) | RVP · GQA (recursive-pattern instances) | accuracy: 55.99 -> 70.23 | Improved |
| let recursive_query return varying types without specifying them in the query (implicit) (ablation) | RVP · COVR val_1000 (recursive-pattern instances) | accuracy: 60.71 -> 53.57 | Hurt |
| fixed-type (string) recursive_query vs non-recursive (ablation) | RVP · GQA vs COVR val_1000 | accuracy: 55.99 -> 65.05 (GQA); 60.71 -> 60.12 (COVR) | Mixed |
| use retrieval-based selection of recursive in-context examples (ablation) | RVP · GQA val_2000 | accuracy: 54.75 -> 53.79 | No clear effect |
| add or remove one in-context example (ablation) | RVP · COVR val_1000 | accuracy: 51.05 -> 51.35 (add 1) | No clear effect |
| replace gpt-3.5-turbo with CodeLlama-13B as code generator (ablation) | RVP · GQA val_2000 | accuracy: 54.75 -> 49.27 | Hurt |
| apply NLP prompting methods (CoT, PoT, sub-functions) instead of recursion (ablation) | visual programming · GQA val_2000 | accuracy: 54.75 (RVP) vs 48.42-52.03 | Hurt |
| recursive decomposition on more complex GQA questions | RVP · GQA 2-step relates | accuracy: 32.37 -> 40.29 (+7.92) | Improved |
| recursive decomposition on single-property GQA questions | RVP · GQA 1-step properties | accuracy: 64.62 -> 62.45 (-2.17) | Hurt |
| recursive code on Compare questions in GQA | RVP · GQA test-dev (Compare) | accuracy: 66.67 -> 53.33 | Hurt |
Closest papers by meaning
- ViUniT: Visual Unit Tests for More Robust Visual Programming (CVPR 2025 · similarity 0.88)
- VQA With No Questions-Answers Training (CVPR 2020 · similarity 0.88)
- VQ-VA World: Towards High-Quality Visual Question-Visual Answering (CVPR 2026 · similarity 0.88)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.87)
- GENOME: Generative Neuro-Symbolic Visual Reasoning by Growing and Reusing Modules (ICLR 2024 · similarity 0.87)
- Rephrase, Augment, Reason: Visual Grounding of Questions for Vision-Language Models (ICLR 2024 · similarity 0.86)
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.
