AI Paper Atlas

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

HALoRA: Low-Rank Adaptation with Hierarchical Budget Allocation for Efficient Vision-Language Alignment

Letian Zhang, GuangHao Meng, XuDong Ren, Jinpeng Wang

AAAI 2026 · Vision-language and multimodal › CLIP and zero-shot · approach: Test-time adaptation of VLMs · 1 citations (OpenAlex)

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
use component-level instead of model-level rank allocator (ablation)HALoRA · Flickr30k retrievalTR@1: 55.6 -> 56.8Improved
use component-level instead of modality-level allocator (ablation)HALoRA · Flickr30k retrievalTR@1: 54.9 -> 56.8Improved
use component-level instead of model-level allocator (ablation)HALoRA · UCF101 zero-shotAcc-1: 26.4 -> 27.2Improved
add gradient-approximation initialization (ablation)HALoRA · Flickr30k retrievalTR (mean of R@1,5,10): 74.2 -> 76.4Improved
add gradient-approximation initialization (ablation)HALoRA · ImageNet-A zero-shotAcc-1: 6.69 -> 6.54Hurt
add gradient-approximation initialization (ablation)HALoRA · UCF101 zero-shotAcc-1: 24.8 -> 27.3Improved
use asymmetric instead of symmetric manual ranks (ablation)LoRA-GA · Flickr30k retrievalTR@1: 53.7 -> 56.6Improved
use asymmetric instead of uniform ranks (ablation)LoRA-GA · Flickr30k retrievalIR@1: 42.0 -> 42.4Improved
use asymmetric ranks instead of uniform (ablation)CLIP-LoRA · 11-dataset few-shot (1-shot, avg)accuracy: 72.5 -> 72.7Improved
use HALoRA instead of AdaLoRALoRA-based PEFT · Flickr30k retrievalTR@1: 52.8 -> 56.8Improved
use plain LoRA instead of adapters (LilT LwA)dual-encoder PEFT · Flickr30k retrievalIR@1: 41.7 -> 40.9Hurt

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