Research map / Vision-language and multimodal / Human and embodied modeling
One Self-Configurable Model to Solve Many Abstract Visual Reasoning Problems
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SCAR, a single unified model for single-choice abstract visual reasoning tasks (RPMs, VAPs, O3) built on a Structure-Aware dynamic Layer whose weights adapt to the task's panel layout, evaluated in STL, MTL and TL settings.
Problem
Abstract visual reasoning tasks (RPMs, visual analogy problems, odd-one-out tests) are usually solved by task-specific models with architectures hard-coded to a particular number and layout of panels, so no general solver exists.
Approach
A Structure-Aware dynamic Layer (SAL) builds its linear weight matrix on the fly by sliding-window unfolding and averaging a shared large weight tensor according to the task's grid structure (r x c), so one model handles panel counts and layouts without assumptions; it is paired with a multi-head variant and an auxiliary rule-prediction loss.
Key results
- SCAR reaches 86.7% on O3 STL, the best among universal and task-specific models in that table, and 92.9% on VAP-S.
- On RPMs SCAR is on par with task-specific SCL (I-RAVEN 94.7% vs 95.5%; PGM-S 46.1% vs 51.1%) and beats CoPINet and SRAN.
- In transfer learning R+V to O3, SCAR reaches 89.0% (+2.3 p.p. over STL), versus 75.5% for SCL+SAL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAL instead of RN in SCAR (ablation) | SCAR · G-set | test accuracy (%): 36.5 -> 82.1 | Helped |
| Use SAL instead of RN in SCAR (ablation) | SCAR · PGM-S | test accuracy (%): 33.2 -> 46.1 | Helped |
| Use SAL instead of RN in SCAR (ablation) | SCAR · I-RAVEN | test accuracy (%): 56.1 -> 94.7 | Helped |
| Use SAL instead of RN in SCAR (ablation) | SCAR · VAP-S | test accuracy (%): 90.5 -> 92.9 | Helped |
| Use SAL instead of RN in SCAR (ablation) | SCAR · O3 | test accuracy (%): 69.5 -> 86.7 | Helped |
| Use SAL instead of LSTM in SCAR (ablation) | SCAR · G-set | test accuracy (%): 43.5 -> 82.1 | Helped |
| Use SAL instead of LSTM in SCAR (ablation) | SCAR · PGM-S | test accuracy (%): 26.0 -> 46.1 | Helped |
| Use SAL instead of LSTM in SCAR (ablation) | SCAR · I-RAVEN | test accuracy (%): 41.7 -> 94.7 | Helped |
| Use SAL instead of LSTM in SCAR (ablation) | SCAR · VAP-S | test accuracy (%): 90.7 -> 92.9 | Helped |
| Use SAL instead of LSTM in SCAR (ablation) | SCAR · O3 | test accuracy (%): 52.4 -> 86.7 | Helped |
| Pre-train on RPM datasets, then fine-tune | SCAR · G-set (MTL) | test accuracy (%): 82.1 -> 91.0 (+8.9) | Helped |
| Pre-train on RPM datasets, then fine-tune | SCAR · PGM-S (MTL) | test accuracy (%): 46.1 -> 53.7 (+7.6) | Helped |
Surprising findings
- Replacing SAL with a standard relation network or LSTM in SCAR drops accuracy substantially (e.g. G-set from 82.1% to 36.5% with RN), showing SAL is critical.
- MTL can cause negative transfer for WReN and RNN, e.g. RNN on I-RAVEN drops 4.9 p.p. after pre-training on RPMs.
Limitations
- Task-specific RPM models (SRAN, CoPINet, RelBase, SCL) cannot be applied to VAP or O3, so cross-task comparison is limited to universal models.
- SCAR is slightly below SCL on PGM-S and I-RAVEN in STL and does not surpass the best task-specific RPM models.
- Requires fixed maximum grid dimensions (R=6, C=60) computed as LCM over the task structures, so only supports the structures considered.
Experimental setup
Datasets: G-set PGM-S I-RAVEN VAP-S O3
Benchmarks: G-set PGM-S I-RAVEN VAP-S O3
Metrics: test accuracy
Baselines: WReN RNN SCL SCL+SAL SRAN CoPINet RelBase STSN RN LSTM
Builds on: SCL WReN Relation Network
Closest papers by meaning
- Beyond Task-Specific Reasoning: A Unified Conditional Generative Framework for Abstract Visual Reasoning (ICML 2025 · similarity 0.91)
- VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 2026 · similarity 0.87)
- SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · similarity 0.87)
- MARVEL: Multidimensional Abstraction and Reasoning through Visual Evaluation and Learning (NeurIPS 2024 · similarity 0.87)
- SoTA with Less: MCTS-Guided Sample Selection for Data-Efficient Visual Reasoning Self-Improvement (NeurIPS 2025 · similarity 0.87)
- VGR: Visual Grounded Reasoning (ICLR 2026 · 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.