MiroMind-M1: Open Math Reasoning Models

• MiroMind-M1 is a series of open, reproducible mathematical reasoning models using a two-stage training pipeline that combines supervised fine-tuning and reinforcement learning with verifiable rewards.
• It employs Context-Aware Multi-Stage Policy Optimization (CAMPO) to optimize token efficiency, enforce adaptive repetition penalties, and support KL-free RL training.
• The release includes complete datasets, enhanced verifiers, and detailed training configurations, advancing openness, transparency, and replicability in mathematical reasoning research.

Searching arXiv for the exact topic to ground the article in current papers. Searching arXiv for "MiroMind-M1" and closely related papers on open mathematical reasoning models. MiroMind-M1 is a series of open, fully reproducible mathematical reasoning LLMs built on the Qwen-2.5 backbone and introduced to advance openness, transparency, and reproducibility in mathematical reasoning (Li et al., 19 Jul 2025). The series centers on a two-stage training pipeline: supervised fine-tuning on a carefully curated corpus of 719K math-reasoning problems with verified chain-of-thought trajectories, followed by reinforcement learning from verifiable rewards on 62K challenging and verifiable problems. Its principal algorithmic contribution is Context-Aware Multi-Stage Policy Optimization (CAMPO), which combines length-progressive training, an adaptive repetition penalty, and diversity filtering to support context-aware RL training without relying on KL loss in the 32B setting. The release includes models, datasets, code, training configurations, evaluation configurations, and verifier improvements, and reports state-of-the-art or competitive performance together with superior token efficiency among Qwen-2.5-based open-source 7B and 32B models on AIME24, AIME25, and MATH benchmarks (Li et al., 19 Jul 2025).

1. Position in open mathematical reasoning

MiroMind-M1 was motivated by a specific deficiency in prior open-source reasoning LLM work: many projects released checkpoints but omitted critical resources such as full datasets, verifiers, and detailed training configurations, making high-quality replication difficult (Li et al., 19 Jul 2025). The series therefore defines its scope through three stated goals. “Openness” refers to release of models, the complete training corpora for SFT and RL, code, and verifier improvements. “Transparency” refers to detailed documentation of data curation, decontamination, and training settings. “Reproducibility” refers to exact hyperparameters, compute settings, and evaluation protocols.

The work treats mathematical reasoning as a preferred domain for RL because it offers precise, verifiable signals for reinforcement learning from verifiable rewards. This choice is methodological rather than merely benchmark-driven: correctness can be checked automatically, enabling stable reward construction and rigorous decontamination against held-out evaluation sets. Within that framing, MiroMind-M1 is presented as a fully open pipeline rather than only a model family.

2. Model series, backbone, and released artifacts

The released series comprises one supervised model and two RLVR models, all Qwen-2.5-based (Li et al., 19 Jul 2025).

Variant	Initialization	Role
MiroMind-M1-SFT-7B	Qwen-2.5-Math-7B	SFT model
MiroMind-M1-RL-7B	MiroMind-M1-SFT-7B	RLVR model
MiroMind-M1-RL-32B	DeepSeek-R1-Distill-Qwen-32B	RLVR model

For SFT, the authors “increase model’s max_position_embeddings to 32,768 {using Linear RoPE scaling}” and “use a cutoff length of 26,000.” For evaluation, the maximum generation length is 32,768 tokens. These long-context settings are integral to the project’s treatment of long reasoning traces rather than a secondary engineering detail.

The released resources include the three model checkpoints; the two datasets MiroMind-M1-SFT-719K and MiroMind-M1-RL-62K; full training and evaluation configurations; the CAMPO implementation; and an improved verifier. Public endpoints listed in the release materials include the GitHub repository at https://github.com/MiroMindAsia/MiroMind-M1, a model endpoint exemplified by https://huggingface.co/miromind-ai/MiroMind-M1-RL-7B, a dataset endpoint exemplified by https://huggingface.co/datasets/miromind-ai/MiroMind-M1-RL-62K, and the project site https://miromind.ai/.

3. Supervised fine-tuning corpus and data curation

The SFT stage uses a math-only corpus with verified chain-of-thought trajectories, assembled entirely from open sources and decontaminated against AIME24, AIME25, and MATH500 (Li et al., 19 Jul 2025). The final dataset is described as 719k traces over 412k questions, with source composition as follows: OpenR1 contributes 191k questions and 418k traces used; OpenThoughts contributes 56k math pairs retained; Light-R1 contributes 75k questions and 76k traces used; Synthetic-1 contributes 362k questions and 638k traces, from which a math-focused subset of 247k is used after removing proof and non-verifiable questions.

