AI Paper Atlas

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

Reversible Primitive–Composition Alignment for Continual Vision–Language Learning

Canran Xiao, Tianxiang Xu, siyuanma, Yiyang Jiang, Haoyu Gao, Yuhan Wu

ICLR 2026 · Vision-language and multimodal › CLIP and zero-shot · approach: Open-vocabulary alignment from VLMs

Read the paper

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
add composed positive to InfoNCE (ablation)COMPO-REALIGN · Retrieval/ITM, compositional DIL + MTILAvg R@1 Image→Text: 56.9 -> 58.8 (−1.9 when removed)Improved
add composed positive to InfoNCE (ablation)COMPO-REALIGN · Retrieval/ITM, compositional DIL + MTILAF: 4.0 -> 3.2 (+0.8 when removed)Improved
add composed positive to InfoNCE (ablation)COMPO-REALIGN · VQACLAvg accuracy: 55.0 -> 56.8 (−1.8 when removed)Improved
add spectral trust region (ablation)COMPO-REALIGN · Retrieval/ITM, compositional DIL + MTILAF: 4.3 -> 3.2 (+1.1 when removed)Improved
add spectral trust region (ablation)COMPO-REALIGN · Retrieval/ITM, compositional DIL + MTILAvg R@1 Image→Text: 57.9 -> 58.8 (−0.9 when removed)Improved
use Cayley orthogonal core instead of linear mix (ablation)COMPO-REALIGN · Retrieval/ITM, compositional DIL + MTILCRR: 0.88 -> 0.91 (−0.03 when removed)Improved
add tiny text buffer (ablation)COMPO-REALIGN · Retrieval/ITM, compositional DIL + MTILAvg R@1 Image→Text: 56.3 -> 58.8 (−2.5 when removed)Improved
add tiny text buffer (ablation)COMPO-REALIGN · CLOVE-sceneAvg accuracy: 62.8 -> 65.1 (−2.3 when removed)Improved
use mean pooling instead of attention pooling (ablation)COMPO-REALIGN · Retrieval/ITM, compositional DIL + MTILAvg R@1 Image→Text: 58.5 vs 58.8 (−0.3)No clear effect
add primitive shaper (phi and A) (ablation)COMPO-REALIGN · Retrieval/ITM, compositional DIL + MTILAvg R@1 Image→Text: 57.6 -> 58.8 (−1.2 when removed)Improved
use temperature 0.07 instead of 0.10 (ablation)COMPO-REALIGN · Retrieval/ITM, compositional DIL + MTILAvg R@1 Image→Text: 57.4 -> 58.8 (−1.4 when drifted)Improved
use Cayley orthogonal core instead of linear mix (ablation)COMPO-REALIGN · VQACLAvg accuracy: 55.6 -> 56.8 (−1.2 when removed)Improved

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