DeepSeek-R1-Zero produces chain-of-thought responses with poor readability and English-Chinese language mixing.
4 events · 1 assessment · 2 decisions
Structured and assessed
First pass. Kept the claim atomic, as a single developer-documented observation. Corrected the arXiv instance metadata (speaker DeepSeek-AI, date 2025-01). Recorded the HF model card as a second instance from the same voice and recorded the shared-authorship relation. Wrote an immaterial source map. Set importance 0.2 and contestation 0.1. Assessed supported (confidence 0.85, credence 0.95). The parent claim on RL illegibility already treats this as supporting evidence, and a first assessment that agrees with the edge needs no notification.
Assessed Supported
verdict confidence 0.85 · credence 0.95
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.
Updated claim instance
Instance f0a4b40a-641e-4366-92bb-f20a2779a1e0 (https://arxiv.org/pdf/2501.12948): was stance=affirms, confidence=0.85; set confidence=1, speaker="DeepSeek-AI", publication="arXiv (2501.12948)", sourceDate="2025-01". The source is the DeepSeek-R1 technical report by DeepSeek-AI, posted to arXiv in January 2025; the passage is a direct assertion of the claim by the model's developers.
Claim entered the graph