Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling
:::info Stub — Full Engineering Breakdown Coming This paper has a linked code implementation and was featured on Hugging Face Papers with 143 upvotes. A full breakdown with production viability rating, implementation notes, and honest limitations is being written. Subscribe to AI Letters → :::
| Authors | Yafu Li et al. |
| Year | 2026 |
| HF Upvotes | 143 |
| arXiv | 2605.13301 |
| Download | |
| Code | https://github.com/Simplified-Reasoning/SU-01 |
Abstract
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.
Engineering Breakdown
Plain English
This paper presents a three-stage training recipe that takes a 30B parameter reasoning model and pushes it to gold-medal performance on math and physics olympiad problems. The approach combines supervised fine-tuning with a reverse-perplexity curriculum (to teach rigorous proof search), followed by two-stage reinforcement learning (first with verifiable rewards, then proof-level RL), and finishes with test-time scaling to squeeze out final performance gains.
Key Engineering Insight
The unified recipe treats reasoning as a learnable, scalable behavior rather than an emergent property—by explicitly training for proof-search rigor early (SFT), then reinforcing correct reasoning paths (RL), then computing longer at inference time, they decouple capability gains from raw model scaling. This is actionable: you can retrofit existing models without retraining from scratch.
Why It Matters for Engineers
For teams building reasoning systems in production (financial modeling, scientific discovery, code generation), this paper shows a reproducible path from base models to expert-level performance without building trillion-parameter systems. The staged approach (SFT → RL → inference-time compute) maps directly to production pipelines you can actually implement and debug incrementally.
Research Context
Prior work showed reasoning models could solve hard problems, but relied on scaling model size or compute at training time. This paper advances the field by proving that structured training recipes—specifically curriculum design and proof-aware RL—unlock olympiad-level reasoning on modest-sized backbones (30B parameters). It positions test-time scaling as a practical alternative to massive model sizes, which has direct implications for cost-effective deployment.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::