Research map / Vision-language and multimodal / Multimodal LLMs
On Test-Time Scaling for Vision-Language Models
Fawaz Sammani, Tzoulio Chamiti, Nikos Deligiannis
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 self-consistency test-time scaling | Qwen3-VL-2B · WeMath | accuracy: 35.46 -> 64.25 | Improved |
| add CoT prompting | Qwen3-VL-4B · LogicVista | accuracy: 40.85 -> 57.59 | Improved |
| add CoT prompting | Qwen3-VL-4B · RealWorldQA | accuracy: 71.76 -> 69.80 | Hurt |
| add CoT prompting | Qwen3-VL-2B · RealWorldQA | accuracy: 65.10 -> 60.65 | Hurt |
| add self-refinement | Qwen2.5-VL-7B · RealWorldQA | accuracy: 70.20 -> 47.97 | Hurt |
| add CoT prompting to small Qwen2.5-VL-7B | Qwen2.5-VL-7B · LogicVista | accuracy: 39.06 -> 38.17 | Hurt |
| use Qwen3-VL-4B with CoT instead of Qwen3-VL-32B baseline | Qwen3-VL-4B · LogicVista | accuracy: 44.42 -> 57.59 | Improved |
| add prompt repetition | Qwen3-VL-2B · WeMath | accuracy: 35.46 -> 39.20 | Improved |
| add Describe-Answer | Qwen3-VL-2B · LogicVista | accuracy: 32.59 -> 30.13 | Hurt |
| reduce token budget from 1024 to 300 for CoT (ablation) | Qwen3-VL-4B · WeMath | accuracy: -11.32 | Hurt |
| reduce token budget from 1024 to 300 for S-CoT (ablation) | Qwen3-VL-4B · RealWorldQA | accuracy: +3.79 | Improved |
| drop image KV cache late in CoT generation (after ~200 steps) (ablation) | LVLM CoT · WeMath etc. | accuracy | No clear effect |
Closest papers by meaning
- Diversity Matters: Revisiting Test-Time Compute in Vision-Language Models (ICML 2026 · similarity 0.93)
- The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · similarity 0.91)
- Are We on the Right Way for Evaluating Large Vision-Language Models? (NeurIPS 2024 · similarity 0.91)
- Unleashing Perception-Time Scaling to Multimodal Reasoning Models (ICLR 2026 · similarity 0.90)
- SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · similarity 0.90)
- Scaling Test-Time Robustness of Vision-Language Models via Self-Critical Inference Framework (CVPR 2026 · similarity 0.90)
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.
