CodeQuant2: Enhanced Outlier Mitigation and Adaptive Clustering for Low-Precision LRMs

Xingyu Liu; Xiangyang Yin; Tianhua Xia; Haiyu Wang; Sai Qian Zhang

cs.CLNov 6, 2026

CodeQuant2: Enhanced Outlier Mitigation and Adaptive Clustering for Low-Precision LRMs

Authors: Xingyu Liu, Xiangyang Yin, Tianhua Xia, Haiyu Wang, Sai Qian Zhang

Organizations: Tandon School of Engineering, New York University · Courant Institute School of Mathematics, Computing, and Data Science, New York University

Abstract

Large reasoning models (LRMs) improve task performance by generating longer intermediate reasoning traces, but this substantially increases inference costs and makes low-precision quantization increasingly important. However, aggressively quantized models often perform poorly on reasoning tasks. We observe larger activation fluctuations during the first decoding steps of reasoning. Calibration that omits these early states may miss their activation fluctuations. To address these reasoning-specific quantization challenges, we propose CodeQuant2, a clustering-based LUT quantization framework for low-precision large reasoning models. CodeQuant2 mitigates activation outliers with rotation and scaling, uses clustering to represent non-uniform weight distributions, and adapts codebook sizes to the quantization difficulty of each layer. Guided by our observation of reasoning-time activation behavior, CodeQuant2 further improves calibration sample selection to better match actual decoding statistics. Combined with our customized LUT CUDA kernel, CodeQuant2 delivers substantial accuracy improvements on reasoning benchmarks while achieving up to $2.85\times$ speedup over BF16.

Figures & tables

Figure 1: Overview of the CodeQuant2 framework. Left: Rotation and scaling are fused into the model for outlier smoothing. Right: CodeQuant2 applies fine-tuned rotation and scaling, allocates clusters adaptively, tunes centroids and assignments, and uses a LUT-based kernel for inference.

Figure 2: Activation behavior during reasoning decoding on GSM8K Cobbe et al. (2021b) , measured on Qwen3-4B-Thinking Team (2025) and averaged over 128 samples. Left: Averaged activation standard deviation across decoding steps with and without thinking. Right: Activation heatmaps at representative decoding steps.

Figure 3: Adaptive cluster allocation and clustering. Left: Importance scores and latency guide cluster allocation. Right: K-means initializes the centroid matrix and per-row binary assignment matrices.

Figure 4: (a) The steps of the LUT kernel, including a one-time LUT precomputation and table lookup. (b) The precomputed LUT is stored in the shared memory in the Streaming Multiprocessors.

Models	Methods	Wiki2 ( $\downarrow$ )	C4 ( $\downarrow$ )	MA500 ( $\uparrow$ )	GSM8K ( $\uparrow$ )	AIME25 ( $\uparrow$ )	AIME26 ( $\uparrow$ )	HE ( $\uparrow$ )	LCB ( $\uparrow$ )	Avg ( $\uparrow$ )
Qwen3- 4B-Thinking	BF16	11.478	18.391	0.958	0.952	0.688	0.708	0.963	0.758	0.838
	QuaRot	17.366	26.879	0.394	0.506	0.004	0.000	0.000	0.002	0.151
	ParoQuant	14.771	23.070	0.720	0.847	0.183	0.142	0.549	0.180	0.437
	SqueezeLLM	14.627	22.642	0.852	0.908	0.192	0.158	0.695	0.158	0.494
	CodeQuant	13.623	21.410	0.930	0.929	0.392	0.367	0.902	0.464	0.664
	Ours	12.890	20.538	0.950	0.938	0.563	0.613	0.945	0.614	0.770

Table 1: Performance under A4W4 setting in terms of perplexity on WikiText2 and C4, and accuracy on Math500 (MA500), GSM8K, AIME25, AIME26, HumanEval (HE), and LiveCodeBench (LCB).

Models	Methods	A4W4				A8W4
		AIME25		AIME26		AIME25		AIME26
		pass@1	pass@8	pass@1	pass@8	pass@1	pass@8	pass@1	pass@8
Qwen3-4B- Thinking	BF16	0.688	0.733	0.708	0.867	0.688	0.733	0.708	0.867
	QuaRot	0.004	0.033	0.000	0.000	0.175	0.233	0.088	0.133
	ParoQuant	0.183	0.267	0.142	0.133	0.283	0.533	0.271	0.433
	SqueezeLLM	0.192	0.333	0.158	0.200	0.438	0.600	0.454	0.600

Table 2: Mathematical reasoning performance on AIME25 and AIME26 under A4W4 and A8W4, evaluated with both pass@1 and pass@8.

Panel of Table 3 — Table 3: Qwen3-4B-Thinking on L40s: decode throughput.

