ScienceArena

EMNLP 2026 · Main Conference

ScienceArena

Benchmarking LLMs in Science Olympiad Competitions

Real olympiad papers in physics, chemistry and biology, graded sub-question by sub-question against the official marking schemes by an automatic grader calibrated with olympiad gold medalists. Every score opens up to the problem, the model's answer and the reason behind each point.

Explore the leaderboard How scoring works
Competitions
11
Models evaluated
19
Sub-questions
413
Disciplines
3

Key findings

Frontier models now compete with olympiad medalists

Scores are on each competition's official scale and compared with the official medal cut-offs and the top human contestant. Click a card to open that leaderboard.

The medalists' comments show where points are still lost: derivations often earn full marks, while misread plots, constraints that run across several parts, and molecular structures that are not uniquely determined cost the most.

Leaderboard

Scores by competition

Pick a group and a competition. In a single competition, click any problem cell to read the problem, the model's answer, the official solution and rubric, and the grader's reasoning.

Scoring note.

Leaderboard

Correct Partial Wrong Missing possible timing/data leakage

Problem-level records with model answer, evaluation feedback, official answer, rubric, and figures.

Method

How ScienceArena works

Olympiad problems are the hardest test of scientific reasoning, but two things kept them out of large-scale evaluation: the papers are PDFs full of formulas, plots and structures, and multi-step solutions need expert graders. ScienceArena removes both obstacles.

  1. From official papers to evaluation data

    Official problems, solutions and marking schemes are parsed, including formulas, plots, molecular structures and apparatus diagrams, into sub-questions that keep the dependencies between parts and the partial-credit rules. Olympiad medalists check every item. New competitions go through the same pipeline.

  2. Grading calibrated by gold medalists

    Olympiad gold medalists graded 525 sub-questions from five models on IPhO 2025 and IChO 2025 against the official rubrics and commented on nearly 400 of them. Their comments became explicit grading rules. After calibration, the automatic grader's totals are within 1 point of the medalists' (out of 30 for IPhO and 60 for IChO).

  3. Every point is traceable

    Open any cell to see the problem and its figures, the official answer and rubric, the model's raw response, and the points awarded with the reason for each. For IPhO and IChO the medalists' comments sit next to the automatic grading.

  4. Release-dated, contamination-aware

    Each competition records when its papers became public, so new contests can be added as they are released. Scores from models released after a paper went public are marked ! for possible leakage; models released before it are tagged pre-release.

Coverage

Competitions covered

International, continental and national olympiads in physics, chemistry and biology. Click a competition to open its leaderboard.

This English edition hides the problem text of the two Chinese-language competitions (CPhO and CChO); their scores and grading notes are shown.