Research map / Vision-language and multimodal / CLIP and zero-shot
Is Less More? Exploring Token Condensation as Training-free Test-time Adaptation
Zixin Wang, Dong Gong, Sen Wang, Zi Huang, Yadan Luo
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 domain anchor tokens and head-wise sorting score together (ablation) | TCA · EuroSAT | accuracy: 68.14 -> 70.43 | Improved |
| add domain anchor tokens alone (ablation) | TCA · EuroSAT | accuracy: 68.14 -> 67.69 | Hurt |
| add head-wise sorting score alone (ablation) | TCA · Pets | accuracy: 88.77 -> 89.23 | Improved |
| add both components (ablation) | TCA · Food101 | accuracy: 85.25 -> 85.31 | No clear effect |
| use diversity-enforced reservoir instead of FIFO (ablation) | TCA · EuroSAT | accuracy: 50.28 -> 70.43 | Improved |
| use diversity-enforced reservoir instead of similarity-enforced (ablation) | TCA · Pets | accuracy: 88.91 -> 89.53 | Improved |
| increase logits correction temperature beta from 0.01 to 5 (ablation) | TCA · EuroSAT | accuracy: 63.64 -> 70.43 | Improved |
| increase logits correction temperature beta from 0.01 to 5 (ablation) | TCA · Flower102 | accuracy: 73.33 -> 70.44 | Hurt |
| increase correction weight lambda from 2 to 8 (ablation) | TCA · EuroSAT | accuracy: 60.15 -> 70.43 | Improved |
| use merging:pruning ratio 2:1 instead of pruning only 0:1 (ablation) | TCA · EuroSAT | accuracy: 69.63 -> 70.43 | Improved |
| use K=2 merging centers instead of K=1 (ablation) | TCA · EuroSAT | accuracy: 66.25 -> 70.43 | Improved |
| add TCA token condensation to CLIP at severity 5 contrast (ablation) | CLIP · CIFAR-100-C contrast | accuracy gain: +18.59% | Improved |
Closest papers by meaning
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.90)
- One Layer's Trash is Another Layer's Treasure: Adaptive Layer-wise Visual Token Selection in LVLMs (CVPR 2026 · similarity 0.90)
- The Illusion of Progress? A Critical Look at Test-Time Adaptation for Vision-Language Models (NeurIPS 2025 · similarity 0.90)
- Rethinking Visual Token Reduction in LVLMs Under Cross-Modal Misalignment (AAAI 2026 · similarity 0.90)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- AdaptVision: Efficient Vision-Language Models via Adaptive Visual Acquisition (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.
