Research map / Architectures and efficiency / Pruning and efficient tuning
An Optimal Control View of LoRA and Binary Controller Design for Vision Transformers
Chi Zhang, Jingpu Cheng, Qianxiao Li
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 |
|---|---|---|---|
| replace full-precision adaptation matrices with binary controls at rank 64 | AdaptFormer-64 · CIFAR-100 | accuracy: 85.90 -> 85.74 | No clear effect |
| replace full-precision adaptation matrices with binary controls at rank 64 | AdaptFormer-64 · SVHN | accuracy: 96.89 -> 97.01 | Improved |
| binary controls at rank 768 instead of full precision | AdaptFormer-768 · Food-101 | accuracy: 88.39 -> 88.31 | No clear effect |
| use high-rank binary control instead of rank-1 full-precision under same compute | AdaptFormer-1 · SVHN | accuracy: 93.04 -> 97.03 | Improved |
| use high-rank binary control instead of rank-4 full-precision under same compute | AdaptFormer-4 · Food-101 | accuracy: 85.42 -> 87.87 | Improved |
| reduce AdaptFormer rank to 1 (ablation) | AdaptFormer · Food-101 | accuracy: drop of 6.45% | Hurt |
| use high-rank binary control (BiC-3072) in place of full-tuning MLP layers | ViT MLP layers · CIFAR-100 | accuracy: marginally lower, difference of 0.11% | No clear effect |
| use binary controllers in distributed training | AdaptFormer-768 · SVHN (5 nodes) | communication cost: 3.5% of full-precision cost | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AdaptFormer-64 — AdaptFormer: Adapting Vision Transformers for Scalable Visual Recognition (NeurIPS 2022) · better on 8 of 10 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
- Parameter-Efficient Fine-Tuning with Controls (ICML 2024 · similarity 0.87)
- RandLoRA: Full rank parameter-efficient fine-tuning of large models (ICLR 2025 · similarity 0.86)
- Efficient Adaptation of Pre-trained Vision Transformer via Householder Transformation (NeurIPS 2024 · similarity 0.85)
- Computational Limits of Low-Rank Adaptation (LoRA) Fine-Tuning for Transformer Models (ICLR 2025 · similarity 0.85)
- Low-Rank Rescaled Vision Transformer Fine-Tuning: A Residual Design Approach (CVPR 2024 · similarity 0.85)
- PhaseLoRA: Control-Regime-Conditioned Low-Rank Adaptation for Continuous-Action Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.84)
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.
