Research map / Vision-language and multimodal / Multimodal LLMs
Learning from Primitive: Probing Visual Reasoning of LVLMs via Counting
Liwei Che, Zhiyu Xue, Yihao Quan, Benlin Liu, Zeru Shi, Ruixiang Tang, Ranjay Krishna, Vladimir Pavlovic
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 |
|---|---|---|---|
| fine-tune on synthetic counting with focus regularizer and head temperature | Qwen2.5-VL-7B · SynReal (OOD counting) | accuracy: 73.48 -> 84.83 | Improved |
| fine-tune on synthetic counting with focus regularizer and head temperature | Qwen2.5-VL-7B · PixMo-Count | accuracy: 58.79 -> 64.15 | Improved |
| fine-tune on synthetic counting with focus regularizer and head temperature | Qwen3-VL-8B · PixMo-Count | accuracy: 58.75 -> 66.98 | Improved |
| fine-tune on synthetic counting with focus regularizer and head temperature | LLaVA-1.5-7B · PixMo-Count | accuracy: 31.88 -> 32.64 | No clear effect |
| fine-tune on synthetic counting with focus regularizer and head temperature | LLaVA-1.5-7B · SynReal (OOD counting) | accuracy: 54.27 -> 60.11 | Improved |
| counting-only fine-tuning transfers to general visual reasoning | Qwen2.5-VL-7B · MMMU, RealWorldQA, MathVista | average accuracy: +1.54 | Improved |
| counting-only fine-tuning transfers to general visual reasoning | Qwen3-VL-8B · MMMU, RealWorldQA, MathVista | average accuracy: +1.93 | Improved |
| counting-only fine-tuning transfers to general visual reasoning | LLaVA-1.5-7B · MMMU, RealWorldQA, MathVista | average accuracy: +0.93 | Improved |
| add LoRA SFT on synthetic counting (ablation) | Qwen2.5-VL-7B · SynReal (OOD counting) | accuracy: 73.48 -> 79.78 | Improved |
| add object-focused attention regularizer to SFT (ablation) | Qwen2.5-VL-7B · SynReal (OOD counting) | accuracy: 79.78 -> 82.34 | Improved |
| add adaptive head temperature tuning (ablation) | Qwen2.5-VL-7B · SynReal (OOD counting) | accuracy: 82.34 -> 84.83 | Improved |
| add adaptive head temperature tuning (ablation) | Qwen2.5-VL-7B · PixMo-Count | accuracy: 63.32 -> 64.15 | Improved |
Closest papers by meaning
- Understanding Counting Mechanisms in Large Language and Vision-Language Models (CVPR 2026 · similarity 0.92)
- Unveiling the Visual Counting Bottleneck in Vision-Language Models (ICML 2026 · similarity 0.91)
- Vision Language Models are Biased (ICLR 2026 · similarity 0.90)
- Benchmarking Multimodal Mathematical Reasoning with Explicit Visual Dependency (NeurIPS 2026 · similarity 0.89)
- VLMs have Tunnel Vision: Evaluating Nonlocal Visual Reasoning in Leading VLMs (NeurIPS 2025 · similarity 0.89)
- Caption This, Reason That: VLMs Caught in the Middle (NeurIPS 2025 · similarity 0.89)
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.
