Does the Linguistic Register of Thinking Matter? Wenyan as a Compact Reasoning Register for Chinese LLMs

Minghan Li; Guodong Zhou

cs.CLOct 14, 2026

Does the Linguistic Register of Thinking Matter? Wenyan as a Compact Reasoning Register for Chinese LLMs

Authors: Minghan Li, Guodong Zhou

Organizations: Soochow University, Suzhou, China

Abstract

Visible reasoning traces of Chinese reasoning-oriented LLMs are typically written in modern vernacular Chinese, which may be verbose under limited decoding budgets. We ask whether the linguistic register of reasoning can be used to improve efficiency. We propose Wenyan Thinking, which emits intermediate reasoning in Classical Chinese (wenyanwen) while retaining a short direct-answer output. Starting from modern-Chinese reasoning traces, we rewrite only the reasoning field into wenyan and fine-tune Qwen3-4B-Thinking-2507 on the rewritten supervision. To distinguish register shift from generic compression, we construct a compact modern-Chinese control using the same source data, teacher family, base model, and SFT recipe. Across budget-sensitive evaluations on MMLU-ProX-Lite-ZH, CEVAL, and AGIEval-CN, Wenyan Thinking substantially reduces reasoning length and missing-answer rates under constrained decoding. Compact rewriting explains much of the efficiency gain, while wenyan consistently outperforms the compact modern-Chinese control on the primary benchmark and remains competitive across the additional benchmarks. These results identify linguistic register as a practical design axis for efficient visible reasoning in Chinese LLMs.

Figures & tables

Figure 1: Overview of Wenyan Thinking. We first convert modern Chinese reasoning traces into concise Wenyan-style rewrites through controlled rewriting, then perform supervised fine-tuning on the Wenyan-rewritten reasoning data to obtain a Wenyan-thinking model. Under limited generation budgets, the resulting model emits shorter reasoning traces, leaving more room for final answer completion and improving decoding efficiency.

Field	Mean	Median	Min	Max
Question	69.8	41	3	5811
Answer	905.3	799	4	7469
Modern reasoning	1457.8	926	41	24025
Wenyan reasoning	420.9	382	4	5794

Table 1: Training data statistics and main SFT configuration. The upper panel reports character lengths in WenyanThinking-10k ; the lower panel reports the full-parameter SFT setup used by the final main system.

Model	1024	2048	4096	6144
Answer-only	41.5	41.5	41.0	41.5
Think	11.4	32.3	50.9	58.8
Concise prompt	13.4	32.0	44.9	47.5
Long modern SFT	37.1	47.8	51.2	53.7
Compact modern	52.2	52.9	53.2	53.4
Wenyan	52.9	55.3	55.4	55.6

Table 2: Exact-match (%) on MMLU-ProX-Lite-ZH under different generation budgets. Wenyan denotes our final Qwen3.5-based main system. Concise prompt is a prompt-only baseline without SFT; Answer-only uses no visible thinking trace. Boldface marks the best accuracy, and underline marks the second best for each budget column.

	Budget = 1024				Budget = 6144
Subject	Think	Long	Compact	Wenyan	Think	Long	Compact	Wenyan
Business	2.5	42.5	60.0	65.0	65.0	60.0	60.0	65.0
Chemistry	5.4	30.4	55.4	53.6	64.3	66.1	57.1	62.5
Computer Sci.	5.0	20.0	45.0	55.0	65.0	40.0	45.0	60.0
Economics	23.8	59.5	66.7	73.8	73.8	73.8	66.7	73.8
Law	0.0	2.1	18.8	31.2	27.1	6.2	20.8	31.2

Table 3: Representative subject-wise exact-match (%) on MMLU-ProX-Lite-ZH . Long denotes long modern-Chinese SFT; Compact denotes the compact modern-Chinese control; Wenyan denotes our final Qwen3.5-based main system. Boldface marks the best accuracy within each subject and budget block.

	Accuracy (%)			Avg. output tokens			Missing-answer rate
Budget	Think	Compact	Wenyan	Think	Compact	Wenyan	Think	Compact	Wenyan
1024	11.4	52.2	52.9	1008.1	254.7	394.2	87.5%	3.1%	7.5%
2048	32.3	52.9	55.3	1859.4	267.5	434.3	59.9%	0.7%	1.9%
4096	50.9	53.2	55.4	2901.8	274.0	458.4	27.1%	0.2%	0.7%
6144	58.8	53.4	55.6	3389.7	274.1	467.9	14.3%	0.0%	0.3%

Table 4: Length and answer-completion analysis under different generation budgets. Compact denotes the modern-Chinese compact control, and Wenyan denotes our final Qwen3.5-based main system. Boldface marks the best value within each metric block: highest accuracy and lowest tokens/rates.

$n$	Faith.	Steps	Read.	Support	Leak. $\downarrow$	Loss $\downarrow$	Major err.
40	4.35	4.30	4.33	4.45	1.00	1.53	0.00

Table 5: LLM-as-judge audit of wenyan rewrite quality. Scores are on 1–5 scales; higher is better except leakage risk and over-compression loss.

