kumarjau1998/AIME2025

AIME 2025 with Oracle Strategy Sketches

Thirty problems from the American Invitational Mathematics Examination (AIME) 2025 I and II, with GPT-6 Astra solutions, curator-approved strategy sketches, and three load-bearing solution steps. Records use the same field names and nested solution format as AOBench and its IMO-ProofBench subset.

Configurations and loading

Configuration	File	Problems
aime2025 (default)	aime2025.jsonl	30
aime2025_i	aime2025_i.jsonl	15
aime2025_ii	aime2025_ii.jsonl	15

All configurations use a test split. The default configuration concatenates AIME I and II in exam/problem order, preserving stable IDs such as AIME2025-I-01.

from datasets import load_dataset

problems = load_dataset("kumarjau1998/AIME2025", "aime2025", split="test")
problem = problems[0]
reference = problem["solutions"][0]
statement = problem["statement"]
solution = reference["solution"]
sketch = reference["sketch"]
steps = reference["steps"]

Record format

Field	Meaning
problem_id	Stable exam/problem identifier.
dataset	aime2025.
statement	Imported problem statement, without rewriting.
domain	algebra, combinatorics, geometry, or number_theory.
source	AIME I or AIME II.
url	Pinned upstream problem-file URL.
country, year, date	United States, 2025, 2025-02.
task	compute_and_prove: compute the requested answer and justify it in the provided reference solution.
answer_type	number; AIME answers are integers from 0 through 999.
subtopic	One primary mathematical subtopic assigned by the curator.
solutions	List containing one AI-generated reference solution.

The term reference identifies the route used to prepare the oracle sketch; it does not mean an official contest solution. Numerical answers occur in the full solutions. There is no additional top-level answer field, matching AOBench's schema.

Solution field	Meaning
solution_id	Stable problem-prefixed route ID.
role	reference.
source_type	ai_generated.
source_url	null: solutions were generated, not imported from a published proof.
generated_by	GPT-6 Astra, using model ID gpt-6-astra.
solution	Complete generated mathematical solution.
sketch	One strategy sketch of 18–24 whitespace-delimited words.
steps	Three ordered load-bearing steps, each 17–20 whitespace-delimited words.
note	Generation and review provenance.

Unlike the non-geometry AOBench/IMO-ProofBench collections, this release includes all AIME 2025 problems, including geometry. Its domains are:

Domain	Problems
algebra	5
combinatorics	10
geometry	11
number_theory	4

Generation and review

1. GPT-6 Astra generated complete solutions from the statements alone using high reasoning effort through a Codex-subscription gateway.
2. Each final answer was checked against the imported answer key. A fresh GPT-6 Astra request reviewed each complete solution.
3. The reviewed route was compressed into an 18–24-word sketch, matching the mathematical prose/notation style used for the study's IMO sketches.
4. A fresh GPT-6 Astra request reviewed each sketch for route fidelity, mathematical usefulness, and answer leakage. Two rejected sketches were revised and reviewed again.
5. The dataset curator reviewed and approved all thirty sketches for publication.
6. Three-step outlines were extracted from the checked solutions and reviewed in fresh GPT-6 Astra requests.

Full solutions and outlines have model-assisted review, using the same model family as the generator. Correlated model errors remain possible. Curator approval applies to the sketches; no independent expert-panel review of the full solutions or outlines is claimed. Agreement with the answer key alone does not certify every proof step. Provenance is recorded in provenance.json.

Intended use and limitations

Use for no-sketch versus oracle-sketch comparisons, strategy-conditioned execution, and test-time compute experiments. Keep full solutions and source metadata outside the solver's accessible inputs when measuring sketch assistance. The three-step outlines are a richer assistance condition and should be reported separately. Freeze the strategy annotations before comparing target-model outcomes.

AIME 2025 is public and may occur in model training data. This dataset is not a clean unseen-problem generalization test. The upstream statements are reproduced as text; formatting quirks are retained. The source answer for AIME II #5 had a degree suffix, which was normalized to the integer 336 when checking answers.

Attribution and licensing

Problems originate from the Mathematical Association of America's AIME 2025 examinations. Statements and answer keys were obtained from OpenCompass's AIME2025 dataset, at revision a6ad95f611d72cf628a80b58bd0432ef6638f958. That upstream dataset declares the MIT license. Its attribution and applicable third-party terms are retained; the upstream metadata is not a claim that this curator owns the contest statements.

The new solution annotations and compilation are released under CC BY 4.0, matching the annotation license used for AOBench. See provenance.json for source URLs, checksums, normalization, and review details.

For the associated discovery/execution study, see the AOBench dataset card. When using this release, cite this dataset repository and credit MAA and OpenCompass for the original problems and upstream transcription.

MathArena statement alignment

Released statements and diagram source are aligned with pinned MathArena datasets. AIME2026 I #10 retains the corrected rotated segment A'C'; I #13 retains nonnegative r and the summation index m. Original generation requests are preserved: the solutions and sketches were generated before this wording update. No solutions, sketches, or outlines were regenerated. Source revisions and per-problem changes are recorded in provenance.json.

MathArena statement transcriptions are licensed CC BY-NC-SA 4.0: https://creativecommons.org/licenses/by-nc-sa/4.0/ Attribution: MathArena, https://huggingface.co/MathArena. The aligned release is distributed under CC BY-NC-SA 4.0. Original annotation licensing and source attributions below remain recorded for provenance.