PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Minghan Li, Guodong Zhou

arXiv:2610.19741v1cs.CLcs.AI

TL;DR

The paper asks whether Chinese LLMs can make visible reasoning more efficient by changing its linguistic register. It trains Wenyan Thinking on Classical Chinese rewrites and finds shorter traces and a favorable budget-sensitive trade-off, with compact rewriting accounting for much of the efficiency gain.

  • Problem

    Chinese reasoning traces usually use modern vernacular, and whether changing their register can improve efficiency under limited decoding budgets has received little direct attention.

    Source: §1 · p. 1

  • Method

    Wenyan Thinking rewrites modern-Chinese reasoning traces into Classical Chinese for visible reasoning while retaining a short direct-answer channel, alongside a compact modern-Chinese control.

    Source: §1 · pp. 1–2, §3.3 · p. 4

  • Results

    Across Chinese benchmarks, Wenyan Thinking shortens traces and reduces missing-answer rates; compact rewriting explains much of the efficiency gain, and wenyan exceeds the compact modern-Chinese control on the primary benchmark.

    Source: §1 · p. 2, §5.1 · p. 6, §8 · p. 8

  • Takeaways & Limitations

    The findings identify linguistic register as a practical design axis for efficient visible reasoning in Chinese LLMs.

    Source: §8 · p. 8

    Evidence is limited to Chinese benchmarks and a 10,000-example rewritten subset, with the primary benchmark’s 588-example Lite subset limiting estimates of small register gaps.

    Source: Limitations · p. 8

Abstract

from arXiv · show

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.

1 Introduction

Visible reasoning can improve difficult-problem performance but consumes tokens and latency, making its register relevant under limited generation budgets. The paper proposes Wenyan Thinking and a matched compact modern-Chinese control to study whether Classical Chinese can shorten reasoning while preserving direct answers.

1 Introduction

  • Long visible reasoning consumes tokens and latency, and can delay, truncate, or incorrectly format the final answer.This motivates treating inference-time reasoning budget as a design dimension.Source: §1 · p. 1
  • The paper asks whether shifting Chinese LLM reasoning from modern vernacular Chinese to Classical Chinese can improve efficiency while retaining task performance.The final answer remains in a short direct-answer format.Source: §1 · p. 1
  • Wenyan Thinking uses Classical Chinese for visible intermediate reasoning while preserving a benchmark-compatible direct-answer output.The approach rewrites modern-Chinese reasoning traces and fine-tunes a Chinese thinking model.Source: §1 · pp. 1–2
  • A compact modern-Chinese control uses the same source examples, teacher family, base model, and training pipeline to distinguish generic shortening from register shift.The comparison is intended to separate compact rewriting from the specific effect of wenyan.Source: §1 · p. 2
  • Wenyan supervision yields a favorable accuracy–budget–latency trade-off under constrained decoding.The paper evaluates this claim against multiple baselines.Source: §1 · p. 2

2 Related Work

Prior work establishes the value of reasoning content and explores ways to improve reasoning efficiency, but generally treats the linguistic form of traces as fixed. This paper positions wenyan as a register-based compression axis for Chinese visible reasoning.

Visible reasoning and chain-of-thought

  • Chain-of-thought research spans prompting, decomposition, self-consistency, tool-mediated reasoning, search, and reasoning-oriented post-training, while treating trace language largely as fixed.The cited work includes Wei et al. (2022), Kojima et al. (2022), and later approaches.Source: §2 · p. 2

Reasoning efficiency and token economy

  • Reasoning-efficiency work compresses content, supervision, decoding, or hidden representations; this paper instead studies linguistic register without architectural changes or specialized tokenizers.The surveyed approaches include length control, token-budget-aware training, rationale compression, prompt compressors, and latent-space reasoning.Source: §2 · p. 2

Multilingual reasoning and Classical Chinese in NLP

  • Multilingual studies mostly compare modern languages, while Classical Chinese NLP has focused on translation, understanding benchmarks, and specialized ancient-text models rather than reasoning as a medium.The paper frames wenyan as a compact surface channel for visible LLM reasoning.Source: §2 · pp. 2–3

Chinese reasoning models

  • Prior work has not systematically asked whether Chinese models should reason visibly in modern vernacular Chinese or whether shifting to wenyan can improve efficiency.This paper opens that question for Chinese reasoning models.Source: §2 · p. 3

3 Method

The method changes the visible reasoning register while keeping the final answer channel short and direct. It builds rewritten supervision and fine-tunes the same thinking-model interface, with compact modern Chinese as a controlled comparison.

