Research map / Architectures and efficiency / Pruning and efficient tuning
Low-Rank Ternary Adaptation for Fine-Tuning Transformers
Alexandru-Dragos Manolache, Yunqiang Li, Jan van Gemert
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 ternary multiplicative Kronecker adapter to ternary SpinQuant backbone | Llama-3.2-1B · 9-task zero-shot average | average accuracy: 33.2 -> 35.3 | Improved |
| add ternary multiplicative adapter to ternary SpinQuant backbone | Llama-3.2-1B · WikiText-2 | perplexity: 86.5 -> 44.6 | Improved |
| add ternary multiplicative adapter to ternary SpinQuant backbone | Llama-3.2-3B · 9-task zero-shot average | average accuracy: 31.9 -> 38.3 | Improved |
| add ternary multiplicative adapter to ternary SpinQuant backbone | Llama-3.2-3B · WikiText-2 | perplexity: 45.6 -> 22.3 | Improved |
| native ternary adapter versus QLoRA merged and requantized to 1.58-bit | Llama-3.2-3B ternary · 9-task zero-shot average | average accuracy: 37.5 -> 38.3 | Improved |
| native ternary adapter versus QLoRA merged and requantized to 1.58-bit | Llama-3.2-3B ternary · WikiText-2 | perplexity: 22.9 -> 22.3 | Improved |
| native ternary adapter versus QLoRA merged and requantized to ternary | ternary ViT-B/16 · ImageNet-100 | top-1 accuracy: 78.9 -> 83.0 | Improved |
| merge and requantize QLoRA adapters to ternary instead of keeping 16-bit adapters (ablation) | QLoRA on ternary ViT-B/16 · ImageNet-100 | top-1 accuracy: 85.3 -> 78.9 | Hurt |
| add ternary adapter to pre-trained ternary model fine-tuned on GSM8K | BitNet b1.58 2B4T · GSM8K | exact match: 60.1 -> 63.0 | Improved |
| add ternary adapter to pre-trained ternary model fine-tuned on GSM8K | Falcon-E-1B-Instruct · GSM8K | exact match: 52.0 -> 55.1 | Improved |
| add ternary adapter to pre-trained ternary model fine-tuned on GSM8K | Falcon-E-3B-Instruct · GSM8K | exact match: 65.4 -> 66.4 | Improved |
| use balanced initialization instead of normalized initialization (ablation) | ternary adapter · Llama-3.2-1B WikiText-2 | perplexity: 52.0 -> 44.6 | Improved |
Closest papers by meaning
- LoTA-QAF: Lossless Ternary Adaptation for Quantization-Aware Fine-Tuning (NeurIPS 2025 · similarity 0.88)
- WeGeFT: Weight‑Generative Fine‑Tuning for Multi‑Faceted Efficient Adaptation of Large Models (ICML 2025 · similarity 0.88)
- FrameQuant: Flexible Low-Bit Quantization for Transformers (ICML 2024 · similarity 0.88)
- The Quantization Benefits of Residual-Free Transformers (NeurIPS 2026 · similarity 0.87)
- Computational Limits of Low-Rank Adaptation (LoRA) Fine-Tuning for Transformer Models (ICLR 2025 · similarity 0.87)
- AffineQuant: Affine Transformation Quantization for Large Language Models (ICLR 2024 · similarity 0.86)
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.