Configuration	MATH-500	GSM8K	AIME25	AIME26	HumanEval	LiveCodeBench	Avg. Acc.	Gain
Base	0.930	0.929	0.392	0.367	0.902	0.464	0.664	–
+ R&S	0.928	0.934	0.450	0.421	0.933	0.557	0.704	+4.00
+ ACA	0.938	0.938	0.542	0.488	0.937	0.599	0.740	+3.64
+ LC	0.950	0.937	0.546	0.567	0.939	0.619	0.760	+1.94
+ RAC	0.950	0.938	0.563	0.613	0.945	0.614	0.770	+1.08

Table 5: Cumulative component ablation on Qwen3-4B-Thinking under A4W4. The base configuration uses fixed random rotation and SmoothQuant scaling. Each subsequent row adds the listed component cumulatively. AIME results are pass@1, and gains are in percentage points.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Models	Methods	ARC ( $\uparrow$ )	PIQA ( $\uparrow$ )	WinoGrande ( $\uparrow$ )	Avg ( $\uparrow$ )
A4W4
Qwen3- 4B-Thinking	BF16	0.937	0.874	0.830	0.881
	QuaRot	0.752	0.562	0.281	0.532
	ParoQuant	0.913	0.789	0.641	0.781
	SqueezeLLM	0.914	0.791	0.681	0.795
	CodeQuant	0.928	0.836	0.772	0.845

Appendix

Table 8: Performance on standard language understanding and commonsense reasoning tasks under A4W4 and A8W4 quantization.

Models	Methods	Wiki2 ( $\downarrow$ )	C4 ( $\downarrow$ )	MA500 ( $\uparrow$ )	GSM8K ( $\uparrow$ )	AIME25 ( $\uparrow$ )	AIME26 ( $\uparrow$ )	HE ( $\uparrow$ )	LCB ( $\uparrow$ )	Avg ( $\uparrow$ )
A8W4
Qwen3- 4B-Thinking	BF16	11.478	18.391	0.958	0.952	0.688	0.708	0.963	0.758	0.838
	QuaRot	14.312	22.492	0.758	0.868	0.175	0.088	0.537	0.137	0.427
	ParoQuant	13.444	21.190	0.900	0.920	0.283	0.271	0.811	0.448	0.606
	SqueezeLLM	12.999	20.260	0.940	0.938	0.438	0.454	0.915	0.552	0.706
	CodeQuant	12.430	19.732	0.946	0.951	0.596	0.638	0.915	0.629	0.779

Appendix

Table 9: Main evaluation results under A8W4 quantization.

LLM Quantization

Large language model quantization reduces numerical precision to lower memory use and accelerate training or inference.

See all 416 papers

Extreme Low-Bit Inference in Reasoning Models: Failure Modes and Targeted Recovery
Jun 1, 2026Ekaterina Alimaskina, Darya Rudas, Denis Shveykin +3LLM Quantization Repetition in Language Generation
BRAIn Lab, Moscow, Russia
Quantization Inflates Reasoning: Token Inflation as a Hidden Cost of Low-Bit Reasoning Models
Jun 24, 2026Xinyu Lian, Walid Krichene, Beichen Huang +4LLM Inference Efficiency LLM Quantization
University of Illinois Urbana-Champaign · Microsoft · Anyscale +1
LAQuant: A Simple Overhead-free Large Reasoning Model Quantization by Layer-wise Lookahead Loss
May 9, 2026Euntae Choi, Sumin Song, Sungjoo YooLLM Quantization Quantization-Aware Training
Seoul National University

Jun 1, 2026J/K move · Enter open · S save

Extreme Low-Bit Inference in Reasoning Models: Failure Modes and Targeted Recovery

Ekaterina Alimaskina, Darya Rudas, Denis Shveykin, Gleb Molodtsov+2

BRAIn Lab, Moscow, Russia

LLM Quantization Repetition in Language Generation Efficient Language Model Reasoning LLM Inference Acceleration Low-Bit Quantization Large Reasoning Models

Open paper AskPDF arXiv

Large Reasoning Models (LRMs) rely on long reasoning traces, making inference expensive. While low-bit quantization reduces per-token decoding cost, we show that aggressive 2-bit inference can fail to deliver end-to-end speedup because instability in the generation process inflates total token count. Instead of merely lowering answer accuracy, 2-bit quantization often produces much longer traces with repetitive loops, budget exhaustion, delayed commitment, and unclosed reasoning segments. We analyze full reasoning traces of Qwen3 reasoning models across mathematical and commonsense benchmarks and show that accuracy degradation is tightly linked to these process-level failures. To address them, we introduce two lightweight controls: FP16 planning, which gives the 2-bit model a short high-precision outline, and loop rescue, which detects repetitive traces and either commits to an earlier answer or falls back to FP16. On MATH-500, loop rescue improves Qwen3-8B accuracy from 17.2% to 74.2%, while planning plus loop rescue improves Qwen3-32B from 65.0% to 87.2%. Overall, our results show that extreme low-bit reasoning becomes practical when its failures are treated as controllable generation pathologies: with lightweight detection and selective FP16 support, 2-bit inference can recover accuracy while preserving real end-to-end speed. Our code is available at: https://github.com/brain-lab-research/quantized-reasoning.

Abstract

LLM Quantization

Explore similar work

Extreme Low-Bit Inference in Reasoning Models: Failure Modes and Targeted Recovery

Quantization Inflates Reasoning: Token Inflation as a Hidden Cost of Low-Bit Reasoning Models

LAQuant: A Simple Overhead-free Large Reasoning Model Quantization by Layer-wise Lookahead Loss