Share

Accepted to ECCV 2026

Shared LoRA Subspaces for almost Strict Continual Learning

Share learns a single shared low-rank subspace that grows with new tasks while keeping continual adaptation compact and stable.

Prakhar Kaushik* Ankit Vaidya* Shravan Chaudhari Rama Chellappa Alan Yuille

* Equal contribution.

Based on the Universal Weight Subspace Hypothesis.

Paper PDF arXiv Code ★

Overview

Share focuses on parameter-efficient continual finetuning without data replay or a growing library of task-specific adapters.

Single Shared Subspace

A shared low-rank basis is updated as new tasks arrive, keeping a compact representation of useful adaptation directions.

Almost Strict Continual Learning

New task directions are integrated while minimizing interference with prior tasks, reducing catastrophic forgetting.

Scalable Efficiency

Many task-specific adapters are replaced by one evolving subspace, reducing parameter and memory overhead across modalities.

Method at a Glance

A compact flow for updating knowledge over time.

01

Initialize Subspace

Start with a low-rank basis that captures core task knowledge.

02

Detect New Directions

Find essential directions needed for each incoming task.

03

Update and Share

Merge the new directions into the shared subspace for reuse.

Task t New directions Shared subspace Task t+1

Figures

Key visuals from the paper.

Method Overview

Diagram of initialization, continual adaptation, merging, and finetuning in Share.

Open method figure

Text-to-Image Results

Text-to-image examples comparing LoRA non-continual and Share continual results.

Open results figure

Results Snapshot

Reported gains span vision, language, 3D pose, and text-to-image tasks.

Efficiency

Up to 100x parameter reduction and 281x memory savings compared to traditional LoRA approaches, while maintaining performance close to jointly trained models.

Scalability

A single Share model can replace many task-specific adapters, supporting continual learning without maintaining separate LoRA weights per task.

BibTeX

Use this entry to cite the paper.

@misc{kaushik2026sharedlorasubspacesstrict,
      title={Shared LoRA Subspaces for almost Strict Continual Learning}, 
      author={Prakhar Kaushik and Ankit Vaidya and Shravan Chaudhari and Rama Chellappa and Alan Yuille},
      year={2026},
      eprint={2602.06043},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.06043}, 
}

Paper

Preview the PDF or open it in a new tab.

Open PDF