3.1 Task Formulation

  • The target format pairs Classical Chinese intermediate reasoning with a short benchmark-required direct answer, while the compact modern-Chinese control rewrites the same source trace concisely.The task formulation hypothesizes shorter wenyan traces with preserved answer quality.Source: §3.1 · p. 3

3.2 Data Construction

  • The 10,000 selected examples retain their questions and answers while rewriting only the reasoning field; the compact control uses the same examples, teacher family, and SFT format.Qwen3.5-397B-A17B is the main rewriting teacher, with step structure intended to be preserved.Source: §3.2 · p. 3
  • Mean reasoning-field length falls from 1457.8 Chinese characters in the source data to 420.9 in the wenyan rewrites before student-model decoding.The final training JSONL contains 10,000 usable records.Source: §3.2 · p. 3

3.3 Supervised Fine-Tuning

  • Full-parameter SFT trains Qwen3-4B-Thinking-2507 to emit rewritten reasoning followed by the preserved answer, supervising completion tokens only.Benchmark prompts retain the short direct-answer format; the compact control uses the same recipe and answer format.Source: §3.3 · p. 4
  • Table 1 summarizes WENYANTHINKING-10K field lengths and the final main system’s full-parameter SFT configuration.The reported implementation uses two training epochs.Source: Table 1 · p. 4, §3.3 · p. 4

4 Experimental Setup

The experiments compare systems across Chinese benchmarks and controlled generation budgets, measuring both accuracy and generation efficiency. A matched compact-reasoning baseline helps distinguish trained rewriting from prompting or answer-only evaluation.

4.1 Benchmarks

  • Evaluation centers on a fixed 588-example MMLU-ProX-Lite-ZH subset and also includes CEVAL-VALID-DIRECT and an AGIEval-CN direct subset.The Lite subset supports budget sweeps while retaining subject coverage across 14 groups.Source: §4.1 · pp. 4–5

4.2 Compared Systems

  • Compared systems include the original thinking model, answer-only and concise-prompt baselines, long and compact modern-Chinese SFT, and Wenyan Thinking.Supporting variants test alternative teachers and rewrite prompt styles.Source: §4.2 · p. 5
  • Table 2 reports exact-match accuracy on MMLU-ProX-Lite-ZH across generation budgets, with Wenyan denoting the final Qwen3.5-based main system.The caption identifies the concise-prompt and answer-only baselines and marks the best and second-best accuracy per budget.Source: Table 2 · p. 5

4.3 Budget-Sensitive Evaluation and Metrics

  • The evaluation varies generation budgets across 1024, 2048, 4096, and 6144 tokens and reports accuracy, reasoning and total output length, missing-answer rate, and evaluation time.Decoding is deterministic; compared systems in a latency table share the same harness and vLLM setup.Source: §4.3 · p. 5

5 Main Results on the Primary Benchmark

On MMLU-PROX-LITE-ZH, Wenyan reaches strong accuracy at shorter generation budgets than long-form thinking, while compact rewriting accounts for much of the efficiency gain. Subject-level results, answer-completion analysis, training-objective comparisons, and a rewrite audit qualify this trade-off.

5.1 Overall Comparison Under Different Generation Budgets

  • 11.4→52.9 at 1024 tokens, 32.3→55.3 at 2048, and 50.9→55.4 at 4096: Wenyan improves over Qwen3-4B-Thinking-2507 under short and medium budgets.Wenyan remains consistently above the compact modern-Chinese control on the primary benchmark, but the margin is modest.Source: §5.1 · p. 6

5.2 Subject-Level Behavior

  • At 1024 tokens, Wenyan outperforms long-form thinking across all shown subjects; at 6144, long-form thinking recovers on some subjects, especially math, while Wenyan remains competitive with shorter traces.The subject-level pattern supports the interpretation that wenyan supervision is most valuable when reasoning budget is scarce.Source: §5.2 · p. 6, Table 3 · p. 7

5.3 Efficiency and Answer Completion

  • At 1024 tokens, the base model emits 1008.1 tokens with an 87.5% missing-answer rate and 11.4% accuracy, versus Wenyan’s 394.2 tokens, 7.5% missing rate, and 52.9% accuracy.At 4096 tokens, Wenyan uses 458.4 tokens versus 2901.8 for the base model; compact modern Chinese is shorter with lower missing-answer rates, while Wenyan has higher primary-benchmark accuracy.Source: §5.3 · p. 6, Table 4 · p. 7

5.4 Does the Answer Channel Drive the Gain?

  • Removing answer-token supervision lowers accuracy by only 0.5–1.2 points, indicating that compact, stable wenyan reasoning style provides the main benefit while answer supervision modestly stabilizes the transition to the final answer.Source: §5.4 · p. 6

5.5 LLM-as-Judge Audit of Rewrite Quality

  • In a 40-example automatic-judge audit, faithfulness, step preservation, readability, and answer support each average above 4.3, with leakage risk at 1.00 and no major-error cases; the authors treat this as supporting evidence, not a substitute for human annotation.Verdicts were 27 pass, 11 minor issue, and 2 major issue; non-pass cases were mostly omissions in open-ended explanatory examples rather than answer-changing errors.Source: §5.5 · p. 6, Table 5 · p. 7

6 Cross-Benchmark Validation

Cross-benchmark evaluations show that Wenyan performs strongly at tighter decoding budgets on CEVAL and supports the efficiency pattern on AGIEval-CN. At larger budgets, it remains competitive, with accuracy and evaluation time varying across compact-reasoning systems.

