DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

• p. 66 Section I code listing

  cd DeepSeek-R1/inference

Judge's note: Page image p.66: the comment reads '# Clone DeepSeek-V3 GitHub repository', followed by 'git clone .../DeepSeek-V3.git' and then 'cd DeepSeek-R1/inference'. Page 65 (text lines 4345-4347) says the inference code is in the DeepSeek-V3 repository. The weights are downloaded to /path/to/DeepSeek-R1, which has no inference folder. The cd path appears to be inconsistent with the clone step.

Second check: another model agreed

• p. 2 Introduction, para 4

  We believe that these instruction-tuned versions will also significantly contribute

Judge's note: Checked the crop of p.2. The wording is as quoted. The intended referent (the distilled models) can be recovered from context, and distilled models are themselves fine-tuned, so this is a minor ambiguity rather than an error.

Second check: another model agreed

• p. 4 Figure 1 caption

  Pass@1 and Cons@16 are described in Supplementary D.1

• p. 41 D.1

  consensus (majority vote) results using 64 samples, denoted as cons@64

• p. 41 D.1

  we also report consensus (majority vote) results using 64 samples, denoted as cons@64

• p. 4 Figure 1 legend and caption

  Pass@1 and Cons@16 are

Judge's note: Checked the Figure 1 caption (p.4) and D.1 (p.41). A search of the text finds Cons@16 only in the figure legend and caption. D.1 does give the majority-vote procedure, so the pointer is partly valid, but the 16-sample setting is not described in the cited section.

Second check: another model agreed

• p. 8 Section 4

  most notably on the AIME benchmark

• p. 9 Table 3, CNMO 2024 row

  88.1

• p. 9 Table 3

  CNMO 2024 (Pass@1) 88.1 58.0

Judge's note: Checked the values in Table 3 (p.9) on the page image. The CNMO drop is larger in both absolute and relative terms, so 'most notably' is loose wording.

Second check: another model agreed

• p. 8 Section 4

  and CMMLU (Li et al., 2024)

• p. 39 D.1 Benchmarks

  MMLU, MMLU-Redux, MMLU-Pro, C-Eval, and CMMLU are multiple-choice

• p. 39 D.1 Benchmarks

  Specifically, MMLU, MMLU-Redux, MMLU-Pro, C-Eval, and CMMLU are multiple-choice benchmarks

• p. 39 D.1 first sentence

  We evaluate models on MMLU (Hendrycks et al., 2021), MMLU-Redux (Gema et al., 2025), MMLU-Pro (Wang et al., 2024), C-Eval (Huang et al., 2023), IFEval

Only the authors can say what is right here.

Judge's note: Checked p.8 and p.39. A text search finds CMMLU only in the two benchmark-list sentences and in the reference list. Tables 3 and 12, which I viewed, have no CMMLU row.

Second check: another model agreed

• p. 9 Table 3, FRAMES row

  82.5

• p. 56 Table 12, FRAMES row

  82.5

Only the authors can say what is right here.

Judge's note: Checked both tables on the page images. In both, R1 is the highest entry in the FRAMES row, yet the significance marking differs. Only the authors can say which marking is correct.

Second check: another model agreed

• p. 9 Section 4 text

  AlpacaEval 2.0 improving by 25% and ArenaHard by 17%

Judge's note: Checked the Dev3 and R1 columns of Table 3 on p.9. The printed figures match point differences, so the text should read 'percentage points'.

Second check: another model agreed

• p. 9 Results text after Table 3

  Marginal improvements occurred in code and mathematics benchmarks

• p. 9 Table 3

  Codeforces (Rating) 1746 → 2029; Aider-Polyglot 44.8 → 53.3

Judge's note: Checked Table 3 on the page image. The math and LiveCodeBench gains are small, but the Codeforces and Aider gains are not marginal, so the wording overgeneralises.

Second check: another model agreed

• p. 35 B.4.4 Training Cost

  scale up to 660B R1-Zero and R1

• p. 13 A.1 DeepSeek-V3

  DeepSeek V3 has 671 billion total parameters

• p. 41 Table 8, # Total Params

  671B

• p. 35 Appendix B.4.4

  confidently scale up to 660B R1-Zero and R1

Only the authors can say what is right here.

Judge's note: Checked the B.4.4 crop and Table 8 on p.41 (Total Params 671B for DeepSeek-R1). Nothing in the paper explains 660B as a different count. The same paragraph gives the smaller model as 30B, while 32B is used elsewhere. Only the authors can say whether 660B is a deliberate approximation.

Second check: another model agreed

• p. 19 Table 4, General row

  66K

• p. 20 B.3.1 General bullet

  Additionally, the dataset includes 12,000 questions focused on evaluating harmlessness.

Only the authors can say what is right here.

Judge's note: Checked Table 4 (p.19) and the General bullet (p.20). The wording could be read as 12k being a subset of the 66k, but 'additionally' and the description of the 66k as helpfulness questions suggest a separate set. Unclear which count is meant.

Second check: another model agreed

• p. 34 B.4.1 heading

  Hyper-Parameters of DeepSeek-R1-Zero-Qwen-32B

• p. 62 F.1

  resulting in Qwen2.5-32B-Zero, as described in B.4.1

• p. 34 B.4.1

  To train DeepSeek-R1-Zero-Qwen-32B

• p. 39 D.1 Benchmarks

  Specifically, MMLU, MMLU-Redux, MMLU-Pro, C-Eval, and CMMLU are multiple-choice

Judge's note: Both crops confirm the two names; F.1 points to B.4.1 for the model's settings. The model is evidently the same, but no sentence links the two names. Side point: the D.1 benchmark list (p.39) leaves out CMMLU and DROP, which the next paragraph then describes. This is a minor list inconsistency.

Second check: another model agreed

• p. 35 B.4.4 Training Cost

  completing the process in about 4 days, or roughly 80 hours

• p. 37 Table 7

  41K

Judge's note: Checked the B.4.4 crop and Table 7 on p.37 (R1 41K, R1-Zero 101K, total 147K). R1-Zero's 198 h × 512 GPUs gives about 101K, which matches the table. R1's 80 h × 512 gives about 41K, so the "4 days" wording is the inconsistent element.

Second check: another model agreed

• p. 40 D.1 Evaluation Prompts

  the HumanEval-Mul dataset covers eight mainstream programming languages

Judge's note: A text search finds HumanEval-Mul only in this sentence. Table 8 on p.41 has no HumanEval-Mul row. The sentence appears to be carried over from the DeepSeek-V3 setup, whose prompts this paragraph says it follows.

Second check: another model agreed

• p. 41 D.2 Standard Benchmark

  DeepSeek-R1 also delivers impressive results on IF-Eval

• p. 41 Table 8, IF-Eval row

  83.3

• p. 41 Table 8

  IF-Eval (Prompt Strict) 86.5 84.3 86.1 84.8 - 83.3

Judge's note: Checked the Table 8 IF-Eval row on the p.41 image. The text does not say what the "improvements" are measured against; it may mean the earlier R1 stages. Next to Table 8, the wording overstates R1's relative standing.

Second check: another model agreed

• p. 43 Human Evaluation paragraph

  DeepSeek-R1 shares the first position alongside OpenAI-o1 and Gemini-Exp-1206 on the style control setting

• p. 43 Figure 11 screenshot

  ChatGPT-4o-latest (2024-11-20)

Judge's note: The Figure 11 screenshot on p.43 shows four models with Rank 1: o1-2024-12-17, Gemini-Exp-1206, ChatGPT-4o-latest (2024-11-20) and DeepSeek-R1. The text names only two co-leaders.

Second check: another model agreed

• p. 47 D.3.2 text

  R1 exhibits a significant performance gap compared to other models on the HarmBench benchmark

• p. 48 Table 9, HarmBench*

  89.3 (35.0)

Only the authors can say what is right here.

Judge's note: Checked Table 9 on the p.48 image. The gap holds only for the without-risk-control figure in parentheses (35.0 vs 67.0/58.0), and that figure is reported only for the DeepSeek models. The text does not say which comparison it means.

Second check: another model agreed

• p. 51 Table 10 and bullet 'Analyzing risk types'

  DeepSeek-R1 performs exceptionally well in handling queries related to Illegal and Criminal Behavior

• p. 51 Table 10, DeepSeek-R1 row

  DeepSeek-R1 19.7 3.8 28.9 8.6 32.5 6.0 16.9

Only the authors can say what is right here.

Judge's note: Table 10 on p.51 gives DeepSeek-R1 unsafe rates of Discrimination 19.7, Illegal 28.9, Harmful 32.5 and Ethical 16.9. The 28.9 for Illegal is higher than every other model in that column (14.1, 12.3, 22.5, 14.5, 17.3). With the risk control system the R1 row is 9.1 / 6.6 / 13.0 / 6.9, and there Illegal and Ethical are its best categories, which fits the text. The bullet does not name the configuration, so a reader looking at the base R1 row would find the claim contradicted. Only the authors can say which configuration they meant.

Second check: another model agreed

• p. 59 E.4 text

  using fewer than 7,000 thinking tokens for simple problems while dedicating more than 18,000 thinking tokens to the most challenging ones

• p. 59 Figure 18

  Avg. Thinking Tokens

• p. 59 E.4 text

  using fewer than 7,000 thinking tokens for simple problems

• p. 59 E.4 text

  more than 18,000 thinking tokens to the most challenging ones

Only the authors can say what is right here.

Judge's note: On p.59 the Figure 18 average-token curve starts at about 8,400 at the easy end and peaks at about 16,200. Only the shaded std-deviation band goes below 7,000 or reaches the 18,000 axis limit. The caption says the curve is smoothed with UnivariateSpline, so unsmoothed per-problem values might reach these extremes. However, the text presents Figure 18 as the evidence and the figure does not show these numbers. The 8,793 overall average could be explained by more problems falling in the easy bins, so I do not count it as a defect. Only the authors can say what the 7,000 and 18,000 figures measure.

Second check: another model agreed

• p. 60 E.5 text

  each stage brings significant improvement on complex coding reasoning problems

• p. 60 Table 14, Medium row

  81.76 81.42 83.45

Judge's note: Table 14 on p.60: Hard rises at every stage (17.09, 23.21, 30.36, 33.16, 34.44). Medium falls slightly from Dev2 to Dev3 (-0.34), and Easy is flat at 100.00 from Dev2 onwards. If 'complex' means the Hard row, the direction holds, but gains such as +1.28 from Dev3 to R1 are called 'significant' with no test shown for this comparison. This is a minor overstatement.

Second check: another model agreed

• p. 61 Section F text

  model performance improves progressively as the parameter size of the student model increases

• p. 61 Table 15, AIME cons@64

  80.0

• p. 61 Table 15

  DeepSeek-R1-Distill-Qwen-7B 55.5 83.3 ... DeepSeek-R1-Distill-Qwen-14B 69.7 80.0

• p. 61 Table 15, AIME cons@64

  DeepSeek-R1-Distill-Qwen-7B 55.5 83.3

Judge's note: Table 15 on p.61: pass@1 columns mostly rise with size within the Qwen family. But cons@64 drops from 83.3 (7B) to 80.0 (14B), Llama-8B is below Qwen-7B on AIME pass@1 (50.4 vs 55.5), and Llama-70B is below Qwen-32B on AIME pass@1 (70.0 vs 72.6) and on CodeForces rating. The trend holds in general, but the sentence states it without qualification.

Second check: another model agreed

• p. 69 Table 21

  output the answer in the format of "The answer is (X)" at the end

• p. 69 Table 21, Evaluation

  Parse the capital letter following “Answer: ” in response

Only the authors can say what is right here.

Judge's note: Page image p.69: the prompt asks for the answer in the format "The answer is (X)", while the evaluation parses the letter after 'Answer: '. The requested output format does not contain 'Answer: ', so the parser as described would not match it. The real parser may accept both formats; only the authors can clarify.

Second check: another model agreed

• p. 73 Table 25, PROMPT

  To write an SQL query that selects the top 10 rows in a database and joins to 3 different tables

Only the authors can say what is right here.

Judge's note: Page 73 image: the PROMPT row gives a full SQL solution and an explanation, so it is a response. The Evaluation row refers to 'the user prompt displayed below' and to assistant A/B answers. The example therefore shows an answer under the PROMPT label. Only the authors can say whether the user question or an example answer was intended.

Second check: another model agreed

• p. 4 Section 2.2, after Eq. 4

  The accuracy, reward and format reward are combined with the same weight.

• p. 2 Introduction, para 2

  The reward signal is solely based on the correctness of final predictions against ground-truth answers

Judge's note: The page image confirms the comma. Eq. 4 sums only two rewards, so the comma is a typo. One reviewer also noted that the Introduction (p.2) says the reward is "solely based on the correctness of final predictions" while Eq. 4 adds a format reward. That point is a minor wording looseness and is noted here as secondary.

Second check: none (judge only)

• p. 85 References

  OpenReview.net, 2023b. URL https://openreview.net/forum?id=1PL1NIMMrw.

• p. 4 Section 2.3

  self-consistency decoding (Wang et al., 2023c)

• p. 64 H.1

  Wang et al. (2023b) introduced self-consistency

• p. 85 References

  OpenReview.net, 2023b. URL https://openreview.net/forum?id=1PL1NIMMrw

Judge's note: The page 85 image shows two identical entries: same authors, title, venue and URL, differing only in the 2023b/2023c suffix. The text layer shows 2023c cited at line 223 (Section 2.3) and 2023b at lines 4284 and 4304. This is a redundant duplicate entry; the citations still lead to the correct work.

Second check: none (judge only)

• p. 11 Section 7 Author List

  individuals listed alphabetically by their first name within each category

• p. 12 Contributors

  Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng

Judge's note: Checked p.11 and p.12. There are further breaks in the order (for example 'Lei Xu' after 'Liyue Zhang').

Second check: none (judge only)

• p. 12 Author List, Contributors

  Shanhuang Chen, Shengfeng Ye, Shiyu Wang

• p. 12 Author List, Contributors

  Shaoqing Wu, Shengfeng Ye, Tao Yun

Judge's note: Both occurrences are visible on the full image of p.12. The second is also out of alphabetical position, which points to a duplication error.

Second check: none (judge only)

• p. 22 Listing 2

  which are large enough to to make incorrect codes exceed the time limit

• p. 22 Listing 2 prompt

  which are large enough to to make incorrect codes exceed the time limit.

Judge's note: Confirmed on the crop. It sits inside a quoted prompt that may be reproduced verbatim, so it is left as an observation.

Second check: none (judge only)

• p. 22

  nputs, which are large enough to to make incorrect code

Judge's note: The crop shows "to to" in the grey prompt box. The text is a quoted prompt template, so it may reproduce the prompt exactly as used. That is why this is an observation and not a correction.

Second check: code check

• p. 35 B.4.2

  For code-start SFT and the second-stage SFT

Judge's note: The crop shows "code-start". Everywhere else the paper calls this stage cold start.

Second check: none (judge only)

• p. 40 D.1 Baselines

  Sepcifically, we use

• p. 40 D.1 Baselines, last paragraph

  Sepcifically, we use 𝑘= 64

• p. 50 D.3.3

  In Tables 10, we consider two metrics

Judge's note: The crop shows the misspelling. One reviewer also raised "In Tables 10"; that point is handled under C028.

Second check: none (judge only)

• p. 44 D.3 intro

  on 6 publicly safety benchmarks

Judge's note: The crop confirms the wording.

Second check: none (judge only)

• p. 50 D.3.3, last paragraph

  In Tables 10, we consider two metrics

Judge's note: The p.50 image confirms the plural for a single table.

Second check: none (judge only)

• p. 51 D.3.3 bullet 1

  unsafe rate aound 5%

• p. 51

  (unsafe rate aound 5%)

Judge's note: The crop confirms the misspelling; the other tiers in the same bullet use "around".

Second check: none (judge only)

• p. 61 Heading F.1

  F.1. Distillation v.s. Reinforcement Learning

Judge's note: Page image p.61 shows the heading exactly as quoted. This is a cosmetic point only.

Second check: none (judge only)

• p. 63 Appendix G.1, 'Iterative pipeline'

  optimal reasoning trajecto-ries for tasks capabilities that cannot be fully realized

Judge's note: The crop of p.63 (Iterative pipeline paragraph) shows the same text, so this is not an extraction artefact. It is likely an editing leftover ('for tasks' or 'capabilities').

Second check: none (judge only)

• p. 84 References

  C. Snell, J. Lee, K. Xu, and A. Kumar. Scaling llm test-time compute optimally

• p. 84 References

  C. V. Snell, J. Lee, K. Xu, and A. Kumar. Scaling LLM test-time compute optimally

• p. 63 Appendix G.2

  (Snell et al., 2024)

• p. 64 Appendix H.2

  (Snell et al., 2025)

Judge's note: Page 84 image shows both entries. They are the arXiv and ICLR versions, with slightly different titles and different author initials. The text cites 2024 in G.2 (main.txt line 4248) and 2025 in H.2 (line 4295). Separate entries for preprint and published versions can be legitimate, so this is a minor consistency point. The Wang 2023b/2023c duplicate raised in reviewer notes 2 and 3 is judged under C056.

Second check: none (judge only)

• p. 68 Table 20, Constraints

  −1 ⩽ K < N ⩽ 100

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: that the original LiveCodeBench/AtCoder problem has lower bounds of 1. Page image p.68 shows a typeset minus sign, so this is not an extraction artefact. Within the paper, a negative K is meaningless for 'take K cards from the bottom', and both samples use positive values. This looks like a typesetting slip in an example prompt.

Second check: none (judge only)

• p. 76 Table 28 caption

  All question in C-Eval are Chinese.

Judge's note: Page 76 crop shows the caption text exactly as quoted. This is a singular/plural agreement error.

Second check: none (judge only)

• p. 78 Table 32, Evaluation

  to parse expressions

Judge's note: The full page 78 image has no footnote area; only the page number '78' follows the table. The text layer also has no footnote 1. The intended content is probably the sympy URL given in the methods (main.txt line 1238).

Second check: none (judge only)

• p. 83 References

  Introducing SWE-bench verified we’re releasing a human-validated subset of swe- bench that more, 2024b.

Judge's note: Page 83 image shows the entry exactly as quoted. The title runs into a scraped page-description fragment ending in 'that more', with no punctuation after 'verified'.

Second check: none (judge only)