Research map / Architectures and efficiency / Quantization
Towards Quantization-Aware Training for Ultra-Low-Bit Reasoning LLMs
Yasuyuki Okoshi, Hikari Otsuka, Daichi Fujiki, Masato Motomura
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 |
|---|---|---|---|
| calibrate with mixed 80/20 reasoning/pre-training data instead of pre-training data only (ablation) | EfficientQAT-style QAT, Qwen3-1.7B 2-bit · 6-task average | accuracy: 10.79 -> 21.56 | Improved |
| calibrate with mixed data instead of reasoning-only data (ablation) | Qwen3-8B 3-bit · 6-task average | accuracy: 70.47 -> 71.13 | Improved |
| calibrate with mixed data instead of reasoning-only data (ablation) | Qwen3-1.7B 2-bit · MATH-500 | accuracy: 20.60 -> 18.68 | Hurt |
| add block-wise calibration before cross-entropy fine-tuning (ablation) | 2-bit Qwen3-1.7B · MATH-500 | accuracy: 1.4 -> 22.70 | Improved |
| replace cross-entropy with teacher-guided reward rectification loss after calibration (ablation) | 2-bit Qwen3-1.7B · MATH-500 | accuracy: 22.70 -> 38.13 | Improved |
| fine-tune with cross-entropy after calibration (ablation) | 2-bit Qwen3-1.7B · MATH-500 | accuracy: 28.57 -> 22.70 | Hurt |
| add KL loss to reward rectification loss (alpha=0.2, beta=1) vs rectification only (ablation) | 3-bit Qwen3-1.7B · MATH-500 | accuracy: 78.2 -> 82.7 | Improved |
| add reward rectification loss to KL loss (alpha=0.2, beta=1) vs KL only (ablation) | 2-bit Qwen3-1.7B · MATH-500 | accuracy: 2.2 -> 48.6 | Improved |
| add reward rectification loss to KL loss at 3-bit (ablation) | 3-bit Qwen3-1.7B · MATH-500 | accuracy: 82.8 -> 82.7 | No clear effect |
| use proposed pipeline instead of GPTQ/AWQ | Qwen3-1.7B 2-bit · 5-benchmark average | accuracy: 4.2 -> 28.4 | Improved |
| use proposed pipeline instead of EfficientQAT | Qwen3-1.7B 2-bit · 5-benchmark average | accuracy: 18.3 -> 28.4 | Improved |
| use proposed pipeline instead of BitDistiller | Qwen3-1.7B 3-bit · 5-benchmark average | accuracy: 39.2 -> 55.2 | Improved |
Closest papers by meaning
- LAQuant: A Simple Overhead-free Large Reasoning Model Quantization by Layer-wise Lookahead Loss (NeurIPS 2026 · similarity 0.90)
- When Reasoning Meets Compression: Understanding the Effects of LLMs Compression on Large Reasoning Models (ICLR 2026 · similarity 0.90)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.90)
- Preserving LLM Capabilities through Calibration Data Curation: From Analysis to Optimization (NeurIPS 2025 · similarity 0.89)
- QuEST: Stable Training of LLMs with 1-Bit Weights and Activations (ICML 2025 · similarity 0.89)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.89)
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.
