Rimo Eval

rimo-eval

RIMO: An Easy-to-Evaluate, Hard-to-Solve Olympiad Benchmark for Advanced Mathematical Reasoning — Chen et al. (2025) (arXiv:2509.07711, 2025)

What this evaluates

This benchmark evaluates advanced mathematical reasoning at the International Mathematical Olympiad level. It probes two distinct capabilities: deriving a unique integer answer from complex problem statements, and constructing step-by-step deductive proofs by solving decomposed sub-problems.

Datasets

• RIMO — total 791; splits: RIMO-N (335), RIMO-P (456)

Metrics

• exact-match grading (primary) — range: [0, 1]
  • Calculated as the proportion of instances where the model's output exactly matches the ground-truth answer or sub-solution string. For RIMO-N, it is a direct string comparison of the final integer. For RIMO-P, it is averaged across all sub-problems in the proof sequence.

Input / output format

Input: For RIMO-N: A rewritten problem statement requiring a single integer answer. For RIMO-P: An original proof problem decomposed into a sequence of 1–4 guided sub-problems.

Output: For RIMO-N: A single integer. For RIMO-P: A sequence of step-by-step sub-solutions corresponding to each guided sub-problem.

Scoring recipe

def compute_rimo_score(predictions, golds, track):
    correct = 0
    for pred, gold in zip(predictions, golds):
        if track == 'RIMO-N':
            if pred.strip() == gold.strip():
                correct += 1
        elif track == 'RIMO-P':
            if all(p.strip() == g.strip() for p, g in zip(pred, gold)):
                correct += 1
    return correct / len(predictions)

Common pitfalls

• RIMO-N grading is strictly deterministic string comparison; any extra text, units, or formatting causes a zero score.
• RIMO-P requires solving sub-problems in the exact prescribed order; models that skip steps or hallucinate intermediate lemmas will fail automated validation.
• The benchmark explicitly avoids symbolic post-processing or learned judges, so relying on external solvers or LLM-as-a-judge will invalidate the protocol.

Evidence (verbatim from paper)

RIMO-N rewrites 335 IMO problems into single-integer-answer formats enabling deterministic exact-match grading, while RIMO-P decomposes 456 proof problems into sub-problems to evaluate step-by-step reasoning via automated, expert-verified sub-solution validation.

Citation

@misc{chen2025rimo,
  title={RIMO: An Easy-to-Evaluate, Hard-to-Solve Olympiad Benchmark for Advanced Mathematical Reasoning},
  author={Chen et al. (2025)},
  year={2025},
  note={arXiv:2509.07711}
}

• arXiv: 2509.07711