Research map / Architectures and efficiency / Pruning and efficient tuning
TRAC: Tensor-Train based Across-layer Compression for Parameter-Efficient Fine-Tuning
Bangguo Ye, Yuanwei Zhang, Xiaoqun Zhang
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 |
|---|---|---|---|
| add vector controllers on shared TT cores (ablation) | TRAC · CoLA (RoBERTa-Base) | Matthews corr.: 61.63 -> 65.2 | Improved |
| make tensor cores trainable instead of frozen (ablation) | TRAC · CoLA (RoBERTa-Base) | Matthews corr.: 38.98 -> 65.2 | Improved |
| make tensor cores trainable instead of frozen (ablation) | TRAC · CIFAR-100 (ViT-Base) | accuracy: 67.47 -> 84.21 | Improved |
| make all cores independently trainable instead of freeze+share (ablation) | TRAC · CoLA (RoBERTa-Base) | Matthews corr.: 65.2 -> 66.17 | No clear effect |
| freeze-share strategy vs all trainable at same TT rank (ablation) | TRAC · RoBERTa-Base MRPC | accuracy: 88.92 -> 88.63 | No clear effect |
| use TT normal initialization instead of TT uniform (ablation) | TRAC · CoLA (RoBERTa-Base) | Matthews corr.: 62.4 -> 65.2 | Improved |
| use TT normal instead of Kaiming normal initialization (ablation) | TRAC · MRPC (RoBERTa-Base) | accuracy: 87.9 -> 89.0 | Improved |
| use asymmetric rank allocation (more to B) instead of symmetric (ablation) | TRAC · CoLA (RoBERTa-Base) | Matthews corr.: 62.6 -> 65.2 | Improved |
| use asymmetric rank allocation instead of symmetric (ablation) | TRAC · E2E (GPT-2) | BLEU: 69.4 -> 70.3 | Improved |
| use 3rd-order instead of 4th-order tensor (ablation) | TRAC · RTE (RoBERTa-Base) | accuracy: 75.3 -> 80.1 | Improved |
| increase tensor order from 3rd to 5th (ablation) | TRAC · training time per epoch | seconds: 776.3 -> 873.7 | Hurt |
| replace LoRA with TRAC | LoRA on ViT-Base · image classification average | accuracy: 90.27 -> 90.85 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- VeRA — VeRA: Vector-based Random Matrix Adaptation (ICLR 2024) · better on 8 of 10 shared results
- VB-LoRA — VB-LoRA: Extreme Parameter Efficient Fine-Tuning with Vector Banks (NeurIPS 2024) · better on 1 of 1 shared result
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
- Distribution-Aware Tensor Decomposition for Compression of Convolutional Neural Networks (NeurIPS 2025 · similarity 0.85)
- Geometry-aware training of factorized layers in tensor Tucker format (NeurIPS 2024 · similarity 0.85)
- LISA: Layerwise Importance Sampling for Memory-Efficient Large Language Model Fine-Tuning (NeurIPS 2024 · similarity 0.85)
- Compress to Impress: Efficient LLM Adaptation Using a Single Gradient Step on 100 Samples (NeurIPS 2025 · similarity 0.85)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.85)
- Compressible Dynamics in Deep Overparameterized Low-Rank Learning & Adaptation (ICML 2024 · 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.
