acceptodds
Under review as a conference paper at ICLR 2027

Davinci-Math: A Routing-Aware Data Pipeline for End-to-End Mathematical Reasoning

Abstract

Training a strong mathematical reasoning model from a base model requires three stages—mid-training, SFT, and RL—with fundamentally different data needs. Most prior work curates data one stage at a time, so locally optimal choices fail to compose end-to-end; we instead take a joint multi-stage view, placing every problem where it is most valuable to the final model. We instantiate this view with daVinci-Math, a unified pipeline built on two pillars: (i) single-source preparation—cleaning, deduplication, and filtering shared across all three stages are applied once to a corpus aggregated from public sources, producing a 2.7M-problem pool; and (ii) stage-aware routing—each problem is labeled via multi-criteria classification and answer verification, then routed to mid-training, post-training, or filtered out, with post-training further split into SFT and a rule-verifiable RL subset. The pipeline yields 62B mid-training tokens, 3.8M SFT trajectories, and 39K RL prompts. We instantiate the full pipeline at both the 3B and 30B-A3B scales: daVinci-Math-3B reaches 86.6 on AIME 2025 and 75.1 on HMMT Feb 2025, and daVinci-Math-30B-A3B reaches 90.4 average over four benchmarks, setting a new state of the art at each scale. Controlled head-to-head comparisons further show that daVinci-Math provides stronger training data than open math datasets at both the mid-training and SFT stages, and ablation studies confirm the effectiveness of our pipeline design.

Primary area:
datasets and benchmarks
Keywords:
LLM reasoning, mathematical dataset

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 5 reviews, above 27% of the others.
Reviewer_UMV6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DaVinci-Math, a pipeline that curates math problem-solving data jointly for midtraining, SFT and RL. Problems from 28 public sources are cleaned and deduplicated, while being filtered to prevent benchmark contamination. Each problem is then labelled by rule-based filters and an LLM classifier and routed to midtraining, post-training or to be dropped. The pipeline yields 62B midtraining tokens, 3.8M SFT trajectories and 39K RL prompts, which are used to train several models. The experiments show a good improvement on a small and weak base model that outperforms newer reasoners, while also achieving near-saturation of the benchmarks on a Qwen3-30B-based model. The authors have also performed numerous ablation studies to show the necessity of the entire pipeline and isolating for confounding variable.s

Critical strengths
  1. The pipeline's motivation makes sense, and seems carefully developed, with diverse data sources, care against data contamination, deduplication, and selection methods. I also appreciate the paper's openness by providing the dataset during submission time. The paper is also clear and well-written.

  2. The ablation studies are well designed, controlling for most variables in question, and motivating the need for all their design choices, including multiple seeds and confidence intervals.

Critical weaknesses
  1. The paper presents no comparison to similar multi-stage training pipelines, which begs the question of whether a novel one is truly required. For example, Olmo 3 [1] presents an open data source that they use for their models, which could be used in-replacement of the one presented here (given the teacher for SFT is held the same), while DeReason [2] proposes a method to split data between SFT and RL stages (this one is not open from what I could understand but the method seems simple enough to implement). Some comparison would significantly strengthen the claims of the paper.

  2. The contribution of distillation versus routing is unclear. From what I could ascertain, other models do not clearly describe their teachers, which might end up weaker than OSS-120B. This leaves open the question of whether the stronger teacher is not what drives the majority of performance (which is why it seems the authors picked a stronger one for the larger model).

References

[1] Olmo, Team, et al. "Olmo 3." arXiv preprint arXiv:2512.13961 (2025).

[2] Hu, Hanxu, et al. "DeReason: A Difficulty-Aware Curriculum Improves Decoupled SFT-then-RL Training for General Reasoning." arXiv preprint arXiv:2603.11193 (2026).

Questions for authors
  1. (Future work?) Several works [3,4,5] highlight the importance of proof correctness contrasted to final-answer correctness. Given the pipeline focuses on verifiable objectives via filtering for SFT+GRPO rewards, did this impact the final models in terms of producing better answers but not necessarily good and correct proofs?

  2. Can you compare the model with its associated teachers on the benchmarks? Given the importance of distillation during the SFT stage it is important presentation-wise to show the capabilities of the teachers themselves.

  3. Please address the remaining concerns from the Weaknesses section.

References

[3] Petrov, Ivo, et al. "Proof or bluff? evaluating llms on 2025 usa math olympiad." arXiv preprint arXiv:2503.21934 (2025).

[4] Mahdavi, Hamed, et al. "Brains vs. bytes: Evaluating llm proficiency in olympiad mathematics." arXiv preprint arXiv:2504.01995 (2025).

[5] Zhang, Ziyin, et al. "Deeptheorem: Advancing llm reasoning for theorem proving through natural language and reinforcement learning." arXiv preprint arXiv:2505.23754 (2025).

Reviewer_aj8TOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DAVINCI-MATH, a unified data pipeline that routes mathematical problems to mid-training, SFT, and RL through filtering, classification, and answer verification. The resulting dataset achieves strong mathematical reasoning performance at both 3B and 30B-A3B scales.

Critical strengths
  1. Substantial engineering effort. The paper integrates 28 public datasets and constructs a large-scale mathematical reasoning corpus.
  2. Strong empirical results. The resulting models achieve competitive performance across multiple mathematical reasoning benchmarks.
  3. Reasonable experimental coverage. The authors provide controlled data comparisons, routing ablations, and evaluations across different model scales.
Critical weaknesses
  1. Unclear motivation and insufficient novelty. The introduction assumes that prior work primarily optimizes individual training stages independently. However, existing studies have already investigated cross-stage data allocation and training interactions (Chen et al., 2025; Akter et al., 2026; Kang et al., 2026). More directly, Zhao et al. (2026) investigate SFT/RL data routing, while He et al. (2026) consider stage-specific data selection with feedback from RL. The paper does not clearly establish what remains unresolved or how its contribution differs from these studies.
  2. Limited methodological innovation. The proposed framework mainly combines existing techniques, including deduplication, LLM-based filtering, answer verification, and heuristic routing. It is unclear what constitutes genuine joint multi-stage optimization beyond integrating these components into one pipeline.
  3. Insufficient evidence for the central claim. The routing ablations exclude RL, and the experiments lack a direct comparison against independently curated stage-specific datasets under matched data and training budgets. Therefore, the reported improvements do not convincingly establish the benefits of joint routing.

Citations:

  1. Chen et al. (2025). Chen, Zui, et al. "Advancing mathematical reasoning in language models: The impact of problem-solving data, data synthesis methods, and training stages." International Conference on Learning Representations. Vol. 2025. 2025.
  2. Akter et al. (2026). Akter, Syeda Nahida, et al. "Front-loading reasoning: The synergy between pretraining and post-training data." International Conference on Learning Representations. Vol. 2026. 2026.
  3. Kang et al. (2026). Kang, Feiyang, et al. "Quagmires in sft-rl post-training: When high sft scores mislead and what to use instead." International Conference on Learning Representations. Vol. 2026. 2026.
  4. Zhao et al. (2026). Zhao, Yang, et al. "Consolidation or Adaptation? PRISM: Disentangling SFT and RL Data via Gradient Concentration." Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.
  5. He et al. (2026). He, Chongyang, et al. "Learning What to Learn: Stage-Specific Data Sets for SFT-then-RL in Small Language Model Reasoning." arXiv preprint arXiv:2606.04466 (2026).
Questions for authors

See weaknesses.

Reviewer_WhvyOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a unified data pipeline for mathematical reasoning. Problems collected from public datasets are cleaned and routed to mid-training, post-training, or removal. The post-training pool supports SFT, with a challenging, rule-verifiable subset used for RL. The authors generate reasoning trajectories, train models with the resulting datasets, and evaluate the pipeline through dataset comparisons and routing ablations.

Critical strengths

The work involves substantial data engineering and training effort. The resulting datasets and models have practical value, and the controlled comparisons under matched training budgets provide useful evidence beyond the final model benchmarks. Evaluating the data across multiple base models and training stages also strengthens the empirical contribution.

Critical weaknesses
  • Table 22 does not clearly isolate the benefit of cross-stage allocation: several ablations also change the SFT pool, so the differences may simply reflect the importance of suitable SFT demonstrations. Given established filtering practices in Qwen2.5-Math (https://arxiv.org/abs/2409.12122) and stage-dependent allocation studied in Front-Loading Reasoning (https://arxiv.org/abs/2510.03264), the additional methodological contribution remains insufficiently established.

  • For Qwen3-8B trained on daVinci-Math, Table 19’s four competition scores average 57.865, whereas Tables 24–25 report 65.21. Although Appendix E describes an independent estimate, the checkpoint or evaluation differences underlying these results need clarification.

Questions for authors
  • Can the authors isolate the benefit of cross-stage allocation while keeping SFT supervision unchanged? A direct comparison would use the same problem pool and teacher, fix SFT data and per-stage budgets, and vary whether mid-training uses the routed pool, a mixture, or SFT-eligible problems. What additional benefit does the proposed allocation provide under these controls?
  • What independent validation supports the reliability of the filtering, answer verification, and decontamination steps? In particular, is there systematic auditing beyond final-answer agreement to assess the quality of retained reasoning trajectories?
  • Do Tables 19 and 24–25 use different checkpoints, training runs, or evaluation settings? Please clarify their correspondence and the source of the numerical difference.
Reviewer_8iJuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Training a math-reasoning model involves three stages: mid-training, SFT and RL. Most prior data work targets one stage, e.g. collecting high-quality SFT traces or RL-suitable problems. The authors argue that data selection should be done across the whole pipeline, deciding for each problem which stage it is most useful for. They aggregate 28 public sources into 2.7M canonical problems and route them, producing 62B mid-training tokens, 3.8M SFT trajectories and 39K RL prompts. Models trained on this data reach 86.6 on AIME 2025 and 75.1 on HMMT Feb 2025 at 3B, and a four-benchmark average of 90.4 at 30B-A3B. The contribution is a data-allocation and routing strategy rather than an algorithm.

Critical strengths
  1. The engineering scale is substantial and the central idea, routing each problem to the stage where it is most useful instead of discarding unverifiable or multiple-choice items, is sensible. The seeded ablations (Table 22) do support the claim that routing helps at small scale: full routing 31.7 vs distribution-matched random 28.2.

  2. The reported results are strong in absolute terms, and the released multi-stage corpus would be useful to the community if the attribution questions below are resolved.

Critical weaknesses
  1. Attribution. It is hard to tell whether the gains come from the routing itself or from more data, stronger filtering and higher-quality teacher trajectories; compute budget, model initialization and training resources also differ across comparisons. In particular, DaVinci trajectories come from GPT-OSS-120B while OMR comes from R1, so matched tokens are not matched teachers, and the NemotronMath baseline is anomalously weak (almost no gain from 1.9B to 28B tokens), which suggests a baseline training problem rather than a data-quality difference. The RL stage, the most expensive part of the "end-to-end" pipeline, is never ablated against alternative RL prompt sets.

  2. The SOTA claim is statistically thin. The benchmarks have 30 questions and results use a single training seed; the authors' own bootstrap CI for the 3B gain is [0.18, 6.64], with HMMT Nov negative; the 30B-A3B margin of +1.07 has no CI and lies within the ±7 per-benchmark noise. For a work that relies on large-scale verified math data, contamination control and the reliability of answer verification also need to be reported much more explicitly.

Questions for authors
  1. Can you compare the 39K RL prompt set against existing RL sets (e.g. DAPO, DeepMath) under identical GRPO settings, so the RL stage of the pipeline is validated?
  2. Please explain the NemotronMath anomaly (essentially flat from 1.9B to 28B tokens), and report SFT comparisons with a matched teacher.
  3. Please report CIs for Table 3 (30B-A3B) and tone down the "new state of the art" language in the abstract accordingly.
  4. Please give per-benchmark decontamination drop rates and the verification protocol in enough detail to assess reliability.
  5. Were baseline numbers re-evaluated under your protocol (80K cap, avg@32) or copied from reports?

I would consider raising my score if (1)–(3) are addressed.

Reviewer_BCXdOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose DAVINCI-MATH, a unified multi-stage data pipeline. They highlight that DAVINCI-MATH is built by single-source preparation, so that data for different stages is processed in the same way; and by stage-aware routing, so the problem is assigned to one of-training, SFT and RL. The authors have clearly declared the dataset as their main contribution and proved their point through strong SOTA experimental results head-to-head.

Critical strengths
  1. The paper is well written and easy to read.
  2. The authors contribute a brand new dataset with a rational data processing pipeline carefully designed to unify training data across different stages.
  3. The authors provide detailed methodology of their data collection process.
  4. The authors provide a comprehensive experimental results I see no major flaw. The main result is strong enough to prove the statements of the paper.
Critical weaknesses
  1. One can argue that the methodology is not novel, but since the authors have emphasized the dataset as their major contribution, I think that's less of a problem.
  2. However, I do hope authors could justify whether their data collection process gets any rational signals from the training process? It could be vibed.
Questions for authors
  1. See weakness 2.
  2. What is the exact difference of the authors' proposed data processing framework, compared to a traditional standard one? I think the authors declare their advantage is unification, but why is unification good? Is there any experimental result that can directly support this claim?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.