Research map / Vision-language and multimodal / CLIP and zero-shot
Efficient Test-Time Scaling for Small Vision-Language Models
Mehmet Onurcan Kaya, Desmond Elliott, Dim Papadopoulos
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 |
|---|---|---|---|
| use input perturbation with greedy decoding instead of temperature sampling (ablation) | SmolVLM2-2.2B test-time scaling · 9-benchmark mean | accuracy: 40.9 -> 44.6 | Improved |
| use token-level instead of answer-level aggregation | SmolVLM2-2.2B · 9-benchmark mean | accuracy: 46.8 -> 47.9 | Improved |
| apply TTAug (token-level aggregation of augmented inputs) | SmolVLM2-2.2B · 9-benchmark mean | accuracy: 43.8 -> 47.9 | Improved |
| use TTAug instead of Self-Consistency | SmolVLM2-2.2B · runtime | Runtime (s): 3.73 -> 2.99 | Improved |
| use simple average instead of majority vote in token-level aggregation (ablation) | TTAug · 9-benchmark mean | accuracy: 46.6 -> 47.9 | Improved |
| use simple average instead of most-confident token (ablation) | TTAug · 9-benchmark mean | accuracy: 45.6 -> 47.9 | Improved |
| use Self-Paraphrasing text augmentation in addition to image augmentation (ablation) | TTAug · 9-benchmark mean | accuracy: 43.5 -> 48.5 | Improved |
| use AugGPT paraphrasing instead of Self-Paraphrasing (ablation) | TTAug · 9-benchmark mean | accuracy: 48.5 -> 46.2 | Hurt |
| add generative (FLUX) image augmentations instead of classical (ablation) | TTAug · 9-benchmark mean | accuracy: 48.3 -> 46.6 | Hurt |
| add model parameter adaptation (TTAdapt) on top of TTAug | SmolVLM2-2.2B · 9-benchmark mean | accuracy: 48.3 -> 50.3 | Improved |
| optimize aggregation weights via marginal entropy minimization instead of TTAug (ablation) | SmolVLM2-2.2B · 9-benchmark mean | accuracy: 48.3 -> 48.3 | No clear effect |
Closest papers by meaning
- The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · similarity 0.93)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.93)
- Diversity Matters: Revisiting Test-Time Compute in Vision-Language Models (ICML 2026 · similarity 0.92)
- Efficient Test-Time Adaptation of Vision-Language Models (CVPR 2024 · similarity 0.92)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.92)
- DOTA: Distributional Test-time Adaptation of Vision-Language Models (NeurIPS 2025 · similarity 0.91)
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.
