Research map / Large language models / Instruction and fine-tuning
CAR-LoRA: Training Compression-Aware and Robust LoRA Adapters for Evolving LLMs
Rana Shahroz, Zhen Tan, Ruichen Zhang, Hua Wei, Tianlong Chen, Charles Fleming
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 |
|---|---|---|---|
| apply BF16-trained LoRA naively to INT4-quantized model | LoRA adapter · MATH | accuracy: 16.5 -> 6.4 | Hurt |
| train one CAR-LoRA adapter across compressions instead of per-scheme QLoRA | CAR-LoRA [FP4] · MATH | accuracy: 16.1 -> 16.7 | No clear effect |
| train with layer skipping simulation and evaluate under layer skipping | CAR-LoRA [LS] · GSM8K | accuracy: 38.9 -> 31.1 | Hurt |
| train on all compression operators instead of quantization only, evaluated under layer skipping (ablation) | CAR-LoRA · SQA (LS), Mistral-7B | accuracy: 53.5 -> 64.4 | Improved |
| train on all compression operators instead of pruning+LS only, evaluated at INT4 (ablation) | CAR-LoRA · SQA (INT4), Mistral-7B | accuracy: 58.1 -> 68.4 | Improved |
| add quantization-aware training to BF16 LoRA, evaluated at INT4 (ablation) | CAR-LoRA · SQA (INT4), Mistral-7B | accuracy: 54.3 -> 67.9 | Improved |
| use CAR-LoRA instead of standard LoRA across 4 model updates | CAR-LoRA · SQA, Llama-3.1-8B | accuracy: 54.9 -> 68.2 | Improved |
| evaluate on quantization unseen in training (FP4 withheld) | CAR-LoRA · GSM8K | accuracy: 38.49 -> 37.42 | Hurt |
| evaluate on unseen INT2 instead of INT3 | CAR-LoRA · SQA | accuracy: 61.7 -> 43.5 | Hurt |
| use CAR-LoRA instead of specialized QLoRA on Mistral-7B/Gemma-2-9B | CAR-LoRA · SQA, Gemma-2-9B INT8 | accuracy: 74.5 -> 74.0 | Mixed |
Closest papers by meaning
- NOLA: Compressing LoRA using Linear Combination of Random Basis (ICLR 2024 · similarity 0.89)
- Compress then Serve: Serving Thousands of LoRA Adapters with Little Overhead (ICML 2025 · similarity 0.89)
- $\textit{Trans-LoRA}$: towards data-free Transferable Parameter Efficient Finetuning (NeurIPS 2024 · similarity 0.88)
- Fix the Structural Bottleneck: Context Compression via Explicit Information Transmission (NeurIPS 2026 · similarity 0.87)
- Train Small, Infer Large: Memory-Efficient LoRA Training for Large Language Models (ICLR 2025 · similarity 0.87)
- Sparse High Rank Adapters (NeurIPS 2024 · similarity 0.87)
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.
