AI Trend Notifier
EN
← wiki

$ cat wiki/papers/2026/2605.13301-olympiad-reasoning.md

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

paperupdated 2026-08-14created 2026-05-16

arxiv: 2605.13301 Authors: 28, first author Yafu Li Submitted: 2026-05-13 · captured from HF Daily 2026-05-16 (#1, 134 upvotes)

TL;DR

A recipe for converting an already post-trained reasoning backbone into an olympiad solver by combining curriculum learning, reinforcement-learning stages and test-time scaling — claimed to reach gold-medal-level performance on maths and physics olympiads with a 30B model (source). The contribution is framed as simple and unified rather than novel: the abstract notes that several systems already reach this level, so the claim is about how cheaply the result can be reproduced.

Method

Three ingredients, named together in the abstract as one recipe (source):

ElementAs described
Backbone30B parameters, already post-trained for reasoning
Supervised stage~340,000 training trajectories
RL stage200 reinforcement-learning steps
Inferencetest-time scaling, stable on trajectories exceeding 100K tokens
The resulting model is called SU-01. Nothing in the abstract describes a new
architecture or objective — the claim rests on the combination and its ordering.

Results

Claimed gold-medal-level performance on IMO 2025 / USAMO 2026 and IPhO 2024 / 2025, plus generalisation to scientific reasoning beyond the competition set (source).

The abstract page carries no numeric scores and no medal counts. The results table is in the 77-page report body, which has not been read for this page. Any figure attributed to this paper must cite that report rather than this page — a qualitative claim restated as a number is the failure this wiki's checks exist to catch.

Significance

  • HF Daily #1 on 2026-05-16 with 134 upvotes — that community's popularity signal on the day, not a quality judgement
  • The simple and unified framing is a claim about reproduction cost, which bears directly on the Test-Time Compute (Inference-Time Compute Scaling) scaling debate: if a 30B backbone plus 200 RL steps reaches this level, the interesting variable is the recipe rather than the parameter count
  • 28 authors across a 77-page technical report indicates a large collaboration

Open Questions

  • The abstract asserts gold-medal-level performance without scores. What does the report's results table actually show, and against which reference solutions?
  • How is compute split between the RL stage and test-time scaling? "Simple" is doing a lot of work if the inference budget is what carries the result.
  • Does the recipe transfer to domains without olympiad-style verifiable answers?

Cite

@article{olympiad_unified_scaling_2026,
  title={Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling},
  year={2026},
  eprint={2605.13301},
  archivePrefix={arXiv}
}

Referenced by

Sources