Does the Linguistic Register of Thinking Matter? Wenyan as a Compact Reasoning Register for Chinese LLMs
The paper tackles the token inefficiency of visible reasoning traces in Chinese LLMs, which are typically generated in modern vernacular Chinese. It proposes Wenyan Thinking, a method that rewrites intermediate reasoning into Classical Chinese (wenyanwen) to produce more compact traces while keeping final answers short and benchmark-compatible. The work matters because it treats linguistic register itself as a controllable compression axis for reasoning, offering a practical way to improve accuracy under tight generation budgets without architectural changes.
The paper presents a focused empirical question and answers it with a well-controlled comparison. The authors are careful to attribute gains jointly to compact rewriting and to the wenyan register, and they avoid overstating the uniqueness of classical Chinese as a reasoning medium. However, the evidence for register-specific benefits beyond generic compression is modest: on the primary benchmark MMLU-ProX-Lite-ZH, the Wenyan model edges out the Compact Modern-Chinese control by only 0.7–2.4 points across budgets, and on CEVAL and AGIEval-CN the two are often statistically indistinguishable or mixed. The authors acknowledge this directly, noting that "Compact rewriting explains much of the efficiency gain" and that the Wenyan–Compact gap should be read as "a modest accuracy advantage on our primary benchmark, not as proof that wenyan is always the shortest or universally best register." The study is therefore best read as a careful proof-of-concept that register choice matters, rather than a definitive demonstration that classical Chinese is intrinsically superior.
The experimental design is the paper’s greatest strength. By constructing a Compact Modern-Chinese control from the "same source examples, teacher family, base model, and downstream SFT recipe," the authors isolate register shift from generic length reduction. The budget-sensitive evaluation protocol—sweeping max_gen_toks over 1024, 2048, 4096, and 6144—is exactly the right frame for reasoning models, and the reported missing-answer rates make the operational mechanism transparent: at 1024 tokens the base model has an 87.5% missing-answer rate versus 7.5% for Wenyan. The observation that "the reasoning and answer channels can be decoupled in format" is supported cleanly by the ablation in Section 5.4, where masking answer supervision drops accuracy by only 0.5–1.2 points.
The primary limitation is scale and generalizability. The entire study rests on a 10,000-example rewritten subset and a single base model family (Qwen3-4B-Thinking-2507), leaving open whether the effect transfers to larger models, other Chinese-capable families, or longer training runs. The primary benchmark is itself a 588-example Lite subset, which the authors admit is "intended for budget sweeps" but limits confidence in small accuracy gaps. More fundamentally, the mechanism underlying any register-specific advantage remains speculative: the paper hypothesizes that wenyan provides "enough connective structure for the student model to learn a stable reasoning channel," yet offers no causal evidence that this structure is uniquely classical rather than an artifact of the Qwen3.5 teacher’s rewriting style. The LLM-as-judge audit is also limited to 40 samples and a single judge, which the authors concede is "supporting evidence, not a replacement for human annotation."
The evidence broadly supports the central claim that compact reasoning supervision improves the accuracy–budget trade-off, but it only weakly supports the stronger claim that the wenyan register itself adds value over equally compact modern Chinese. On MMLU-ProX-Lite-ZH, Wenyan consistently outperforms the compact control, yet the margin is small enough that the authors wisely retreat to the narrower interpretation that wenyan is "a viable compact register." The comparisons to related work are fair: the authors distinguish their register-shift approach from prompt compression, latent-space reasoning, and length-controlled SFT, and they appropriately note that "Our work complements them by studying a linguistic register as the compression axis, without requiring architectural changes or specialized tokenizers." The sensitivity analyses in Appendix A.3, which vary teacher model and prompt style, further strengthen the empirical grounding by showing that results are not tied to a single rewriting pipeline.
Reproduction is partially feasible but faces several blockers. The training data (WenyanThinking-10k) is derived from Chinese-DeepSeek-R1-Distill-data-110k and rewritten via the Qwen3.5-397B-A17B API; the paper states the rewrite prompts in Appendix B, but the exact API calls, temperature settings, and failure-repair logic for the 87 initially failed records are not specified. Full-parameter SFT hyperparameters are reported in Table 1, including learning rate $1\times 10^{-5}$, 2 epochs, bf16, and DeepSpeed ZeRO-3, yet the authors explicitly decline to specify "the exact GPU topology," which complicates wall-clock latency comparisons. No code repository or model checkpoint is mentioned, and the evaluation relies on a custom Lite subset whose stratified selection protocol is described only in brief. An independent team could replicate the pipeline from the prompts and hyperparameters, but exact reproduction of the missing-answer-rate dynamics would require access to the same distilled-source slice and API teacher outputs.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.