Research map / Vision-language and multimodal / Multimodal LLMs
SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant
Guohao Sun, Can Qin, JIAMINAN WANG, Zeyuan Chen, Ran Xu, Zhiqiang Tao
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 visual self-questioning as extra training task (ablation) | SQ-LLaVA 7B · 5 VQA benchmarks, 558K PT / 665K IT | average accuracy: 66.4 -> 67.5 | Improved |
| add prototype extractor (ablation) | SQ-LLaVA 7B · 5 VQA benchmarks, 558K PT / 665K IT | average accuracy: 66.9 -> 67.5 | Improved |
| add LoRA on the vision encoder (ablation) | SQ-LLaVA 7B · 5 VQA benchmarks, 558K PT / 665K IT | average accuracy: 67.3 -> 67.5 | Improved |
| add prototype extractor on larger data scale (ablation) | SQ-LLaVA 7B · 5 VQA benchmarks, 1200K PT / 700K IT | average accuracy: 69.6 -> 69.6 | No clear effect |
| add all three components versus baseline (ablation) | SQ-LLaVA 7B · 5 VQA benchmarks, 558K PT / 665K IT | average accuracy: 65.9 -> 67.5 | Improved |
| add all three components on larger data (ablation) | SQ-LLaVA 7B · 5 VQA benchmarks, 1200K PT / 700K IT | average accuracy: 67.6 -> 69.6 | Improved |
| SQ-LLaVA versus LLaVA-v1.5 on LLaVA-Bench | SQ-LLaVA 7B · LLaVA-Bench (in the wild) | accuracy: 63.4 -> 66.3 | Improved |
| SQ-LLaVA versus LLaVA-v1.5 on POPE hallucination | SQ-LLaVA 7B · POPE | accuracy: 85.9 -> 87.7 | Improved |
| SQ-LLaVA versus ShareGPT4V on VQAv2, VizWiz, MM-Vet | SQ-LLaVA 7B · VQAv2 / VizWiz / MM-Vet | accuracy: MM-Vet 37.6 vs 32.5; VizWiz 57.2 vs 54.0 | Mixed |
| SQ-LLaVA zero-shot COCO captioning versus LLaVA-v1.5 | SQ-LLaVA · COCO zero-shot captioning | CIDEr: 111.46 -> 110.77 | No clear effect |
Closest papers by meaning
- Enhancing Large Vision Language Models with Self-Training on Image Comprehension (NeurIPS 2024 · similarity 0.92)
- Towards Self-Refinement of Vision-Language Models with Triangular Consistency (NeurIPS 2025 · similarity 0.91)
- Rephrase, Augment, Reason: Visual Grounding of Questions for Vision-Language Models (ICLR 2024 · similarity 0.90)
- Self-Training Large Language Models for Improved Visual Program Synthesis With Visual Reinforcement (CVPR 2024 · similarity 0.90)
- Tackling Vision Language Tasks through Learning Inner Monologues (AAAI 2024 · similarity 0.90)
- Rethinking Token Reduction for Large Vision-Language Models (CVPR 2026 · 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.
