Research map / Vision-language and multimodal / Visual question answering
Meta-Adaptive Prompt Distillation for Few-Shot Visual Question Answering
Akash Gupta, Amos Storkey, Mirella Lapata
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 test-time finetuning (K<=30 steps) of distilled prompts instead of ICL | MAPD · VL-ICL (4 tasks) | accuracy: average increase of 21.2% over all datasets | Improved |
| add meta-task training (first-order MAML) for prompt distillation (ablation) | attention-mapper prompt distillation · VL-ICL (4 tasks, FT) | accuracy: Open-MI 68.0 -> 77.9 (NoMeta-task PD -> MAPD) | Improved |
| use MAML bi-level optimization instead of multi-task loss (ablation) | MAPD · VL-ICL (4 tasks, FT) | accuracy: Open-MI 74.6 -> 77.9; TextOCR 22.9 -> 26.4 | Improved |
| apply LoRA to all LLM layers instead of first 16 layers (ablation) | LoRA baseline · VL-ICL Open-MI | accuracy: 67.3 -> 55.1 | Hurt |
| add LoRA on attention-mapper layers to LoRA on first 16 LLM layers (ablation) | LoRA baseline · VL-ICL (4 tasks) | accuracy: Open-MI 67.3 -> 69.1; CLEVR 30.0 -> 28.7 | Mixed |
| use MAPD instead of best LoRA configuration | few-shot adaptation · VL-ICL (4 tasks) | accuracy: average 7.7% gain | Improved |
| add soft prompts to CLIP projection (ablation) | MAPD projection layer · VL-ICL | accuracy: average gain of 16.3% across architectures | Improved |
| replace 2-layer MLP projector with attention mapper (ablation) | MAPD projection layer · VL-ICL | accuracy: additional average gain of 13.1% | Improved |
| increase number of soft prompts (ablation) | MAPD · VL-ICL | mean accuracy | Improved |
| increase number of soft prompts (ablation) | In-Context PD · VL-ICL | mean accuracy | Hurt |
| use MAPD instead of Multi-Task PD for task induction subtask | Operator Induction · Operator Induction task induction | exact match: +11.7% | Improved |
| swap LLM and vision encoder (3B, Vicuna, SigLIP, Qwen3-8B) with MAPD FT (ablation) | MAPD · Open-MI | accuracy: FT over ICL: +24.6, +37.06, +12.02, +18.04 | Improved |
Closest papers by meaning
- VL-ICL Bench: The Devil in the Details of Multimodal In-Context Learning (ICLR 2025 · similarity 0.90)
- FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering (NeurIPS 2025 · similarity 0.89)
- LIVE: Learnable In-Context Vector for Visual Question Answering (NeurIPS 2024 · similarity 0.89)
- Rethinking Token Reduction for Large Vision-Language Models (CVPR 2026 · similarity 0.88)
- Learning to Learn Better Visual Prompts (AAAI 2024 · similarity 0.88)
- AdaptVision: Efficient Vision-Language Models via Adaptive Visual Acquisition (CVPR 2026 · 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.
