NaturalProofs
Models retrieve or generate references needed for a mathematical theorem's proof. Ranking and sequence metrics compare predicted references with those used in the source proof.
The released tasks evaluate reference selection rather than complete proof generation. Sources include ProofWiki, the Stacks Project, and textbooks.
- Released
- April 2021
- Kind
- Written proofs
- Level
- University
- Scored
- Checked automatically
- Answers
- Short answer
- Tasks
- All public
- Measures
- Capability
- Subject
- Mathematics
- Language
- English
- Interaction
- Single turn
- Items from
- Real work and use, Exams and contests
- Refresh
- One fixed set
- Run by
- Anyone
In the authors’ words
Understanding and creating mathematics using natural mathematical language - the mixture of symbolic and natural language used by humans - is a challenging and important problem for driving progress in machine learning. As a step in this direction, we develop NaturalProofs, a multi-domain corpus of mathematical statements and their proofs, written in natural mathematical language. NaturalProofs unifies broad coverage, deep coverage, and low-resource mathematical sources, allowing for evaluating both in-distribution and zero-shot generalization. Using NaturalProofs, we benchmark strong neural methods on mathematical reference retrieval and generation tasks which test a system's ability to determine key results that appear in a proof. Large-scale sequence models show promise compared to classical information retrieval methods, yet their performance and out-of-domain generalization leave substantial room for improvement. NaturalProofs opens many avenues for research on challenging mathematical tasks.
BERT (P/S) + joint tops the leaderboard with 32.5
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The paper reports no overall retrieval score, so this is the plain mean of ProofWiki and Stacks test mAP. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders.
- 1 BERT (P/S) + joint 32.5
- 2 BERT (P+S) + joint 29.8
- 3 BERT (P/S) + pair 18.9
- 4 BERT (P+S) + pair 16.1
- 5 TF-IDF 9.91
Source: The NaturalProofs paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | BERT (P/S) + joint | 32.5 | Official leaderboard | ||
| 2 | BERT (P+S) + joint | 29.8 | Official leaderboard | ||
| 3 | BERT (P/S) + pair | 18.9 | Official leaderboard | ||
| 4 | BERT (P+S) + pair | 16.1 | Official leaderboard | ||
| 5 | TF-IDF | 9.91 | Official leaderboard |
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Math
- AIME 2026 In use Models solve the 30 integer-answer problems from the 2026 AIME I and II math competitions. Scored by exact answer match. GPT-5.5 100%
- IMO-AnswerBench In use 400 olympiad problems with short verifiable answers, robustified against memorization. Scored by answer match. GLM-5.2 91%
- HMMT 2026 In use Models solve the 33 final-answer problems from the February 2026 Harvard-MIT Mathematics Tournament. Scored by exact answer match. GPT-5.5 98.5%
- PolyMath In use Models solve 9,000 parallel math problems in 18 languages across four difficulty levels, from school math to frontier math. Final answers are checked.
Qwen3.7-Max 86.5%
- MathArena (Apex, IMO, USAMO) In use Evaluates models on newly released math competitions after release. Final-answer or human/LLM-judged proof scores. Claude Opus 4.8 81.3%
- MathArena Apex Shortlist In use Models solve 47 hard final-answer problems that MathArena selected from 2025 competitions, where the best models scored about 50% at selection. Scored by accuracy over repeated attempts. GPT-5.5 98.4%
Data audit
All 8 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
- Pass Independent runs Scores from the benchmark’s paper (5 entries).
- Pass Models scored 5 entries on its leaderboard, none a model we track.
- Note Newest score No scores yet.
- Note Frontier models No models scored yet.
- Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 5 entries at the last read, none a model we track.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.