Research map / Large language models / Instruction and fine-tuning
OTARo: Once Tuning for All Precisions Toward Robust On-Device LLMs
Shaoyuan Chen, Zhixuan Chen, Dawei Yang, Zhihang Yuan, Qiang Wu
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 |
|---|---|---|---|
| fine-tune with OTARo (BPS+LAA) instead of fixed-precision fine-tuning | SEFP LLaMA3-8B · zero-shot avg, E5M4 | accuracy: 62.30% -> 63.34% | Improved |
| fine-tune with OTARo instead of FP16 fine-tuning | SEFP Qwen3-8B · zero-shot avg, E5M3 | accuracy: 58.60% -> 61.51% | Improved |
| fine-tune with OTARo instead of fixed-precision fine-tuning | SEFP LLaMA3.2-3B · zero-shot avg, E5M3 | accuracy: 54.78% -> 56.03% | Improved |
| fine-tune with OTARo instead of fixed-precision fine-tuning | SEFP LLaMA3-8B · zero-shot avg, E5M3 | accuracy: 59.84% -> 60.67% | Improved |
| fine-tune with OTARo instead of fixed-precision fine-tuning | SEFP LLaMA3.2-1B · WikiText2 | PPL: average reduction of 0.16 PPL | Improved |
| sample bit-widths uniformly instead of fixed precision (ablation) | SEFP fine-tuning · WikiText2 LLaMA3.2-1B | PPL | Hurt |
| use BPS sampling instead of uniform sampling (ablation) | SEFP fine-tuning · WikiText2 LLaMA3.2-1B | PPL | Improved |
| add LAA to BPS (ablation) | OTARo · zero-shot avg | accuracy | Improved |
| set exploration coefficient lambda=5 vs 3,4,6,7 (ablation) | BPS · zero-shot E5M8 tasks | accuracy | Improved |
| set delay step N=10 vs 5 or 20 (ablation) | LAA · zero-shot E5M8 tasks | accuracy | Improved |
| convert FP16 to SEFP-E5M4 | LLaMA3-8B · memory and decoding | memory and throughput: 15.20 GB -> 4.77 GB; 39.00 -> 95.65 token/s | Improved |
| FP16 fine-tuning then SEFP quantization vs fine-tuning for SEFP | LLaMA3-8B · zero-shot avg, E5M4 | accuracy: 63.29% (E5M5) -> 61.23% (E5M4) | Hurt |
Closest papers by meaning
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.88)
- DP-LLM: Runtime Model Adaptation with Dynamic Layer-wise Precision Assignment (NeurIPS 2025 · similarity 0.87)
- Any-Precision LLM: Low-Cost Deployment of Multiple, Different-Sized LLMs (ICML 2024 · similarity 0.87)
- HALO: Hadamard-Assisted Lower-Precision Optimization for LLMs (NeurIPS 2025 · similarity 0.87)
- ShiftAddLLM: Accelerating Pretrained LLMs via Post-Training Multiplication-Less Reparameterization (NeurIPS 2024 · similarity 0.86)
- LowRA: Accurate and Efficient LoRA Fine-Tuning of LLMs under 2 Bits (ICML 2025 · 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.