Model	1024	2048	4096	6144
Think	41.0	63.0	75.6	79.3
Long modern	63.4	69.4	70.6	73.0
Compact modern	72.7	72.8	72.9	73.1
Wenyan	73.5	74.8	74.7	74.7

Table 6: Exact-match (%) on CEVAL-valid-direct under different generation budgets. Compact modern rewrites the reasoning traces into concise modern Chinese rather than wenyan. Boldface marks the best accuracy, and underline marks the second best for each budget column.

	Accuracy (%)				Time (Latency)
Model	1024	2048	4096	6144	1024	2048	4096	6144
Think	14.20	24.37	33.49	35.89	24:20	1:06:19	2:35:33	3:36:37
Long modern	40.89	47.82	52.22	52.97	19:59	40:59	1:39:32	3:02:33
Compact modern	48.94	49.45	49.80	49.83	05:28	06:11	06:48	07:34
Wenyan	49.92	51.55	51.43	51.27	07:54	08:53	10:14	11:20

Table 7: Zero-shot accuracy (%) and total evaluation time on the AGIEval-CN direct subset under different generation budgets. The subset excludes Gaokao MathCloze. Boldface marks the best (highest accuracy, lowest latency), and underline marks the second best for each budget column.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Variant	$n$	Faith.	Steps	Read.	Support	Leak. $\downarrow$	Loss $\downarrow$	Major err.
Wenyan rewrite	40	4.35	4.30	4.33	4.45	1.00	1.53	0.00

Appendix

Table 8: LLM-as-judge rewrite-quality audit. Faith., Steps, Read., and Support denote faithfulness, step preservation, readability/reversibility, and whether the rewrite supports the final answer.

Group	Model	Exact-match accuracy (%)
Group	Model	1024	2048	4096	6144
Baselines / non-Wenyan	Qwen3-4B-Instruct-2507 answer-only	41.5	41.5	41.0	41.5
	Qwen3-4B-Thinking-2507	11.4	32.3	50.9	58.8
	Qwen3-4B-Thinking-2507 concise prompt	13.4	32.0	44.9	47.5
	Baihua-SFT	37.1	47.8	51.2	53.7
	Compact modern Chinese control	52.2	52.9	53.2	53.4

Appendix

Table 9: Complete MMLU-ProX-Lite-ZH leaderboard across the evaluated systems. Values are exact-match accuracy (%). The concise-prompt baseline is prompt-only and uses no additional SFT. The compact modern-Chinese control uses the same source examples, teacher family, base model, and SFT recipe as the Wenyan system, but rewrites the reasoning traces into concise modern Chinese rather than Classical Chinese. Boldface marks the best accuracy.

Group	Model	Exact-match accuracy (%)
Group	Model	1024	2048	4096	6144
Baselines / non-Wenyan	Qwen3-4B-Thinking-2507	41.0	63.0	75.6	79.3
	Baihua-SFT	63.4	69.4	70.6	73.0
	Compact modern Chinese control	72.7	72.8	72.9	73.1
Wenyan variants	DeepSeek-V3.2 normal	73.9	74.3	74.9	73.3
	Ours: Qwen3.5-397B-A17B normal	73.5	74.8	74.7	74.7

Appendix

Table 10: Complete CEVAL-valid-direct leaderboard across the evaluated systems used in the final paper. Values are zero-shot exact-match accuracy (%). The compact modern-Chinese control uses concise modern Chinese reasoning traces rather than Wenyan, and is included to disentangle the effect of compact rewriting from the effect of shifting the reasoning register. Boldface marks the best accuracy.

Category	Non-Wenyan / Controls												Wenyan Variants
	Qwen3-4B-Thinking-2507				Long Vernacular SFT				Compact modern Chinese control				Wenyan (DeepSeek-V4-Flash teacher)				Wenyan Default (Ours)
	1024	2048	4096	6144	1024	2048	4096	6144	1024	2048	4096	6144	1024	2048	4096	6144	1024	2048	4096	6144
Avg.	14.20	24.37	33.49	35.89	40.89	47.82	52.22	52.97	48.94	49.45	49.80	49.83	52.15	52.69	52.08	51.87	49.92	51.55	51.43	51.27
Gaokao Biology	17.14	24.29	30.00	30.00	74.29	77.62	79.52	79.52	73.33	74.29	73.33	73.33	81.43	78.57	77.14	79.05	74.76	78.57	73.81	74.76
Gaokao Chemistry	7.73	15.94	25.12	34.30	45.89	53.14	60.39	59.90	53.14	51.69	52.17	52.17	57.00	57.49	56.52	54.11	54.11	59.42	56.04	57.49
Gaokao Chinese	12.20	22.76	22.36	24.39	51.22	58.54	61.79	60.98	55.69	55.28	54.88	54.88	63.01	61.79	57.32	60.16	54.47	58.54	55.28	54.47

Appendix

Table 11: Per-subtask zero-shot accuracy (%) on the AGIEval-CN direct subset under different generation budgets. The subset excludes Gaokao MathCloze because it does not fit our unified direct-answer evaluation protocol. We separate non-Wenyan baselines/controls from Wenyan-based variants for clarity. Boldface marks the best accuracy.

