Research map / Vision-language and multimodal / CLIP and zero-shot
Unlocking Few-Shot Capabilities in LVLMs via Prompt Conditioning and Head Selection
Adhémar de Senneville, Xavier Bou, Jérémy Anger, Rafael Grompone von Gioi, Gabriele Facciolo
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 Task prompt conditioning (ablation) | HEC-T · 10-way 4-shot zero-shot | accuracy: 82.34 -> 84.09 | Improved |
| add Domain prompt conditioning (ablation) | HEC-T · 10-way 4-shot zero-shot | accuracy: 82.34 -> 87.08 | Improved |
| add Class prompt conditioning (ablation) | HEC-T · 10-way 4-shot zero-shot | accuracy: 82.34 -> 88.45 | Improved |
| add Domain prompt conditioning (ablation) | HEC-V · 10-way 4-shot few-shot | accuracy: 90.43 -> 92.78 | Improved |
| add Class prompt conditioning (ablation) | HEC-V · 10-way 4-shot few-shot | accuracy: 90.43 -> 94.14 | Improved |
| use Domain prompt instead of Task prompt for probing | Qwen2-VL probing · 12 datasets 4-shot | accuracy: 75.6 -> 79.9 | Improved |
| use text-heads (HEC-T) instead of summary token | Qwen2-VL · 12 datasets 10-way 0-shot | accuracy: 81.1 -> 86.2 | Improved |
| use HEC-T on LLaVA-OV instead of baseline | LLaVA-OV · 12 datasets 10-way 0-shot | accuracy: 79.7 -> 82.1 | Improved |
| add text-heads to vision-heads (HEC-VT over HEC-V) | HEC · 12 datasets 4-shot vision-text | accuracy: 82.4 -> 83.0 | Improved |
| average probabilities instead of majority vote across heads (ablation) | HEC-V · 10-way 4-shot ensemble | accuracy: 93.85 -> 94.00 | No clear effect |
| learn optimal weights instead of mean for head ensemble (ablation) | HEC-V · 10-way 4-shot ensemble | accuracy: 94.00 -> 94.04 | No clear effect |
| select domain heads instead of task heads (ablation) | HEC-V · 10-way 4-shot head selection | accuracy: 93.79 -> 94.18 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ProKeR — ProKeR: A Kernel Perspective on Few-Shot Adaptation of Large Vision-Language Models (CVPR 2025) · 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
- Making Large Vision Language Models to Be Good Few-Shot Learners (AAAI 2025 · similarity 0.91)
- VladVA: Discriminative Fine-tuning of LVLMs (CVPR 2025 · similarity 0.90)
- Large Language Models are Good Prompt Learners for Low-Shot Image Classification (CVPR 2024 · similarity 0.90)
- Bridge the Modality and Capability Gaps in Vision-Language Model Selection (NeurIPS 2024 · similarity 0.89)
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.89)
- The Neglected Tails in Vision-Language Models (CVPR 2024 · 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.
