Research map / Vision-language and multimodal / CLIP and zero-shot
SyncVL: Synchronizing Vision ⟷ Language Using Unsupervised Adaptation
Maria Marrium, Muhammad Haris Khan, Sajid Javed, Arif Mahmood
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 V2T synchronizer without OF (ablation) | CLIP · CIFAR-10 | accuracy (%): 91.30 -> 86.37 | Hurt |
| add V2T synchronizer with OF (ablation) | CLIP · CIFAR-10 | accuracy (%): 91.30 -> 93.06 | Improved |
| add T2V synchronizer to V2T without OF (ablation) | SyncVL · CIFAR-10 | accuracy (%): 86.37 -> 95.86 | Improved |
| add proposed objective function (OF) to V2T+T2V (ablation) | SyncVL · CIFAR-10 | accuracy (%): 95.86 -> 98.65 | Improved |
| add alignment reward to consistency reward (ablation) | SyncVL · CIFAR-10 | accuracy (%): 94.18 -> 96.32 | Improved |
| add KD (latent) reward on top of consistency and alignment (ablation) | SyncVL · CIFAR-10 | accuracy (%): 96.32 -> 98.65 | Improved |
| replace MLP autoencoder with two-block transformer synchronizer (ablation) | SyncVL · CIFAR-10 | accuracy (%): 84.72 -> 98.65 | Improved |
| increase synchronizer depth from two to four transformer blocks (ablation) | SyncVL · CIFAR-100 | accuracy (%): 81.14 -> 79.91 | Hurt |
| add pseudo-label re-grouping (ablation) | SyncVL · DTD | accuracy (%): 58.86 -> 59.92 | Improved |
| add iterative synchronizer training (ablation) | SyncVL · DTD | accuracy (%): 59.92 -> 63.28 | Improved |
| use GRPO-based SyncVL instead of baseline formulation for reward gradients (ablation) | SyncVL · CIFAR-10 | gradient cosine similarity: -0.17 -> 0.44 | Improved |
| apply SyncVL to CLIP RN50 under poor initialization (ablation) | CLIP RN50 · DTD | accuracy (%): 41.70 -> 61.94 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- POUF — POUF: Prompt-Oriented Unsupervised Fine-tuning for Large Pre-trained Models (ICML 2023) · better on 1 of 1 shared result
- LafTer — LaFTer: Label-Free Tuning of Zero-shot Classifier using Language and Unlabeled Image Collections (NeurIPS 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- From Pixels to Words -- Towards Native Vision-Language Primitives at Scale (ICLR 2026 · similarity 0.89)
- VladVA: Discriminative Fine-tuning of LVLMs (CVPR 2025 · similarity 0.88)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- TGV-KV: Text-Grounded KV Eviction for Vision-Language Models (ICML 2026 · similarity 0.88)
- Unveiling Encoder-Free Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- SynerGen-VL: Towards Synergistic Image Understanding and Generation with Vision Experts and Token Folding (CVPR 2025 · similarity 0.88)
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.