Data processing is explicitly structured around contamination control and response diversity. Deduplication uses N-gram overlap on concatenated question-response pairs, while allowing multiple correct responses per question. Decontamination uses N-gram overlap removal against AIME24, AIME25, and MATH500. Verification is source-specific in places: OpenR1 uses MathVerify and an LLM-based judge, while the overall curation emphasizes correctness and long-form CoT.

The SFT configuration fixes several training choices that are later treated as reproducibility-critical: 3 epochs; peak learning rate $5.0 \times 10^{-5}$ with cosine scheduler; warmup ratio 10%; batch size 128; no-packing strategy; LlamaFactory implementation; cutoff length 26,000; and initialization from Qwen-2.5-Math-7B. The paper reports that no-packing was empirically superior to packing and neat-packing in this setup. In the ablation table “No-Packing usually leads to better performance,” packing yields AIME24 35.41, AIME25 26.66, MATH500 89.06; neat-packing yields 32.50, 26.25, 88.80; and no-packing yields 38.12, 29.37, 90.40. A separate ablation reports that selecting long-trajectory examples instead of random examples at equal counts improves results at both 30k and 50k training subsets.

Under the main SFT evaluation protocol—avg@k over repeated runs with $k=64$ for AIME24 and AIME25, $k=5$ for MATH-500, temperature 0.6, top-p 0.95, and maximum generation length 32,768—MiroMind-M1-SFT-7B obtains AIME24 60.4, AIME25 45.0, and MATH-500 94.6. The paper states that this outperforms the DeepSeek-R1 distilled model trained from the same initial checkpoint and MiMo-7B-SFT, which records 58.7, 44.3, and 93.0.

4. RLVR dataset construction and CAMPO

The RL stage uses MiroMind-M1-RL-62K, a curated RLVR corpus assembled from NuminaMath-1.5 (896K), Skywork-OR1-RL-Data math subset (105K), Big-Math (50K, HARP + reformulated), and DAPO-Math-17K (17K) (Li et al., 19 Jul 2025). Inclusion-exclusion filtering reduces approximately 1M candidates by approximately 94% to 62,118 verifiable, curated math problems. Filtering proceeds through style and format constraints, exact and near-duplicate removal via a 10-gram similarity threshold, difficulty-aware selection through offline rollouts, exclusion of fully correct and fully incorrect problems, and exclusion of target answers longer than 20 characters. The corpus is also decontaminated against AIME24/2025 and MATH500.

The reward function is correctness-based: $r(o,a)$ indicates correctness through $\mathrm{is\_equivalent}(a,o)$. Correctness is computed by an improved cascade math verifier that fixes edge cases involving units, $\pi$ and degrees, percentages, and small numerical tolerances. The paper treats verifier quality as central to RL stability and sample efficiency, and reports that better verifier accuracy improves RL stability and performance while rewarded correct responses show shorter reasoning traces.

CAMPO, the project’s main algorithmic contribution, formalizes RL as

$$\max_{\theta} J(\pi_{\theta}) \triangleq \mathbb{E}_{(q,a)\sim \mathcal{D}} \mathbb{E}_{o\sim \pi_{\theta}(\cdot|q)} [r(o,a)].$$

Its distinguishing mechanisms are multi-stage length-progressive training, an adaptive repetition penalty $f(o_i)$, and filtering to questions whose $G$ rollouts are neither all-correct nor all-incorrect. The group-normalized advantage is defined as

$$\hat{A}_{i,t} \triangleq \frac{r(o_i,a)-f(o_i)-\mathrm{mean}(\{r(o_i,a)-f(o_i)\}_{i=1}^{G})} {\mathrm{std}(\{r(o_i,a)-f(o_i)\}_{i=1}^{G})}.$$

The stage index $k=64$0 controls decoupled clipping through $k=64$1 and $k=64$2, while $k=64$3 is the repetition penalty, defined as the proportion of the output sequence within a detected repeating loop, with earlier repetition penalized more.

Length progression is explicit. For the 32B model, CAMPO uses three stages with maximum response lengths 16,384, 32,768, and 49,152. Responses exceeding the current stage’s maximum length are assigned zero reward. For the 7B model, CAMPO uses two stages: a 16K cap for 1,200 steps, then 32K thereafter. This suggests that context control is used as a training signal rather than only as an inference constraint.

5. Training configuration, benchmark results, and token efficiency

The RL configurations are fully specified in the release materials (Li et al., 19 Jul 2025). For the 32B model, the setup omits KL loss; uses learning rate $k=64$4, clipping ratio 0.2, temperature 1.0, $k=64$5 rollouts per prompt, batch size 32, and mini-batch size 32; trains strictly on-policy; and uses compute described as 16×8 A100 GPUs. For the 7B model, the same RL data are used, initialization is from MiroMind-M1-SFT-7B, the two-stage 16K→32K schedule is applied after 1,200 steps, and distributed training is implemented with VERL on 8 nodes × 8 NVIDIA A800, for 64 GPUs total. RL evaluation uses temperature 1.0 to match RL training.

The principal benchmark results are as follows:

Model	Benchmark set	Scores
MiroMind-M1-SFT-7B	AIME24 / AIME25 / MATH-500	60.4 / 45.0 / 94.6
MiroMind-M1-RL-7B	AIME24 / AIME25 / MATH500	73.4 / 57.8 / 96.7
MiroMind-M1-RL-32B	AIME24 / AIME25 / MATH500	77.5 / 65.6 / 96.4

The comparison baselines reported in the same table place MiroMind-M1-RL-7B above DeepSeek-R1-Distill-Qwen-7B (55.5, 39.2), Light-R1-7B-DS (59.1, 44.3), and Skywork-OR1-7B (72.2, 54.6). For 32B Qwen2.5-series models, DeepSeek-R1-Distill-Qwen-32B records AIME24 70.8, AIME25 52.1, MATH500 95.8; Skywork-OR1-32B-Preview records 77.1, 68.2, 97.5; and MiroMind-M1-RL-32B records 77.5, 65.6, 96.4.

A notable secondary result is token efficiency. Mean response lengths are computed over 64 runs, both overall and conditioned on correct answers. The paper reports that MiroMind-M1-RL-32B generates significantly shorter responses than Skywork-OR1-32B-Preview across maximum response length budgets from 4k to 16k, with minimal impact on accuracy; under shorter budgets, MiroMind-M1-RL-32B outperforms Skywork-OR1-32B-Preview. For correct, rewarded outputs, MiroMind-M1-RL-32B also produces shorter reasoning chains than Skywork-OR1-32B-Preview. The paper attributes this to the combination of CAMPO’s repetition penalty, the cascade verifier, and multi-stage length control.

6. Reproducibility, limitations, and research significance

Reproducibility is a primary design feature rather than an afterthought (Li et al., 19 Jul 2025). The release includes full SFT and CAMPO configurations, evaluation scripts, the improved verifier, and both training datasets. The practical usage guidance is also explicit: for SFT-style evaluation, use temperature 0.6, top-p 0.95, and maximum generation length 32,768; for RL-trained models, use temperature 1.0 to align with RL sampling. The paper notes that prompt templates are not specified and that the paper does not state license details or specific prompt templates; usage is therefore tied to repository and model-card licenses once published.

Several limitations are stated directly. The 32B model lags Skywork-OR1-32B-Preview on AIME25, 65.6 versus 68.2, which the authors attribute to Skywork’s broader mixture of math and code, whereas MiroMind-M1-RL training focused solely on math. Training efficiency remains constrained by rollout cost and by long-tail generations that slow training because of synchronization and GPU idle time. Evaluation stability is also a documented issue: despite averaging over 64 runs, AIME24 and AIME25 scores can have “two-sided standard deviation … exceed 8%,” reflecting the small test size of 30 questions. In the RL objective, KL-free training is said to encourage long-chain reasoning but also to increase drift risk; the repetition penalty is introduced for stability, but the authors note that further exploration may improve robustness.

In methodological terms, MiroMind-M1’s significance lies less in a single benchmark number than in the completeness of the released stack. A plausible implication is that the project functions as a reference implementation for open mathematical RLVR: verified SFT data, curated RL data, an explicit verifier, long-context training settings, and a documented CAMPO procedure are all released together. Within the Qwen-2.5-based open 7B/32B math RLM landscape, the reported results position MiroMind-M1 as both a competitive benchmark system and a reproducible substrate for subsequent work on data curation, verifier design, token-efficient reasoning, and KL-free policy optimization.