DeepSeek-R1-Zero produces chain-of-thought responses with poor readability and English-Chinese language mixing.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
DeepSeek-R1-Zero is the model DeepSeek-AI trained with large-scale reinforcement learning directly on DeepSeek-V3-Base, with no supervised fine-tuning first. Its developers report that its reasoning traces are often hard to read and sometimes mix English and Chinese within a single chain of thought. The report (arXiv 2501.12948, January 2025) says this, and so does the R1-Zero model card, which also mentions endless repetition. The developers' design choices back the report: to fix these problems they built DeepSeek-R1 with a "cold start" of curated readable reasoning data and a reward for language consistency.
The evidence is the developers' testimony about their own model, and it is qualitative. The report gives no measured rate of language mixing for R1-Zero. No source disputes the observation, and it fits what one would expect when an outcome-only reward places no constraint on the language or form of intermediate reasoning. What has not been located is a quantified, independent measurement of R1-Zero's outputs, and that is the main reason this falls short of verified.
Full reasoning: the evidence and decisions behind this verdict
Prior: R1-Zero is optimized only on answer correctness and format tags, so nothing in the reward penalizes switching languages or terse, idiosyncratic reasoning. A bilingual base model drifting between English and Chinese under such pressure is expected. The R1 paper is the obvious primary source, and my prior was that it states exactly this.
Evidence checked: the arXiv HTML of the R1 report (arxiv.org/html/2501.12948v1) states in the abstract, introduction and section 2.2.4 that R1-Zero "encounters challenges such as poor readability, and language mixing." Section 2.3 adds that R1-Zero's content "is often not suitable for reading," and that this motivated the cold-start data. The recorded instance passage, which mentions combining English and Chinese within a single chain-of-thought response, matched the stored text verbatim. The model card (huggingface.co/deepseek-ai/DeepSeek-R1-Zero) repeats the claim and adds endless repetition. Both sources come from the same developer, so they are one voice.
Weighing: the developer has no evident motive to invent a flaw in its own model, and the corrective steps it describes (a language-consistency reward, cold-start data) only make sense if the problem was real. The paper also reports that this reward slightly lowers benchmark performance, a cost the developers accepted to fix the issue. The claim has no counter-evidence. The limits are that "poor readability" is a qualitative judgment and that no independent quantification was located in this light pass. Hence supported with high credence rather than verified. Independent measurements of R1-Zero's language-mixing rate would raise this to verified, or would qualify "poor readability" if mixing turned out to be rare.
The claim is kept atomic. It is a single observation about one model, and nothing beneath it needs its own node.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this claim has been said, linked to its canonical form.
Although DeepSeek-R1-Zero demonstrates excellent reasoning capabilities, it faces chal- lenges such as poor readability and language mixing, occasionally combining English and Chinese within a single chain-of-thought response.
Introduction, describing limitations of the pure-RL model.
The source's own evidence bears what it asserts. The developers' own report. It describes the problem qualitatively and backs it indirectly by introducing a language-consistency reward to correct it; it gives no quantified rate of mixing for R1-Zero itself. Read through search excerpts of the HTML version rather than in full.
However, DeepSeek-R1-Zero encounters challenges such as endless repetition, poor readability, and language mixing.
Model card for the released R1-Zero weights, summarizing the model's limitations and motivating DeepSeek-R1.
How these sources relate
- deepseek-ai/DeepSeek-R1-Zero and https://arxiv.org/pdf/2501.12948 share an author. Both the technical report and the Hugging Face model card are published by DeepSeek-AI and carry the same description of R1-Zero's limitations.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Oct 28, 2026. Every judgment on this page is accompanied by a reasoning trace.