Research map / Vision-language and multimodal / Multimodal LLMs
Take A Step Back: Rethinking the Two Stages in Visual Reasoning
Mingyu Zhang, Jiting Cai, Mingyu Liu, Yue Xu, Cewu Lu, Yong-Lu Li
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 |
|---|---|---|---|
| share the reasoner across tasks while keeping separate encoders (vs sharing both or only the encoder) (ablation) | two-stage visual reasoning framework · RAVEN | accuracy: 35.86 -> 55.72 | Improved |
| share the reasoner across tasks instead of separate ad-hoc models (ablation) | ResNet-18 + MLP · RAVEN, CVR, SVRT, Bongard-HOI, Bongard-LOGO | accuracy: RAVEN 53.40 -> 55.72; Bongard-HOI 61.64 -> 57.65 | Mixed |
| deepen the symbol encoder (ResBlocks) beyond a task-specific inflection point (ablation) | symbol encoder · RAVEN | accuracy: 56.49 -> 39.12 | Hurt |
| deepen the symbol encoder from ResBlock 0 to ResBlock 8 (ablation) | symbol encoder · SVRT | accuracy: 51.96 -> 87.33 | Improved |
| use Transformer reasoner instead of MLP (ablation) | shared reasoner · VQAv2 | accuracy: 50.06 -> 32.64 | Hurt |
| use GCN reasoner instead of MLP (ablation) | shared reasoner · CVR, SVRT, Bongard-LOGO | accuracy: CVR 72.40 -> 78.75 | Improved |
| use Neuro-Symbolic reasoner instead of MLP (ablation) | shared reasoner · RAVEN | accuracy: 55.72 -> 27.89 | Hurt |
| train the shared reasoner on more datasets from different domains (ablation) | shared reasoner · Filtered-CoPhy Collision (3,000 samples) | keypoint error: 13.09 -> 4.88 | Improved |
| use ImageNet-pretrained symbol encoder (ablation) | symbol encoder · RAVEN, CVR, SVRT | accuracy: RAVEN 53.56 -> 52.21; CVR 72.40 -> 74.82 | No clear effect |
| replace separated encoders with CLIP as a general shared encoder (ablation) | One-for-All · RAVEN | accuracy: 55.72 -> 23.59 | Hurt |
| use MiniGPT-4 as shared reasoner instead of lite MLP (ablation) | One-for-All · RAVEN | accuracy: 55.7 -> 21.6 | Hurt |
| train One-for-All jointly rather than separately per task (ablation) | One-for-All · RAVEN, CVR, SVRT, Bongard-HOI | reasoning consistency (F1): RAVEN 76.94 -> 77.89 | Improved |
Closest papers by meaning
- CodeDance: A Dynamic Tool-integrated MLLM for Executable Visual Reasoning (CVPR 2026 · similarity 0.89)
- CogFlow: Bridging Perception and Reasoning through Knowledge Internalization for Visual Mathematical Problem Solving (ICLR 2026 · similarity 0.89)
- MARVEL: Multidimensional Abstraction and Reasoning through Visual Evaluation and Learning (NeurIPS 2024 · similarity 0.89)
- Revisiting the Necessity of Lengthy Chain-of-Thought in Vision-centric Reasoning Generalization (CVPR 2026 · similarity 0.88)
- No Labels, No Problem: Training Visual Reasoners with Multimodal Verifiers (ICLR 2026 · similarity 0.88)
- Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 2026 · 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.
