Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling

2026-05-15 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

The SU-01 model, developed by Shanghai AI Laboratory and collaborators, achieves gold-medal-level performance on International Mathematical Olympiad (IMO) and International Physics Olympiad (IPhO) problems. This 30B-A3B reasoning backbone is trained using a "simple and unified recipe" involving three stages: rigorous Supervised Fine-Tuning (SFT) with a reverse-perplexity curriculum on 340K sub-8K-token trajectories, a two-stage Reinforcement Learning (RL) pipeline progressing from verifiable rewards to proof-level RL, and finally, test-time scaling. SU-01 supports stable reasoning on trajectories exceeding 100K tokens and demonstrates strong generalization to scientific reasoning beyond mathematics and physics, outperforming similar-sized models and approaching commercial systems like Gemini 3.1 Pro Thinking on benchmarks like IMO-ProofBench and FrontierScience-Research.

Key takeaway

Research Scientists developing advanced reasoning models should consider adopting a modular, multi-stage training pipeline like SU-01's. By combining rigorous SFT with a reverse-perplexity curriculum, two-stage RL for behavior scaling, and test-time self-verification, you can push broadly capable backbones to expert-level performance in complex domains like mathematical and scientific Olympiads, while preserving generalization across scientific tasks.

Key insights

A compact 30B-A3B model achieves Olympiad-level reasoning through a multi-stage training and inference pipeline.

Principles

• Specializable-generalist models excel with targeted training.
• Reverse-perplexity curricula stabilize SFT on post-trained models.
• Multi-stage RL refines reasoning from correctness to proof quality.

Method

The SU-01 recipe involves SFT with a reverse-perplexity curriculum, followed by two-stage RL (coarse with verifiable rewards, refined with proof-level generative rewards), and concludes with test-time self-verification and refinement for enhanced performance.

In practice

• Use long-form SFT to instill disciplined proof-search patterns.
• Employ RL with verifiable rewards for efficient reasoning improvement.
• Apply test-time scaling for elevating trained models to expert levels.

Topics

• Olympiad Reasoning
• Supervised Fine-tuning
• Reinforcement Learning
• Test-Time Scaling
• SU-01 Model

Code references

• Simplified-Reasoning/SU-01

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Student

Related on AIssential

• Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling — A unified recipe scales reasoning models to gold-medal Olympiad performance via SFT and two-stage RL.
• Learning What to Learn: Stage-Specific Data Sets for SFT-then-RL in Small Language Model Reasoning — Data strategy for SFT-then-RL pipelines should align with each stage's distinct role in skill acquisition and consolidation.
• Scaling Behaviors of LLM Reinforcement Learning Post-Training: An Empirical Study in Mathematical Reasoning — Larger LLMs demonstrate superior efficiency and performance in RL post-training for mathematical reasoning.
• Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability — Reasoning SFT can generalize cross-domain, but it is conditional on optimization, data quality, and model capability.
• DeReason: A Difficulty-Aware Curriculum Improves Decoupled SFT-then-RL Training for General Reasoning — Difficulty-aware data partitioning for SFT-then-RL training significantly enhances LLM general reasoning.
• Provable Benefits of RLVR over SFT for Reasoning Models: Learning to Backtrack Efficiently — RLVR's outcome-based rewards enable LLMs to learn efficient backtracking, unlike SFT, leading to exponential inference compute gains.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.