{"id":"4e8558ee-bca1-4d69-a852-93bb698cd7e6","arxiv_id":"2506.02338","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A pipeline converts short-CoT LLM outputs into o1-style long chain-of-thought rationales using 1K seed reasoning flows, and SFT on the resulting dataset improves downstream RLVR cold-start.","lead":"The authors create a 100K long chain-of-thought dataset using ordinary short-reasoning LLMs guided by just 1K examples of OpenAI o1's reasoning outlines. Models fine-tuned on this data show larger reinforcement-learning gains than uninitialized base models, offering a cheaper path to open-source reasoning models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2-3x RLVR gain claim is confounded by a length reward that the LC initialization is explicitly trained to satisfy; removing it is needed before the central claim is supported.","rationale":"The abstract's central quantitative claim is the 2-3x larger RLVR gain. That claim is supported only by a comparison between the base model and the LC-initialized model under a reward mixture that includes a length reward matching o1-mini thought lengths. The dataset construction pipeline explicitly matches thought budgets to o1-mini, so the LC model is tailor-made to score well on that length reward, while the base model is not. Under GRPO, this changes the effective optimization objective for the two models, so the comparison does not cleanly isolate cold-start quality. The reader's rationale does mention reward design as a confound, but the reader's weakest-assumption slot is assigned to outline transfer; I view the reward confound as the more decisive risk to the central claim. If the length-reward ablation in the concrete test shows the gap collapses, the headline result would need to be reframed as demonstrating compatibility with a particular reward design, not a general foundation for RL. If the gap persists, the central claim is substantially strengthened. The dataset may still be valuable, and the issue is addressable with an ablation, so I do not see grounds to change the reader's conditional verdict.","tokens_in":16263,"tokens_out":6055,"duration_ms":55722,"concrete_test":"Run the Section 5.3 RLVR experiment on Qwen-2.5-0.5B and Qwen-2.5-0.5B-LC with the length reward removed, keeping only Answer Reward and Format Reward and holding all hyperparameters fixed. If the 2-3x relative gain on MATH500 and GPQA persists, the length reward is not driving the result; if it shrinks or vanishes, the headline claim is an artifact of matching the length reward. As a further control, run the same ablated-reward RL on a 0.5B model SFT on R1-distilled long-CoT data to test whether the advantage is specific to the Long CoT Collection or generic to any long-CoT initialization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive weakness is in the RLVR comparison (Section 5.3, Figure 1). The reward is answer + format + length, where the length reward is 1 - |min(x,y)/max(x,y) - 1| with y being o1-mini thought length. The LC model was SFT on rationales generated with thought budgets matched to o1-mini (Sections 3.2.1 and 6.1), so it starts with thought lengths near y. The base Qwen-2.5-0.5B starts with short thoughts, so its length reward is low and varies with output length. Under GRPO group-relative advantages, the base model receives a strong gradient to lengthen outputs rather than to solve, whereas the LC model receives a flatter length signal and a clearer answer signal. Thus, the reported 2-3x larger gains may measure reward-shaping compatibility rather than a better cold-start foundation. The paper does not include an RL baseline initialized on R1-distilled or any other long-CoT SFT data, so the claim that the Long CoT Collection is a 'reliable foundation' for RL is not yet controlled. This concern is more load-bearing than the outline-transfer question because even imperfect transfer of o1's reasoning strategies could still yield a useful RL initialization, whereas a reward confound directly invalidates the paper's primary quantitative result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Long CoT Collection, a 100K-instance dataset of long chain-of-thought (CoT) rationales generated by short-CoT LLMs (GPT-4o) under the guidance of a 1K seed dataset of o1 reasoning flows and thought budgets. The authors propose a three-stage pipeline (retrieval, reasoning-flow generation, step-by-step rationale generation) that transfers o1's reasoning structure to short-CoT LLMs while allowing control over thought length. They claim the resulting dataset is comparable in quality to DeepSeek-R1, improves SFT performance on GPQA and MMLU-Pro, and acts as a reliable foundation for reinforcement learning, with models initialized on it achieving 2-3x larger RLVR gains than base-model initialization. The paper publicly releases code, data, and models.","tokens_in":16528,"tokens_out":8442,"duration_ms":72984,"significance":"If the claims hold, the work provides a method to create long-CoT reasoning data without direct distillation from existing LRMs, which is a meaningful step toward independent open-source LRM development. The thought-budget controllability is a genuine and useful contribution, and the detailed pipeline description, public artifact release, and