6.1 CEVAL

  • On CEVAL, Wenyan is strong at 1024–2048 tokens and remains competitive at larger budgets, while the original Qwen3-4B-Thinking-2507 model needs 4096–6144 tokens to overtake compact-reasoning systems.The reported pattern points to early answer completion as the main benefit under realistic budgets.Source: §6.1 · p. 7, Table 6 · p. 7

6.2 AGIEval-CN

  • At 4096 tokens on AGIEval-CN, Wenyan reaches 51.43 accuracy in 10:14, compared with the base thinking model’s 33.49 in 2:35:33.Compact modern Chinese is faster, while Wenyan trades additional generation time for higher accuracy at 1024–2048 tokens.Source: §6.2 · pp. 7–8, Table 7 · p. 8

7 Discussion

The discussion presents compact reasoning as central to efficiency while arguing that wenyan can retain a modest accuracy advantage over compact modern Chinese in selected settings. It frames reasoning-register selection as a design variable, not a universally optimal recipe.

7.1 Main Takeaways and Interpretation

  • Compact, logic-preserving rewriting shortens reasoning traces and improves the accuracy–budget–latency trade-off under constrained decoding.The compact modern-Chinese control indicates that shortening is an important mechanism; the wenyan system also supports register as a meaningful design variable.Source: §7.1 · p. 8
  • Teacher choice and rewrite style can dominate register effects on some benchmarks, so register shift is a design axis rather than a one-size-fits-all recipe.The discussion says effective supervision must be compact, learnable, and faithful to the original reasoning steps.Source: §7.1 · p. 8
  • Wenyan achieves higher accuracy than compact modern Chinese on the primary benchmark and low-budget CEVAL, although compact modern Chinese is often shortest and has the lowest missing-answer rate.The paper characterizes wenyan as a compact-but-expressive point on the compression frontier, not as the shortest or universally best register.Source: §7.1 · p. 8

7.2 Broader Implications and Risks

  • Other languages with diglossic or historical varieties may offer analogous compact reasoning registers, and intermediate-computation surface forms merit systematic study.These are broader implications proposed by the paper, beyond its Chinese-focused evidence.Source: §7.2 · p. 8

8 Conclusion

The paper proposes Wenyan Thinking, which uses Classical Chinese for visible intermediate reasoning while retaining a short direct-answer format. Across three Chinese reasoning benchmarks and controlled generation budgets, it reports shorter traces, fewer missing answers, and a strong accuracy–budget–latency trade-off.

8 Conclusion

  • Across three Chinese reasoning benchmarks under controlled generation budgets, Wenyan Thinking shortens emitted traces, reduces missing-answer rates, and yields a strong accuracy–budget–latency trade-off.The compact modern-Chinese control confirms that shortening is central; wenyan is presented as a practical design axis for efficient visible reasoning.Source: §8 · p. 8
  • Wenyan Thinking emits visible intermediate reasoning in Classical Chinese while keeping the final answer in a short direct-answer format.The conclusion describes this as a training setting for Chinese LLMs.Source: §8 · p. 8
  • The wenyan system offers a competitive compact register that often trades slightly longer traces for higher accuracy.This conclusion does not claim wenyan is the shortest option.Source: §8 · p. 8

Limitations

The study’s evidence is bounded by its Chinese benchmark focus, a 10,000-example rewritten training subset, and a 588-example primary Lite evaluation. The authors also note that compact reasoning is harder to audit than longer vernacular traces and that automatic judging does not replace full human evaluation.

Limitations

  • The study focuses on Chinese reasoning benchmarks and a 10,000-example rewritten training subset; broader task coverage and additional base-model families remain future work.The primary benchmark is a 588-example Lite subset, and larger-scale evaluation would better estimate small gaps between compact registers.Source: Limitations · p. 8
  • Compact reasoning can be harder to audit than longer vernacular traces, and the automatic LLM-as-judge audit does not replace full human evaluation of faithfulness, step preservation, or readability.The authors identify those dimensions as needing fuller human assessment.Source: Limitations · p. 9
  • The answer-only baseline uses an instruct sibling rather than a strictly identical no-thinking decoding of the same base model, and evidence is strongest for Chinese.Whether analogous registers help in other languages remains open.Source: Limitations · p. 9

A.1 Data, Evaluation, and Decontamination Details

The appendix documents evaluation scope and decontamination, a rewrite-quality audit, complete benchmark tables, and the prompt variants used for data generation. It supplies supplementary evidence and implementation detail without changing the paper’s central claim.