Group	Model	Accuracy (%)				Time (Latency)
Group	Model	1024	2048	4096	6144	1024	2048	4096	6144
Non-Wenyan / Controls	Qwen3-4B-Thinking-2507	14.20	24.37	33.49	35.89	24:20	1:06:19	2:35:33	3:36:37
	Long Vernacular SFT Reasoning	40.89	47.82	52.22	52.97	19:59	40:59	1:39:32	3:02:33
	Compact modern Chinese control	48.94	49.45	49.80	49.83	05:28	06:11	06:48	07:34
Wenyan Variants	Wenyan (DeepSeek-V4-Flash teacher)	52.15	52.69	52.08	51.87	07:09	07:56	10:42	13:21
Wenyan Variants	Wenyan Default (Ours)	49.92	51.55	51.43	51.27	07:54	08:53	10:14	11:20

Appendix

Table 12: Zero-shot accuracy (%) and total evaluation time on the AGIEval-CN direct subset under different generation budgets. The subset excludes Gaokao MathCloze. We separate non-Wenyan baselines/controls from Wenyan-based variants for clarity. Boldface marks the best (highest accuracy, lowest latency).

Subject	1024	2048	4096	6144
Biology	63.9	63.9	63.9	63.9
Business	65.0	65.0	65.0	65.0
Chemistry	53.6	62.5	62.5	62.5
Computer Science	55.0	60.0	60.0	60.0
Economics	73.8	73.8	73.8	73.8
Engineering	25.0	29.2	29.2	31.2

Appendix

Table 13: Subject-wise exact-match (%) of the final main system (Qwen3.5-397B-A17B normal Wenyan prompt) under all generation budgets. Boldface marks the peak accuracy achieved for each subject across the different generation budgets.

Training target during SFT	1024	2048	4096	6144
Full completion ( thinking + ¡/think¿ + answer )	52.9	55.3	55.4	55.6
Thinking only ( thinking + ¡/think¿ )	52.4	54.1	54.6	54.9
Drop from full completion	-0.5	-1.2	-0.8	-0.7

Appendix

Table 14: Exact-match (%) on MMLU-ProX-Lite-ZH : full-completion vs. thinking-only supervision across generation budgets.

Efficient Language Model Reasoning

Efficient language model reasoning studies how to improve problem-solving accuracy while reducing inference, memory, latency, and training costs.

See all 370 papers

Cost of Reasoning in non-English Languages: A Case Study on Japanese
Jul 11, 2026Yuu JinnaiCross-Lingual Reasoning in Language Models Multilingual Language Model Evaluation
CyberAgent / Tokyo, Japan
Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens
Feb 13, 2026Wei-Lin Chen, Liqian Peng, Tian Tan +5LLM Inference Efficiency Inference-Time Scaling
University of Virginia · Google
Draft-Thinking: Learning Efficient Reasoning in Long Chain-of-Thought LLMs
Feb 28, 2026Jie Cao, Tianwei Lin, Juekai Lin +7Efficient Language Model Reasoning CoT Reasoning
Zhejiang University · Tencent

Jul 11, 2026J/K move · Enter open · S save

Cost of Reasoning in non-English Languages: A Case Study on Japanese

Yuu Jinnai

CyberAgent / Tokyo, Japan

Cross-Lingual Reasoning in Language Models Multilingual Language Model Evaluation RL for Language Model Reasoning Large Reasoning Models Multilingual Language Models Structured Reasoning in Language Models+1

Open paper AskPDF arXiv

Reasoning Language Models (RLMs) achieve their strongest performance when they reason in English, the language for which reasoning-oriented training data is most abundant. However, reasoning trace is a clue for model interpretability and safety, and useful in practice for both the model users and for model developers. Thus, it is desirable to be able to develop a model that reasons in a language of the user's choice, while still maintaining strong reasoning performance. To this end, we study the feasibility of training a model that reasons in Japanese. We develop a Japanese-reasoning variant of Qwen-3-Swallow-8B, which is a Japanese LLM continually pretrained from Qwen-3-8B, with GRPO and evaluate it across coding, math, and science benchmarks. The study shows that reasoning-language control is feasible by training a Japanese continually pretrained model with GRPO. However, its performance is at best on par with strong English-reasoning baselines on several benchmarks. We also evaluate the trained model on Japanese cultural benchmarks and observe that the model's performance is worse than the baseline models, suggesting that the reasoning in Japanese does not immediately improve performance on culturally relevant tasks for free.

Abstract

Efficient Language Model Reasoning

Explore similar work

Cost of Reasoning in non-English Languages: A Case Study on Japanese

Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens

Draft-Thinking: Learning Efficient Reasoning in Long Chain-of-Thought LLMs