BoN/SFT results are strengths. However, the headline RLVR claim is currently confounded by the length-reward design, and the quality comparison against R1 is biased by question selection. The paper itself concedes in Section 7 that SFT on the dataset is not dramatically better than direct distillation, so the RL-initialization claim is the central differentiator and must be supported by properly controlled experiments. With additional controls, the dataset could be a valuable community resource.","major_comments":[{"comment":"The headline RLVR comparison is confounded by the length reward. The length reward is defined as 1 - |min(x,y)/max(x,y)-1| with y being o1-mini's thought length, while the Long CoT Collection is constructed using thought budgets matched to the same reference (Sections 3.2.1 and 6.1). Consequently, the LC-initialized policy already starts with outputs near the length target, whereas the base Qwen-2.5-0.5B does not. Under GRPO's group-relative advantages, the base model's early learning signal is dominated by length matching rather than by solving the problem, so the reported 2-3x larger gains likely measure reward-shaping compatibility rather than a better cold-start foundation. The authors should add RLVR runs with answer+format rewards only (no length reward) and, ideally, a control initialization on R1-distilled or other long-CoT SFT data. Without these controls, the central claim that the collection is a 'reliable foundation for RL' is not supported.","section":"Section 5.3, Figure 1"},{"comment":"The quality comparison against R1 is based on 100 questions selected 'for which R1-generated solutions have the correct answer.' This conditioning removes all cases where R1 is incorrect, biasing the comparison in R1's favor and preventing any conclusion about the dataset's quality on the underlying distribution. To support the claim that the Long CoT Collection is 'comparable to—or slightly below—R1,' the evaluation should be run on a random sample or stratified by R1 correctness, with results reported separately for cases where R1 is correct and incorrect.","section":"Section 4.1, Figure 4"},{"comment":"Table 3 reports that self-verification triggers appear far less frequently in the generated rationales than in R1: 'Wait' appears in only 4% of Ours versus 100% of DeepSeek-R1, and 'Verif-' in 27% versus 60%. This is hard to reconcile with the paper's claim that the dataset contains 'rich reasoning triggers' that 'help explore diverse reasoning paths.' The o3-mini-based qualitative evaluation may reward outline-like structure rather than genuine reasoning behavior. A behavioral test, such as measuring the frequency of self-correction or the accuracy improvement on problems requiring verification, would provide more direct evidence that the reasoning-flow transfer preserves o1's strategies.","section":"Section 4.1, Table 3"},{"comment":"The RLVR experiment is conducted only on Qwen-2.5-0.5B, while the SFT and BoN experiments use 7B-8B models. The abstract and Section 5.3 make an unqualified claim that 'models initialized on our data achieve 2-3x larger gains with RLVR,' which is not demonstrated at the model scale where the dataset is most likely to be deployed. While the limitation section acknowledges the GPU constraint, the claim should be qualified to the 0.5B setting or replicated at a larger scale before generalizing.","section":"Section 5.3 and Limitations"}],"minor_comments":[{"comment":"'dataset volumns' should be 'dataset volumes.'","section":"Section 2"},{"comment":"'the parable answer span' should be 'the parseable answer span' or 'the answer span.'","section":"Section 5.3"},{"comment":"The caption says 'Pearson correlation (R2)'—Pearson correlation is conventionally denoted r, while R2 is the coefficient of determination; please clarify which statistic is reported.","section":"Figure 7 caption"},{"comment":"The notation Dref ∈ {q, Sref, bref} is ambiguous; it should be Dref = {(q, Sref, bref)} or similar to indicate a set of tuples.","section":"Section 3.1.1"},{"comment":"The dataset is referred to as both 'Magpie-Reasoning-150K' and 'magpie-reasoning-V1'; use a single consistent name throughout.","section":"Appendix A.1"},{"comment":"'NunimaMath-CoT' should be 'NuminaMath-CoT.'","section":"Table 5 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper's core contribution is a dataset and a pipeline, and the SFT/BoN results are solid. The 2-3x RLVR claim is likely to be the main point of contention in review; the length-reward confound is real and should be addressed with additional experiments before publication. The authors should also consider whether the comparison against R1 in Section 4.1 meets the bar for a fair evaluation, as the current selection criterion is likely to draw criticism. This is a potentially useful resource, but the evidence for its unique value as an RL initialization is not yet convincing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nWhat you should know: this paper builds a 100K long-CoT dataset using GPT-4o, a short-CoT model, guided by just 1K o1 reasoning-flow outlines. The construction pipeline is the real contribution: retrieve a domain- and budget-matched outline from the seed set, generate a higher-level reasoning flow for the new question, then expand each step into prose. That combination is new, and the SFT results are credible—models trained on the collection beat their base versions on GPQA and MMLU-Pro by a few points, and the head-to-head quality comparison against R1 is not embarrassing. The thought-budget control (100/50/25%) is also a practical feature, with the expected accuracy trade-off. This is a resource the open-source reasoning community can build on.\n\nThe soft spot is the RLVR claim in Figure 1, and it is load-bearing. The 2-3x gain compares LC-initialized Qwen-2.5-0.5B against the base model under a reward that includes a length term defined as closeness to o1-mini's thought token count. The LC model was SFT on rationales whose budgets were matched to o1-mini, so it starts near the length target. The base model starts with short thoughts, so under GRPO's group-relative advantages it gets a strong gradient to lengthen outputs rather than to solve. The comparison mostly shows that LC SFT makes the model compatible with this particular reward, not that it is a better reasoning foundation. To support the \"reliable foundation for RL\" claim, they need an RL baseline initialized on R1-distilled or another long-CoT SFT dataset, or a reward ablation without the length term. The outline-transfer question—whether GPT-4o inherits o1's strategies or just the surface pattern—is real but secondary; even imperfect transfer can yield a useful initialization, and the SFT gains suggest something substantive is being learned.\n\nThe paper is honest in places: the conclusion admits SFT on their data does not dramatically beat direct distillation. The data/code release link is a placeholder, which matters for a dataset paper. The quality evaluation uses o3-mini as judge, a mild same-family concern but not damning.\n\nBottom line: this deserves a serious referee. The dataset and pipeline are worth engaging with; the RLVR claim needs either re-analysis or re-running with a proper control before it can be taken at face value. I would bring it to a reading group if we are discussing open reasoning datasets.\n\nRecommendation: send to peer review, with the RL confound as the major revision point.","headline":"A genuinely useful dataset and construction pipeline for distillation-free long CoT, but the headline RLVR gain is mostly a reward-compatibility artifact and needs a controlled comparison.","tokens_in":17094,"tokens_out":1718,"would_cite":true,"duration_ms":16022,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a 100K long chain-of-thought dataset can be generated by short-CoT LLMs, guided by 1K reasoning-flow outlines from o1, and that this dataset matches R1-level quality while serving as a cold-start foundation for…","keywords":["long chain-of-thought","reasoning flow","dataset construction","reinforcement learning with verifiable rewards","LLM reasoning","thought budget","overthinking","cold-start"],"falsifier":"Take the Long CoT Collection and shuffle the order of the outline steps before expansion, or replace the o1-extracted outlines with generic templates while keeping everything else identical; if the downstream SFT-then-RLVR gains on MATH500 and GPQA remain the same, the specific reasoning flow is not the causal ingredient, whereas if they drop sharply, the flow is doing the work.","tokens_in":16048,"feed_emoji":"🧠","tokens_out":5304,"duration_ms":42411,"temperature":0.7,"pith_summary":"The paper asks whether long chain-of-thought (CoT) reasoning data needed for training reasoning models can be produced by ordinary short-CoT LLMs, without distilling from an existing large reasoning model like R1. It answers yes: a 100K-instance dataset, the Long CoT Collection, is generated by GPT-4o guided by 1K reasoning-flow outlines harvested from o1. The dataset reaches quality judged comparable to R1's outputs and, when used to cold-start a small policy before reinforcement learning with verifiable rewards, yields 2-3x larger gains than starting from the base model. This matters because it decouples open-source reasoning-model development from dependence on closed long-CoT teachers.","feed_headline":"Short-CoT models mint a long-CoT dataset that triples RL gains","feed_subtitle":"Reasoning-flow outlines from o1 guide GPT-4o to produce 100K rationales, cold-starting RLVR.","key_machinery":"The reasoning flow, a sequence of outline phrases such as 'Analyzing the problem statement' and 'Okay, let's verify', summarizes the high-level reasoning strategy. The pipeline extracts 1K such flows with thought budgets from o1, retrieves domain- and budget-matched demonstrations for each new question, has GPT-4o generate a flow by imitating the retrieved structure, expands each flow step into an actual long CoT rationale conditioned on previous and next steps, and filters out wrong-answer rationales. The flow carries the teacher's verification and exploration strategies, and the number of outline steps provides the thought-budget control knob.","core_discovery":"The central claim is that o1's reasoning competence can be captured at the level of a 'reasoning flow'—a compact sequence of outline steps describing how a problem is decomposed and checked—and that this outline, plus a thought-budget estimate, is enough to make a short-CoT LLM (GPT-4o) produce long CoT rationales that rival R1's in quality. The resulting Long CoT Collection is not only useful for supervised fine-tuning; it is a reliable cold-start foundation for RLVR, with models initialized on it showing 2-3x greater performance gains on MATH500 and GPQA than base-model-initialized policies. The paper also demonstrates thought-budget controllability, letting the data generator adjust how much thinking is spent per problem, which addresses LRM overthinking.","pith_inferences":["If reasoning-flow outlines can be obtained cheaply from any capable model or even from text heuristics, the pipeline could be applied to domains beyond math and general reasoning, such as coding or scientific QA, without needing a long-CoT teacher.","The thought-budget mechanism suggests a potential curriculum: train first on high-budget data, then distill or compress to low budget, which might yield models that think just long enough and mitigate overthinking.","A direct test of the paper's causal story would be to shuffle the outline order or replace o1 flows with generic problem-solving templates; if downstream RL gains persist, then the specific o1 strategies are not the active ingredient.","Because the dataset is generated by GPT-4o, its ceiling is bounded by GPT-4o's base competence; the method's value likely grows as the generator model's base reasoning improves."],"forward_implications":["Long CoT datasets can be built by short-CoT LLMs, so future open reasoning models need not distill from R1 or other closed long-CoT systems.","Thought budget can be set at data-generation time; training on 100%-budget data beats 50% and 25% budgets on MATH500, suggesting the budget knob trades compute against reasoning quality.","Cold-starting RLVR with the Long CoT Collection gives 2-3x larger performance gains than starting from the base model, reducing the sparse-reward cold-start problem.","The collection transfers to general reasoning benchmarks (GPQA, MMLU-Pro), not just math.","Models trained on the collection show higher length, answer, and format rewards than the base in a LIMO-based reward analysis, indicating a better starting policy for RL."],"supporting_citations":[{"why":"introduced R1 and the cold-start SFT-before-RL recipe that this paper targets","marker":"(DeepSeek-AI et al., 2025)"},{"why":"analyzed the role of RLVR after SFT and motivated the reward functions used in the RL phase","marker":"(Yeo et al., 2025)"},{"why":"supplied the Magpie-Reasoning-V1 questions and domain annotations underlying the 1K seed and 100K collection","marker":"(Xu et al., 2024)"},{"why":"provided o1 as the teacher model whose reasoning flows and thought budgets are extracted","marker":"(OpenAI, 2024)"},{"why":"provided GRPO, the RLVR algorithm used for the downstream reinforcement learning","marker":"(Shao et al., 2024)"},{"why":"Bespoke-7B serves as an open-data direct-distillation baseline that the paper's models are compared against","marker":"(Labs, 2025)"},{"why":"documented underthinking in o1-like LLMs, motivating the thought-budget control analysis","marker":"(Wang et al., 2025)"},{"why":"provided the LIMO dataset used in the reward comparison between the initialized model and baselines","marker":"(Ye et al., 2025)"}],"fun_headline_variants":["Short-CoT LLMs mint long-CoT dataset","Reasoning-flow outlines drive long-CoT data","Cold-start RL with synthetic long CoTs","Long CoT Collection: 100K rationales for RL","Triple RL gains with generated long-CoT data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a compact reasoning-flow outline plus a thought-budget number captures enough of o1's reasoning competence that GPT-4o's step-by-step expansion preserves the real strategies rather than producing plausible filler; if the outline loses the substance of o1's reasoning, the dataset inherits only surface patterns and the RL gains would not come from learned reasoning strategies.","fun_headline_variants_meta":{"raw":{"variants":["Short-CoT LLMs mint long-CoT dataset","Reasoning-flow outlines drive long-CoT data","Cold-start RL with synthetic long CoTs","Long CoT Collection: 100K rationales for RL","Triple RL gains with generated long-CoT data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00027,"raw_usage":{"total_tokens":1624,"prompt_tokens":944,"completion_tokens":680,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":619}},"tokens_in":560,"tokens_out":680,"duration_ms":5911,"temperature":1.0,"reasoning_tokens":619,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:26:29.042429+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the Long CoT Collection and shuffle the order of the outline steps before expansion, or replace the o1-extracted outlines with generic templates while keeping everything else identical; if the downstream SFT-then-RLVR gains on MATH500 and GPQA remain the same, the specific reasoning flow is not the causal ingredient, whereas if they drop sharply, the flow is doing the work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provided o1 as the teacher model whose reasoning flows and thought budgets are extracted"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Bespoke-7B serves as an open-data direct-distillation baseline that the paper's models are compared against"}],"review_version":1}