A.1 Data, Evaluation, and Decontamination Details

  • MMLU-ProX-Lite-ZH contains 588 examples across 14 subject groups, while the AGIEval-CN direct subset excludes Gaokao MathCloze because it does not fit the unified direct-answer protocol.The decontamination screen found zero exact question overlaps across 296,681 scanned evaluation questions; a high-overlap screen flagged repeated logs of one AGIEval-style geography example, not affecting the primary evaluation.Source: §A.1 · p. 11

A.2 LLM-as-Judge Rewrite Quality Audit

  • The rewrite audit reports 27 pass, 11 minor-issue, and 2 major-issue verdicts among 40 examples, with no major factual or logical errors identified.The two major-issue cases compressed exploratory reasoning into more direct traces; Table 8 identifies the audit dimensions as faithfulness, step preservation, readability/reversibility, and answer support, among others.Source: §A.2 · p. 11, Table 8 · p. 11

A.3 Complete Leaderboard of MMLU-ProX-Lite-ZH Over Evaluated Systems

  • Table 9 provides a complete MMLU-ProX-Lite-ZH leaderboard covering the main system, compact modern-Chinese controls, prompt-style variants, and teacher-source variants.Its caption identifies exact-match accuracy and describes the matched compact modern-Chinese control; the supplied passages do not provide cell values.Source: §A.3 · p. 12, Table 9 · p. 13

A.4 Complete AGIEval-CN Results

  • The appendix reports AGIEval-CN direct-subset per-subtask results and an overall accuracy–latency summary, excluding Gaokao MathCloze under the unified direct-answer protocol.The supplied captions describe the tables’ scope but do not provide their result values.Source: §A.4 · p. 12, Table 11 · p. 14, Table 12 · p. 14

A.5 MMLU-ProX-Lite-ZH Full Subject-by-Budget Results of the Main System

  • Table 13 reports subject-wise exact-match results for the final Qwen3.5-397B-A17B normal Wenyan prompt system under all four generation budgets.The appendix presents this as a consistency check that reported improvements are not limited to a hand-picked subset of subjects.Source: §A.5 · p. 12, Table 13 · p. 14

B Prompt Templates

  • The prompt appendix includes normal, faithful, and ultra-compact Wenyan rewriting variants, plus a compact modern-Chinese control that explicitly avoids wenyan.The faithful prompt prioritizes preserving reasoning steps and structure; the ultra-compact prompt retains only the minimal sufficient reasoning chain, while the modern-Chinese control preserves key steps while removing redundancy.Source: Figure 2 · p. 12, B Prompt Templates · pp. 12, 15, Figure 3 · p. 15

C Ablation on the Supervised Target (Full Results)

The supervised-target ablation finds a modest benefit from answer-channel supervision, while the case study contrasts verbose conversational reasoning with compact modern-Chinese and Wenyan traces. The examples retain the same correct answer while varying the density and structure of their reasoning.

C Ablation on the Supervised Target (Full Results)

D Case Study: Reasoning Trace Comparison on an MMLU-ProX Physics Example

  • All five systems answer the representative MMLU-ProX physics question correctly, despite differing reasoning styles.The example asks for the escape-velocity formula.Source: D Case Study: Reasoning Trace… · p. 16

D.1 Problem Statement

D.2 Chinese Reasoning Comparison

  • Wenyan faithful reasoning preserves step-by-step derivation and verification, while the ultra-compact variant retains only a minimal derivation and option check.The normal-prompt Wenyan trace is short and dense, retaining the main derivation and lightweight option filtering.Source: §D.2 · pp. 16–17

D.3 English Translation Comparison

The translated examples show distinct reasoning styles across vernacular, compact modern Chinese, and Wenyan outputs. The case-study observation finds that compact rewriting removes conversational repetition, while Wenyan expresses reasoning in a distinct compact register.

D.3 English Translation Comparison

  • The translated compact-control trace derives the escape-velocity formula and selects option C.It states the energy relation, cancels the object’s mass, and obtains the velocity expression.Source: §D.3 · p. 18
  • The vernacular-reasoning SFT example begins with uncertainty and plans to check the options, illustrating its conversational style.The example continues into derivation and verification in the following passage.Source: §D.3 · p. 18

D.4 Brief Observation

  • Compact modern-Chinese reasoning removes colloquial repetition while retaining the essential derivation; Wenyan variants range from dense traces to more structured verification.The ultra-compact variant keeps a minimal derivation and final option check, while the faithful variant preserves more step-by-step structure.Source: §D.4 · p. 19

Related papers

Loading 2610.19741v1…