AI Paper Atlas

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

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 5 citations (OpenAlex)

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 changedWhereMeasuredEffect
fine-tune on synthetic counting with focus regularizer and head temperatureQwen2.5-VL-7B · SynReal (OOD counting)accuracy: 73.48 -> 84.83Improved
fine-tune on synthetic counting with focus regularizer and head temperatureQwen2.5-VL-7B · PixMo-Countaccuracy: 58.79 -> 64.15Improved
fine-tune on synthetic counting with focus regularizer and head temperatureQwen3-VL-8B · PixMo-Countaccuracy: 58.75 -> 66.98Improved
fine-tune on synthetic counting with focus regularizer and head temperatureLLaVA-1.5-7B · PixMo-Countaccuracy: 31.88 -> 32.64No clear effect
fine-tune on synthetic counting with focus regularizer and head temperatureLLaVA-1.5-7B · SynReal (OOD counting)accuracy: 54.27 -> 60.11Improved
counting-only fine-tuning transfers to general visual reasoningQwen2.5-VL-7B · MMMU, RealWorldQA, MathVistaaverage accuracy: +1.54Improved
counting-only fine-tuning transfers to general visual reasoningQwen3-VL-8B · MMMU, RealWorldQA, MathVistaaverage accuracy: +1.93Improved
counting-only fine-tuning transfers to general visual reasoningLLaVA-1.5-7B · MMMU, RealWorldQA, MathVistaaverage accuracy: +0.93Improved
add LoRA SFT on synthetic counting (ablation)Qwen2.5-VL-7B · SynReal (OOD counting)accuracy: 73.48 -> 79.78Improved
add object-focused attention regularizer to SFT (ablation)Qwen2.5-VL-7B · SynReal (OOD counting)accuracy: 79.78 -> 82.34Improved
add adaptive head temperature tuning (ablation)Qwen2.5-VL-7B · SynReal (OOD counting)accuracy: 82.34 -> 84.83Improved
add adaptive head temperature tuning (ablation)Qwen2.5-VL-7B · PixMo-Countaccuracy: 63.32 -> 64.15Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy