{"id":"c0819d7d-d2de-4f0a-92a3-889dc203e624","arxiv_id":"2507.08267","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"Ten-epoch SFT then GRPO raises AIME/MATH-500 accuracy while reducing solution length, achieving 8th place on AIMO's private set.","lead":"An extended 10-epoch supervised fine-tuning stage followed by GRPO reinforcement learning improves both accuracy and token efficiency of DeepSeek-R1-Distill-Qwen models on math benchmarks. The recipe placed 8th among 2,212 teams on AIMO's private set, suggesting it is a practical and competitive training blueprint.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 10-epoch SFT claim is undercut by the paper's own 1.5B results and by the lack of a contamination check against the public AIME/MATH-500 test sets, so the central 'accuracy breakthrough' claim is not yet secure.","rationale":"The reader's weakest_assumption is that the evaluation numbers reflect genuine generalization rather than overlap between the public test sets and the curated SFT data, with no deduplication or contamination check reported. This is exactly the most load-bearing vulnerability of the paper's central claim. The paper's strongest evidence is the AIMO public/private ranking, which is genuinely leak-free external validation, and the 14B results do show a consistent accuracy improvement from +SFT(10 epochs) and a large token reduction from the subsequent GRPO phase. However, the central claim in the abstract is unqualified: 'extending SFT for as many as 10 epochs is crucial for performance breakthroughs' and 'the primary role of GRPO is to optimize solution length.' Both conditions need scrutiny. For the GRPO efficiency claim, the paper has direct support: across nearly all rows in Table 1, +SFT(10 epochs)+RL reduces token counts relative to +SFT(10 epochs) (e.g., 14B AIME2024: 10,268 to 7,932) while maintaining or slightly improving accuracy. That part of the claim is internally consistent, though it would be stronger with confidence intervals given only 64 samples and no variance reported. For the SFT-accuracy claim, the internal evidence is mixed: the 1.5B model shows no SFT accuracy gain on AIME (26.0 vs 27.8 on 2024; 22.1 vs 22.3 on 2025). The paper itself acknowledges this in Section 4.3, yet the abstract does not condition the 'crucial' claim on model size or data regime. The contamination risk is real because the SFT sources are public corpora (OpenR1 Math, open-r1-math-220k, Light-R1-SFT) that are known to be scraped from public benchmark and forum data, and the paper reports no overlap check with the exactly the benchmarks used for evaluation. If contamination exists, it would inflate exactly the SFT accuracy gains that the narrative depends on. I agree with the reader's assessment; the appropriate verdict remains CONDITIONAL, pending either the release of the promised code/data (which the paper says will be open-sourced but does not ship) or a contamination check and scale-conditioned restatement of the claims. The AIMO private-set result is genuinely strong evidence that the overall recipe works in a leak-free setting, but it does not by itself validate the internal attribution to 10-epoch SFT versus GRPO, nor the unqualified generalization across scales.","tokens_in":9230,"tokens_out":2666,"duration_ms":23167,"concrete_test":"Run a contamination check: for every problem in AIME 2024, AIME 2025, and MATH-500, compute the maximum token-overlap or normalized longest-common-subsequence similarity against each of the 7,900 SFT training problems (and against the Light-R1 GRPO training set), flagging any pair above a threshold (the standard is >= 13-token contiguous overlap or an embedding similarity of >0.9). If any AIME/MATH-500 problem (or a near-duplicate) is present in the SFT or GRPO training data, re-run Table 1 and Table 2 with those problems removed; if the +SFT(10 epochs) accuracy gains shrink substantially relative to the Original row, the central 'breakthrough' claim must be restated as conditional on excluding contaminated items.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that extended SFT (10 epochs) is 'crucial for performance breakthroughs' and that GRPO's primary role is token-efficiency improvement. Two conditions must hold: (a) the accuracy gains from +SFT(10 epochs) reflect genuine generalization rather than train/test overlap, and (b) the 'prolonged SFT is crucial' claim holds across the tested scales. Condition (a) is not established: Section 3.1 curates SFT data from OpenR1 Math and Light-R1 public corpora, and Section 4.1 uses AIME 2024/2025 and MATH-500 as test sets, but the paper reports no deduplication or contamination check (no n-gram overlap analysis, no exclusion of AIME/MATH-500 problems during dataset construction). If AIME 2024/2025 or MATH-500 problems (or near-duplicates) appear in the 7,900 SFT triplets, the reported SFT accuracy gains in Table 1 and Table 2 are inflated and the GRPO 'preserves accuracy while shortening length' narrative is weakened. Condition (b) is contradicted by the paper's own data: in Table 1, for the 1.5B model on AIME 2024, +SFT(10 epochs) actually decreases accuracy from 27.8 to 26.0, and on AIME 2025 it decreases from 22.3 to 22.1; the 'breakthrough' appears mainly at 14B. The paper acknowledges this (Section 4.3: 'for the 1.5B model, the 10-epoch SFT did not yield a substantial accuracy increase') but still frames '10 epochs is crucial' as a general recipe in the abstract without conditioning on model scale. The AIMO result (28/50 private, 8th place) is strong external evidence that the recipe can produce good competition performance, so the concern is not that the method fails outright, but that the paper's headline causal attribution to 10-epoch SFT is not supported for all scales and is vulnerable to contamination.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage post-training recipe for mathematical LLMs: an extended full-parameter SFT stage (10 epochs) on a curated dataset of 7,900 high-difficulty problem-solution-answer triplets, followed by a GRPO stage with a composite reward (format, cosine similarity against reference traces, and length penalty). The central claims are that prolonged SFT is crucial for accuracy breakthroughs and that, in this framework, GRPO mainly improves token efficiency while preserving the SFT peak accuracy. The authors evaluate on AIME 2024, AIME 2025, and MATH-500 across 1.5B, 7B, and 14B DeepSeek-R1-Distill-Qwen models, and report a private-set score of 28/50 (8th of 2,212 teams) on the AIMO 2 competition. The paper also reports a per-problem analysis and ablation of reward functions.","tokens_in":9690,"tokens_out":3492,"duration_ms":38419,"significance":"If the claims hold, the recipe is practically valuable: it decouples accuracy maximization (SFT) from inference-efficiency optimization (GRPO) and is validated in a competitive, genuinely leak-free setting. The strongest evidence is the AIMO private-set result, which is external and not subject to public-benchmark contamination concerns. The commitment to release code, checkpoints, and datasets also supports reproducibility. However, the paper's central '10 epochs is crucial' claim is not uniformly supported by its own tables, and the absence of a contamination check leaves the headline accuracy gains on public benchmarks uncertain. As submitted, the evidence supports a scale-dependent, more cautiously worded version of the recipe rather than the broad claim in the abstract.","major_comments":[{"comment":"The SFT data are curated from public corpora (OpenR1 Math, Light-R1-SFT Data) and evaluated on AIME 2024/2025 and MATH-500, but no deduplication, n-gram overlap analysis, or exclusion of test-set problems is reported. If any AIME/MATH-500 problems or near-duplicates are present in the 7,900 training triplets, the accuracy gains in Tables 1 and 2, and the subsequent claim that RL preserves accuracy, are inflated. The authors should add a contamination check (e.g., normalized n-gram overlap, embedding similarity, and manual review of top matches) and report results with matched items removed.","section":"§3.1, §4.1, Tables 1-2"},{"comment":"The abstract states that extending SFT to 10 epochs is 'crucial for performance breakthroughs' without conditioning on model scale. Table 1 shows the opposite for the 1.5B model: +SFT(10 epochs) decreases AIME 2024 accuracy from 27.8 to 26.0 and AIME 2025 accuracy from 22.3 to 22.1. The paper acknowledges this in §4.3 but does not reconcile it with the abstract-level claim. The claim should be restricted to 7B/14B or explicitly framed as scale-dependent.","section":"Abstract, §4.3, Table 1"},{"comment":"Pass@1 is reported as the mean over 64 sampling runs, but no standard deviations, confidence intervals, or bootstrap estimates are provided. AIME 2024 and AIME 2025 each contain only 30 problems, so differences of a few percentage points (e.g., 65.2 vs 63.3 for 14B on AIME 2024) are within sampling noise. This is load-bearing for several comparisons, including the claim that RL preserves accuracy after SFT. The authors should report CIs or bootstrap intervals for the headline numbers.","section":"§4.1, Tables 1-2"},{"comment":"The GRPO reward combines format, cosine-similarity, and length-penalty terms with multiple hand-set scaling choices (e.g., reward ranges 0.1-1.0 for correct and -1.0 to -0.1 for incorrect, a 30,000-token maximum, and a length penalty whose coefficient is not stated). The conclusion that GRPO's primary role is token-efficiency optimization could be an artifact of these particular weights. The authors should give the full reward formula and coefficients, and include at least a small sensitivity analysis over penalty strengths.","section":"§3.2, §4.6, Figure 2"}],"minor_comments":[{"comment":"The sentence 'To test this, For this analysis, we evaluated three distinct reward configurations' contains a duplicated phrase and should be rewritten.","section":"§4.6"},{"comment":"The format-reward regular expression is printed as r\"^.*?oxed\\{(.*?)\\}.*?</think>.*?$\", which appears to be missing the backslash before 'oxed' and should read r\"^.*?\\boxed\\{(.*?)\\}.*?</think>.*?$\"; as typeset it is not a valid way to match \\boxed.","section":"§3.2"},{"comment":"The evaluation protocol says pass@1 is averaged over 64 sampling runs but does not report sampling temperature or top-p; these are needed for reproducibility and for interpreting the token-length comparisons.","section":"§4.1"},{"comment":"The phrase 'consistently and significantly boosts' in the introduction is stronger than the data support, especially given the 1.5B results and the lack of statistical significance testing; 'significantly' should either be removed or backed by tests.","section":"§1, §5"},{"comment":"The Wei et al. reference contains a malformed author segment ('brian ichter, Xia, F.,') and should be corrected to the standard ACL anthology author list.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The AIMO private-set ranking is the most credible piece of evidence and should be foregrounded, but the manuscript currently overstates the 10-epoch SFT result. I would not reject on the contamination concern alone, since the AIMO result is external, but the public-benchmark claims need either a contamination analysis or a clear caveat. The paper is also quite thin methodologically for a journal submission; the authors should be asked to provide the missing statistical and reproducibility details."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWorth a look if you build math reasoners. The concrete contribution is a two-stage recipe — 10-epoch SFT on ~7,900 hard traces, then GRPO with cosine/length rewards — and it ships with a real external validation: 28/50 on AIMO private, 8th of 2,212 teams. That result is the strongest thing in the paper, because the private set is leak-controlled. The SFT-epoch ablation in Figure 1 is also a genuinely useful datapoint: token length balloons after one epoch while accuracy dips, and accuracy climbs with more epochs at 7B/14B. I also buy the division-of-labor claim at those scales: GRPO's main measured effect is shorter outputs, not higher accuracy.\n\nThe soft spots are real but not fatal. First, the headline '10 epochs is crucial' is contradicted by the paper's own 1.5B numbers: AIME 2024 actually drops from 27.8 to 26.0, and AIME 2025 from 22.3 to 22.1. The paper acknowledges this in Section 4.3 but still frames it as a general recipe in the abstract. Second, there is no contamination check: SFT data is curated from OpenR1 Math and Light-R1 public corpora, while AIME 2024/2025 and MATH-500 are public benchmarks. Without deduplication or n-gram overlap analysis, part of the accuracy gain could be memorization. That is a genuine gap, though the AIMO result provides some independent evidence the model generalizes. Third, there are no error bars on the AIME results and sampling temperature/top-p are unstated, which makes small deltas (e.g., 52.0 vs 51.0) hard to interpret. Also, they promise code, checkpoints, and data but only provide a GitHub URL, so reproducibility is currently a promise.\n\nOverall, this is an honest empirical study within an existing SFT+RL program, not a fundamental advance. It deserves a serious referee, but the claims need scaling and contamination caveats, and the artifact release should be required before publication is final.\n\nRecommendation: referee it, with requests for the artifacts and a softened 'accuracy breakthrough' claim.","headline":"Useful empirical recipe with a credible AIMO result, but the '10 epochs is crucial' claim is oversold at smaller scales and unguarded against contamination.","tokens_in":10267,"tokens_out":1832,"would_cite":true,"duration_ms":19598,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-stage training recipe—extended supervised fine-tuning followed by a short GRPO phase—produces mathematical LLMs that are both more accurate and more token-efficient, with SFT carrying the accuracy gains and reinforcement learning…","keywords":["mathematical reasoning","supervised fine-tuning","GRPO","token efficiency","reinforcement learning","AIMO","chain-of-thought","reward design"],"falsifier":"Search the curated 7,900-example SFT corpus, along with the OpenR1 and Light-R1 sources, for near-duplicates of AIME 2024/2025 and MATH-500 test problems; if a substantial fraction of test items appear with their solution traces, the reported accuracy gains from 10-epoch SFT would be explained by memorization. Re-running the recipe on a deduplicated corpus would settle whether the gains persist.","tokens_in":9024,"feed_emoji":"🧮","tokens_out":6501,"duration_ms":59843,"temperature":0.7,"pith_summary":"This paper proposes and defends a practical training recipe for mathematical reasoning LLMs: first fine-tune on high-quality solved problems for as many as 10 epochs, then apply a short GRPO reinforcement learning stage. The central claim is that these stages play complementary roles—the long SFT phase is what pushes accuracy upward, while GRPO's main contribution is cutting the length of generated solutions without giving back that accuracy. If correct, this separates accuracy and inference efficiency into two independently optimizable objectives rather than a trade-off. The authors support the recipe with results on AIME 2024/2025 and MATH-500 across three model scales, and with a top-8 private-set finish among 2,212 teams on the leak-free AIMO competition.","feed_headline":"10-epoch SFT lifts math accuracy; RL then cuts solution length","feed_subtitle":"A two-stage recipe—long supervised fine-tuning then GRPO—tops AIME and places 8th on the leak-free AIMO private set.","key_machinery":"The load-bearing mechanism is the two-stage recipe itself. Stage 1 is full-parameter SFT for 10 epochs on 7,900 high-difficulty problem-solution-answer triplets assembled from OpenR1 Math, the openr1 hard subset, and Light-R1-SFT data, selecting correct solutions with short token lengths and substituting ground-truth answers from R1 traces where missing. Stage 2 is GRPO for 50 steps on Light-R1 second-stage data, with a reward function combining a format regular-expression check, a cosine-similarity reward that scores generated traces against reference correct traces (scaled 0.1 to 1.0 for correct, -1.0 to -0.1 for incorrect), and a length penalty. The empirical pattern of accuracy rising across SFT epochs while GRPO compresses output length is what carries the argument that the two methods are complementary.","core_discovery":"The paper's central discovery is that accuracy and inference efficiency in math LLMs can be decoupled across two training stages. Starting from DeepSeek-R1-Distill-Qwen models, the authors find that supervised fine-tuning pushed to 10 epochs—longer than typical cold-start SFT—improves pass@1 on AIME and MATH-500, whereas a single epoch can even hurt accuracy while lengthening outputs. A subsequent short GRPO phase, using a reward that combines format adherence, cosine similarity to reference correct traces, and an explicit length penalty, preserves or slightly improves that accuracy while substantially reducing mean output tokens. The paper interprets this as SFT setting the performance ceiling and GRPO optimizing solution length, and reports that the recipe transfers across 1.5B, 7B, and 14B models, with the largest gains at 14B. The strongest validation claim is 29/50 on the AIMO public set and 28/50 on the private set, equivalent to 4th and 8th place among 2,212 teams.","pith_inferences":["If the decoupling is real, the same two-stage scheme—long supervised training for competence, then RL for brevity—may transfer to other reasoning domains such as code generation or theorem proving, where over-long outputs are also costly.","The cosine reward's dependence on reference correct traces from a strong teacher (DeepSeek-R1) means the recipe's efficiency gains may not reproduce when only weaker teachers are available; a testable extension is to run Stage 2 with a self-generated reference set.","Because the paper reports no contamination check, the AIME/MATH-500 gains are vulnerable to training/test overlap; comparing accuracy on a freshly written, non-public problem set would be a stronger test of the recipe than any existing benchmark.","The finding that one SFT epoch sharply drops accuracy while many epochs recover suggests a non-monotonic learning curve worth characterizing mechanistically, for instance through loss landscape or representation analysis across epochs."],"forward_implications":["Practitioners building specialized math models should expect peak accuracy only after many SFT epochs, contrary to the common one-to-three epoch cold-start practice.","GRPO can be deployed primarily as a token-compression stage rather than an accuracy booster, enabling models that are both accurate and cheap to run at inference.","The recipe's benefit grows with model scale: 14B models gain both accuracy and efficiency, while 1.5B models show little accuracy improvement, suggesting a capacity threshold for the effect.","The reward design matters: a cosine-similarity reward tied to reference traces plus an explicit length penalty beats a binary accuracy reward on the accuracy-efficiency frontier.","Success on the leak-free AIMO private set supports the claim that the recipe generalizes beyond public benchmarks, though the single competition result is a limited sample."],"supporting_citations":[{"why":"Supplies the DeepSeek-R1-Distill-Qwen base models and the binary accuracy reward baseline used in the ablation.","marker":"DeepSeek-AI, 2025"},{"why":"Introduces GRPO, the reinforcement learning algorithm used in Stage 2.","marker":"Shao et al., 2024"},{"why":"Provides the cosine similarity reward component that scores generated traces against reference correct traces.","marker":"Yeo et al., 2025"},{"why":"Provides the Light-R1-SFT Data used both as one SFT source and as the GRPO training set.","marker":"Wen et al., 2025"},{"why":"Open R1 is the source of the OpenR1 Math and openr1 hard subsets from which the SFT data is curated.","marker":"Face, 2025"},{"why":"Defines the AIMO competition and its leak-free private test set, the paper's headline validation benchmark.","marker":"Frieder et al., 2024"},{"why":"Introduces the MATH-500 benchmark used for a second evaluation axis.","marker":"Hendrycks et al., 2021"}],"fun_headline_variants":["SFT to the limit, RL for brevity: math LLM recipe","Two-stage recipe: long SFT peaks accuracy, RL trims length","10-epoch SFT then GRPO: top math accuracy with token efficiency","Complementary training: SFT boosts math scores, RL slashes tokens"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation numbers reflect genuine generalization rather than overlap between the public test sets and the SFT training data; the paper reports no deduplication or contamination check against AIME 2024/2025 and trusts the AIMO private set to be leak-free.","fun_headline_variants_meta":{"raw":{"variants":["SFT to the limit, RL for brevity: math LLM recipe","Two-stage recipe: long SFT peaks accuracy, RL trims length","10-epoch SFT then GRPO: top math accuracy with token efficiency","Complementary training: SFT boosts math scores, RL slashes tokens"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000666,"raw_usage":{"total_tokens":3083,"prompt_tokens":1035,"completion_tokens":2048,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":1967}},"tokens_in":651,"tokens_out":2048,"duration_ms":16760,"temperature":1.0,"reasoning_tokens":1967,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:22:44.797263+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Search the curated 7,900-example SFT corpus, along with the OpenR1 and Light-R1 sources, for near-duplicates of AIME 2024/2025 and MATH-500 test problems; if a substantial fraction of test items appear with their solution traces, the reported accuracy gains from 10-epoch SFT would be explained by memorization. Re-running the recipe on a deduplicated corpus would settle whether the gains persist.","supporting_citations":[{"cited_title":"Demystifying long chain-of-thought reasoning in LLM s","cited_arxiv_id":null,"evidence_quote":"Provides the cosine similarity reward component that scores generated traces against reference correct traces."},{"cited_title":"C., Buzzard, K., Gowers, T., Liu, P","cited_arxiv_id":null,"evidence_quote":"Defines the AIMO competition and its leak-free private test set, the paper's headline validation benchmark."},{"cited_title":"Measuring mathematical problem solving with the MATH dataset","cited_arxiv_id":null,"evidence_quote":"Introduces the MATH-500 benchmark used for a second evaluation axis."}],"review_version":1}