Research map / Architectures and efficiency / Pruning and efficient tuning
ABBA-Adapters: Efficient and Expressive Fine-Tuning of Foundation Models
Raghav Singhal, Kaustubh Ponkshe, Rohit Vartak, Praneeth Vepakomma
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 |
|---|---|---|---|
| use ABBA r=32 instead of LoRA | Llama-3.2 1B · CommonSense170K average | accuracy: 71.49 -> 75.03 | Improved |
| use ABBA r=32 instead of HiRA | Llama-3.2 3B · CommonSense170K average | accuracy: 82.56 -> 84.08 | Improved |
| use ABBA r=32 instead of LoRA | Mistral-7B · GSM8K | accuracy: 61.94 -> 66.26 | Improved |
| use ABBA r=32 instead of full FT | Mistral-7B · GSM8K (20K) | accuracy: 63.87 -> 66.26 | Improved |
| use ABBA instead of full FT with larger 40K data | Mistral-7B · GSM8K (40K) | accuracy: 66.28 -> 67.04 | No clear effect |
| use ABBA instead of LoRA | Llama-3.2 1B · HumanEval | Pass@1: 20.73 -> 25.61 | Improved |
| use ABBA instead of HiRA | Llama-3.1 8B · HumanEval | Pass@1: 48.17 -> 49.39 | Improved |
| initialize with SVD of W0 on one pair and LoRA init on the other instead of both-pairs sqrt(W0) SVD (ablation) | ABBA · GSM8K (Mistral-7B) | accuracy: 64.53 -> 66.26 | Improved |
| use SVD-plus-LoRA init instead of SVD-plus-ones init (ablation) | ABBA · GSM8K (Mistral-7B) | accuracy: 57.39 -> 66.26 | Improved |
| use four chained adapter pairs instead of two (ablation) | ABBA · GSM8K (Mistral-7B) | accuracy: 66.26 -> 64.84 | Hurt |
| raise alpha from 32 to 64 (ablation) | ABBA · CommonSense170K average (Llama-3.2 3B) | accuracy: 84.08 -> 76.36 | Hurt |
| raise alpha from 4 to 16 (ablation) | ABBA · CommonSense170K average (Llama-3.2 3B) | accuracy: 82.43 -> 84.04 | 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.
- DoRA — DoRA: Weight-Decomposed Low-Rank Adaptation (ICML 2024) · better on 2 of 2 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
- RefLoRA: Refactored Low-Rank Adaptation for Efficient Fine-Tuning of Large Models (NeurIPS 2025 · similarity 0.88)
- Linearization Explains Fine-Tuning in Large Language Models (NeurIPS 2025 · similarity 0.87)
- VB-LoRA: Extreme Parameter Efficient Fine-Tuning with Vector Banks (NeurIPS 2024 · similarity 0.87)
- HydraLoRA: An Asymmetric LoRA Architecture for Efficient Fine-Tuning (NeurIPS 2024 · similarity 0.87)
- LoRA-One: One-Step Full Gradient Could Suffice for Fine-Tuning Large Language Models, Provably and Efficiently (ICML 2025 · similarity 0.87)
- PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models (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.
