AI Paper Atlas

Research map / Vision-language and multimodal / CLIP and zero-shot

Prompt-Robust Vision-Language Models via Meta-Finetuning

Haohui Liang, Runlin Huang, Yingjun Du, Yujia Hu, Weifeng Su, Cees G. M. Snoek

ICLR 2026 · Vision-language and multimodal › CLIP and zero-shot · approach: Prompt learning for CLIP

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 changedWhereMeasuredEffect
apply Promise meta-finetuning on top of existing prompt learnersIVLP, MaPLe, MMRL · Base-to-new (11 datasets avg)Harmonic mean H: MaPLe 78.55 -> 80.29; MMRL 81.20 -> 82.33; IVLP 77.66 -> 79.16Improved
add adaptive prompt weighting (ablation)Promise on MaPLe · Base-to-new (11 datasets avg)Harmonic mean H: 79.24 -> 80.29Improved
add adaptive prompt weighting (ablation)Promise on MMRL · Base-to-new (11 datasets avg)Harmonic mean H: 81.53 -> 82.33Improved
use token-specific learning rate instead of fixed rate (ablation)Promise on MaPLe · Base-to-new (11 datasets avg)Harmonic mean H: 79.08 -> 80.29Improved
use token-specific learning rate instead of task-specific (MetaSGD) rate (ablation)Promise on MMRL · Base-to-new (11 datasets avg)Harmonic mean H: 81.48 -> 82.33Improved
use both inner and outer loops rather than only one (ablation)Promise on MaPLe · Base-to-new (11 datasets avg)Harmonic mean H: 79.08 -> 80.29 (vs inner-only)Improved
use full Promise instead of simple variance regularization (ablation)MMRL · Base-to-new (11 datasets avg)Harmonic mean H: 81.44 -> 82.33Improved
replace hand-crafted CLIP templates with LLM-generated templates for meta-finetuning (ablation)Promise on MaPLe · Base-to-new (11 datasets avg)Harmonic mean H: 80.29 -> 81.74Improved
increase template pool from 60 to 80 or 100 (ablation)Promise on MMRL · Base-to-new (11 datasets avg)Harmonic mean H: 82.33 -> 82.39 (80), 82.29 (100)No clear effect
reduce template pool from 60 to 30 (ablation)Promise on MMRL · Base-to-new (11 datasets avg)Harmonic mean H: 82.33 -> 81.89Hurt
allow 20% overlap between inner and outer prompt sets (ablation)Promise · Base-to-new (11 datasets avg)Harmonic mean H: 80.31 -> 80.07No clear effect
apply Promise to SigLIP and BLIP-2 backbonesMaPLe, MMRL · Base-to-new SigLIP/BLIP-2 (11 datasets avg)Harmonic mean H: SigLIP MaPLe 79.50 -> 80.84; BLIP-2 MMRL 81.07 -> 82.30Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 3 on most shared results.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy