AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Towards Quantization-Aware Training for Ultra-Low-Bit Reasoning LLMs

Yasuyuki Okoshi, Hikari Otsuka, Daichi Fujiki, Masato Motomura

ICLR 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of LLMs

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 changedWhereMeasuredEffect
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 averageaccuracy: 10.79 -> 21.56Improved
calibrate with mixed data instead of reasoning-only data (ablation)Qwen3-8B 3-bit · 6-task averageaccuracy: 70.47 -> 71.13Improved
calibrate with mixed data instead of reasoning-only data (ablation)Qwen3-1.7B 2-bit · MATH-500accuracy: 20.60 -> 18.68Hurt
add block-wise calibration before cross-entropy fine-tuning (ablation)2-bit Qwen3-1.7B · MATH-500accuracy: 1.4 -> 22.70Improved
replace cross-entropy with teacher-guided reward rectification loss after calibration (ablation)2-bit Qwen3-1.7B · MATH-500accuracy: 22.70 -> 38.13Improved
fine-tune with cross-entropy after calibration (ablation)2-bit Qwen3-1.7B · MATH-500accuracy: 28.57 -> 22.70Hurt
add KL loss to reward rectification loss (alpha=0.2, beta=1) vs rectification only (ablation)3-bit Qwen3-1.7B · MATH-500accuracy: 78.2 -> 82.7Improved
add reward rectification loss to KL loss (alpha=0.2, beta=1) vs KL only (ablation)2-bit Qwen3-1.7B · MATH-500accuracy: 2.2 -> 48.6Improved
add reward rectification loss to KL loss at 3-bit (ablation)3-bit Qwen3-1.7B · MATH-500accuracy: 82.8 -> 82.7No clear effect
use proposed pipeline instead of GPTQ/AWQQwen3-1.7B 2-bit · 5-benchmark averageaccuracy: 4.2 -> 28.4Improved
use proposed pipeline instead of EfficientQATQwen3-1.7B 2-bit · 5-benchmark averageaccuracy: 18.3 -> 28.4Improved
use proposed pipeline instead of BitDistillerQwen3-1.7B 3-bit · 5-benchmark averageaccuracy: 39.2 -> 55.2Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy