TL;DR
SU-01, a 30B-A3B model, achieves gold-medal-level olympiad reasoning through a three-stage pipeline (SFT, RL, test-time scaling). It reaches 35 points on IMO 2025 and USAMO 2026, matching the IMO gold line and exceeding USAMO gold by 10 points—matching the highest human score among 340 competitors.
SU-01, a 30B-A3B model, achieves gold-medal-level olympiad reasoning through a three-stage pipeline (SFT, RL, test-time scaling). It reaches 35 points on IMO 2025 and USAMO 2026, matching the IMO gold line and exceeding USAMO gold by 10 points—matching the highest human score among 340 competitors.
论文原始摘要
Recent progress in reasoning models has substantially advanced long-horizon mathematical and scientific problem solving, with several systems now reaching gold-medal-level performance on International Mathematical Olympiad (IMO) and International Physics Olympiad (IPhO) problems. In this paper, we introduce a simple and unified recipe for converting a post-trained reasoning backbone into a rigorous olympiad-level solver. The recipe first uses a reverse-perplexity curriculum for SFT to instill rigorous proof-search and self-checking behaviors, then scales these behaviors through a two-stage RL pipeline that progresses from RL with verifiable rewards to more delicate proof-level RL, and finally boosts solving performance with test-time scaling. Applying this recipe, we train a 30B-A3B backbone with SFT on around 340K sub-8K-token trajectories followed by 200 RL steps. The resulting model, SU-01, supports stable reasoning on difficult problems with trajectories exceeding 100K tokens, while achieving gold-medal-level performance on mathematical and physical olympiad competitions, including IMO 2025/USAMO 2026 and IPhO 2024/2025. It also demonstrates strong generalization of scientific reasoning to domains beyond mathematics and physics.
Paper Collector 中文速览
提出统一训练配方实现奥数金牌级推理
方法概述
先通过逆向困惑度课程SFT灌输证明搜索和自检行为，再用两阶段RL从可验证奖励扩展到证明级RL，最后测试时扩展提升性能
核心贡献
提出逆向困惑度SFT课程+两阶段RL+测试时扩展的统一配方，训练出稳定支持100K+token推理的金牌级竞赛模型
原始来源与核验范围
核验仅覆盖论文正文中的主张、证据片段和参考文献关系。NGJOO 未独立运行作者代码、重做实验或验证真实部署效果；页面中的可复现性分数是文档完整度评估，不是复现实验结果。
47
已证实
6
证据不足
0
无法验证
N/A
可复现性
核心问题
Can a compact 30B-A3B model achieve gold-medal-level performance on mathematical olympiad problems through a simple, unified training pipeline?
核心方法
The authors develop a three-stage pipeline starting from P1-30B-A3B: (1) Supervised Fine-Tuning with reverse-perplexity curriculum on 338K trajectories to reshape reasoning behavior, (2) Reinforcement Learning split into coarse RL using GSPO for answer-seeking and refined RL with generative proof rewards for proof quality, and (3) Test-Time Scaling with iterative solve-verify-refine loops allocating additional inference compute.
方法组件
- Inference is organized as an iterative solve-verify-refine procedure with maximum generation length of 160,000 tokens.
- Explicit stopping rules include 5 consecutive verification passes for acceptance and 10 consecutive failures for early termination.
- Each problem allows at most 10 independent runs with 30 solve-verify-refine cycles per run.
- The DeepSeekMath-V2 reward server uses speculative decoding with 3-step scheme and 4 draft tokens per step.
- Different maximum generation lengths are used for different benchmarks, ranging from 131K to 256K tokens.
- The comparison uses public technical reports, official model cards, and official benchmark reports.
- The larger-model group contains DeepSeek-V3.2, GPT-5.5, and Gemini 3.1 Pro Thinking.
- The similar-size group contains GLM-4.7-Flash, Nemotron-Cascade-2-30B-A3B, Qwen3.6-35B-A3B, Gemma-4-31B, and SU-01.
- GPT-5.5-High is the high-reasoning variant of GPT-5.5, and Gemini 3.1 Pro Thinking is from the Gemini 3.1 Pro family.
- The section lists model-generated solutions for human-expert scoring in Table 4.
论点验证
已证实 (90%)
Using a 30B-A3B model, we build a modular pipeline: SFT reshapes reasoning behavior, RL scales solving capability, and TTS allocates additional inference compute to the hardest proof-search problems.
已证实 (75%)
The design follows a specializable-generalist view: rather than building a narrow olympiad solver, we specialize a broadly capable post-trained model toward expert-level proof reasoning while preserving transfer across scientific domains.
已证实 (90%)
We then order the examples by reverse perplexity so that each pass starts with trajectories most mismatched to the initial policy before consolidating on more familiar examples.
The reverse-perplexity curriculum is fully specified in Section 2.3 and empirically validated in Section 6.3 with controlled comparisons showing descending-PPL ordering outperforms random and low-PPL-first orderings with specific numerical results.
The reverse-perplexity curriculum is fully specified in Section 2.3 and empirically validated in Section 6.3 with controlled comparisons showing descending-PPL ordering outperforms random and low-PPL-first orderings with specific numerical results.
已证实 (85%)
Refined RL then shifts the target from answer correctness to proof quality. It combines a proof-level generative reward model for scoring complete proofs, self-refinement prompts for training critique-and-repair behavior, and experience replay for preserving rare successful trajectories on hard problems.
证据不足 (50%)
On answer-verifiable benchmarks, the resulting model, SU-01, nearly matches the strongest similar-size baseline, Qwen3.6-35B-A3B, across AnswerBench, AMO-Bench, AIME 2025/2026, and FrontierScience-Olympiad.
已证实 (90%)
On proof-oriented evaluation, SU-01 reaches 57.6% on IMO-ProofBench with direct generation and 70.2% with TTS, substantially outperforming similar-size models and approaching competitive commercial systems such as Gemini 3.1 Pro Thinking.
证据不足 (50%)
SU-01 obtains the best similar-size overall score on FrontierScience-Research, suggesting that the recipe generalizes scientific reasoning toward research-style problems beyond olympiad benchmarks.
证据不足 (50%)
Direct SU-01 already exceeds the IPhO gold lines for both 2024 and 2025, and clears the bronze-medal lines on IMO 2025 and USAMO 2026.
已证实 (90%)
With test-time scaling, it reaches 35 points on both mathematical olympiads, meeting the IMO 2025 gold line and exceeding the USAMO 2026 gold line by 10 points.
Specific numerical results are provided: 35 points on both IMO 2025 and USAMO 2026 with TTS, meeting the IMO gold line exactly and exceeding USAMO gold line by 10 points. The paper provides context for these achievements.
Specific numerical results are provided: 35 points on both IMO 2025 and USAMO 2026 with TTS, meeting the IMO gold line exactly and exceeding USAMO gold line by 10 points. The paper provides context for these achievements.
已证实 (90%)
On USAMO 2026, this matches the highest reported human total among 340 competitors, indicating that the overall recipe can elicit top-level human-like olympiad reasoning from a compact 30B-A3B model.
The paper provides specific context: USAMO 2026 had 340 competitors with maximum score of 35, and SU-01's TTS score of 35 matches this highest reported human total. This is a concrete, verifiable claim with specific numbers.
The paper provides specific context: USAMO 2026 had 340 competitors with maximum score of 35, and SU-01's TTS score of 35 matches this highest reported human total. This is a concrete, verifiable claim with specific numbers.
已证实 (85%)
SU-01 can sustain reasoning trajectories beyond 100K tokens, condition on its own drafts and error analyses, and repeatedly verify and repair candidate proofs.
已证实 (80%)
We choose P1-30B-A3B (Chen et al., 2025) as the initial model because it already shows competitive performance in scientific reasoning, including both mathematics and physics.
This is a design choice with referenced justification. The paper cites Chen et al. 2025 for P1-30B-A3B's competitive performance in scientific reasoning. As a design choice with external citation, this is adequately supported.
This is a design choice with referenced justification. The paper cites Chen et al. 2025 for P1-30B-A3B's competitive performance in scientific reasoning. As a design choice with external citation, this is adequately supported.
证据不足 (40%)
We empirically find that applying SFT to a post-trained backbone is more efficient than training the same reasoning behavior from a base model.
The paper states this as an empirical finding but provides no controlled experiment comparing SFT on post-trained vs. base model. No ablation study or comparative data is shown. This is an assertion without supporting experimental evidence.
The paper states this as an empirical finding but provides no controlled experiment comparing SFT on post-trained vs. base model. No ablation study or comparative data is shown. This is an assertion without supporting experimental evidence.
已证实 (95%)
After filtering, the SFT mixture contains 338K trajectories with responses shorter than 8K tokens.
This is a straightforward design choice with specific numbers stated in multiple locations. The 338K trajectories and 8K token limit are clearly specified.
This is a straightforward design choice with specific numbers stated in multiple locations. The 338K trajectories and 8K token limit are clearly specified.
已证实 (85%)
We then filter noisy generations and remove trajectories longer than 8,192 tokens. This filtering step keeps the supervised signal focused on rigorous and usable reasoning traces, while avoiding extremely long outputs that are more likely to introduce truncation or unstable optimization.
This is a design choice with clear justification. The filtering criteria (noisy generations, >8192 tokens) and rationale (focused signal, avoiding truncation/instability) are explicitly stated.
This is a design choice with clear justification. The filtering criteria (noisy generations, >8192 tokens) and rationale (focused signal, avoiding truncation/instability) are explicitly stated.
证据不足 (45%)
For trajectories capped at 8,192 tokens, we empirically find that four epochs are usually sufficient to recover most of the model capability after the initial behavioral shift, provided that the data mixture and learning rate are well controlled.
The paper states this as an empirical finding but provides no controlled experiment comparing different epoch counts. No ablation showing performance at 1, 2, 3, 4+ epochs is presented. This is an assertion without supporting experimental data.
The paper states this as an empirical finding but provides no controlled experiment comparing different epoch counts. No ablation showing performance at 1, 2, 3, 4+ epochs is presented. This is an assertion without supporting experimental data.
证据不足 (45%)
In practice, we find that a truncation rate below 5% is a useful sign that the model has largely adapted to the target reasoning style.
已证实 (90%)
To make long-CoT SFT more stable, we use a reverse-perplexity training curriculum.
The reverse-perplexity curriculum is a methodological contribution that is fully specified with mathematical formulation and empirically validated through controlled comparison in Section 6.3.
The reverse-perplexity curriculum is a methodological contribution that is fully specified with mathematical formulation and empirically validated through controlled comparison in Section 6.3.
已证实 (85%)
We split this stage into two levels. Coarse RL converts the SFT reasoning pattern into stronger answer-seeking behavior under reliable, mostly verifiable reward signals, improving search, coverage, and direct solving performance on hard tasks. Refined RL then specializes the policy toward complete, auditable proof construction.
This is a design choice that is fully specified. The two-level RL structure (coarse and refined) is described in detail with specific objectives and mechanisms for each level.
This is a design choice that is fully specified. The two-level RL structure (coarse and refined) is described in detail with specific objectives and mechanisms for each level.
已证实 (95%)
The final RL pool contains 8,967 verifiable prompts and 16,287 non-verifiable prompts.
This is a straightforward design choice with specific numbers clearly stated.
This is a straightforward design choice with specific numbers clearly stated.
... 共 53 个论点
可复现性评估
较低可复现性 (0%)
缺失的复现细节
- Training hyperparameters (learning rate, batch size, number of training steps/epochs, optimizer settings, weight decay, warmup schedule)
- Model architecture details for SU-01 (layer configuration, hidden dimensions, attention mechanism, parameter count)
- Training data details (dataset composition, size, preprocessing steps, data sources for refined RL)
- Random seeds for reproducibility of experiments
- Complete hardware specifications for training (GPU type, number of nodes, memory requirements)
- Training duration and computational cost
- Model weights/checkpoints availability
- Exact implementation details of the refined RL algorithm (reward function formulation, RL objective, policy gradient specifics)
- Verification model architecture and training details (DeepSeekMath-V2 reward server setup beyond deployment specs)
- Data splits for training/validation/test sets
局限与证据边界
- The failures show a clear limitation: the model can miss subtle structural constraints, as in the invalid column-permutation reduction in IMO P6, or leave gaps in delicate global strategy arguments, as in USAMO P2.
- Overall, the model performs well when a problem admits a rigid formal representation, but is less reliable when the core challenge is preserving combinatorial structure or proving a finely tuned process invariant.
本分析由 PDF 阅读助手 自动生成，仅供参考，不构成学术评审意见。验证结论和可复现性评估基于论文文本自动分析，可能存在偏差。原始论文请参阅 arXiv。
分析时间：2026-05-15T13:18:26+00:00 · 数据来源：Paper Collector