INSAIT-Institute/ProofRank-outputs

ProofRank: Evaluation Outputs

Companion artifact to the NeurIPS 2026 (Evaluations and Datasets Track) paper "Not All Proofs Are Equal: Evaluating LLM Proof Quality Beyond Correctness".

This dataset contains the LLM-judge outputs behind every number reported in the paper, so that all results can be recomputed without re-querying any model. It covers the thirteen evaluated models on the 382 benchmark problems, across all five proof-quality metrics (conciseness, computational ease, cognitive simplicity, diversity, and adaptivity) plus the correctness verification.

The benchmark problems themselves (with the human solution summaries and requested techniques) are released separately as the ProofRank dataset.

Contents

Each configuration of this repository is one evaluation artifact:

Config(s)	Rows	Description
answer_checker*, completeness_checker*	~51.1k	Final-answer and completeness judgments (GPT-OSS-120B) for the main, diversity, adaptivity, and conciseness-prompting runs
verbosity_rephrase, verbosity_rephrase_2/3	9,936	Rephraser outputs used for the compressibility-based conciseness metric, for the default and the two concise prompts
verbosity_verifier	3,425	Validity judgments of the rephrasings
technique_verifier	5,691	Judgments of whether a solution follows the requested technique (adaptivity, Sec. 3.2)
summary_diversity_clustering_main, _with_human	1,525	Technique clusterings of the sampled solutions, without and with the human solutions
topic_classifier	382	Topic labels per problem
pairwise_judgments	~173.4k	Pairwise computational-ease / cognitive-simplicity verdicts, per evaluation setting
pairwise_raw	~49.3k	The underlying raw pairwise judge files, needed to recompute the Bradley-Terry ratings from scratch
correctness_raw	600	GPT-5.4 reference judgments and the corresponding checker judgments on the 200 sampled solutions used to validate the correctness filter
clustering_raw	755	Clusterings of the human solutions (and human + initial LLM solutions), used for diversity problem selection
correctness_labels	~24.6k	Derived per-solution correctness labels (informational; re-derived from the checker outputs during evaluation)
diversity_samples, human_solutions, all_solutions, raw_problems	~19.1k	The compiled solution samples and problem/solution data files consumed by the evaluation scripts

Note. The text of the human-written solutions in human_solutions and human_solutions_sample is not redistributed: each such field reads "Hidden due to the Art of Problem Solving terms of use". The LLM-written summaries of these solutions (core_idea, solution_summary, family, backbone, closing_engine) are still provided.

Steering and validation runs

These ship as raw judge trees rather than cached tables, because the ablation analysis scripts read outputs/ directly. The prompt-control runs cover eleven of the thirteen models - StepFun-3.5-Flash and Grok-4.1-Fast were not re-solved under the perturbed prompts.

Config	Rows	Description
anti_prompt_raw	~50.4k	Checker and rephraser outputs for the four prompt-control runs, which re-solve the benchmark under a prompt pushing against one metric
anti_prompt_pairwise_raw	~417.4k	Pairwise verdicts pitting each perturbed run against the unperturbed one
single_prompt_raw	~11.2k	Checker and clustering outputs for the single-prompt diversity variant, which asks for four methods in one call
validation_raw	2,991	Judge and metric validation: the GPT-5.4 adaptivity-verifier subset and the clustering-subset runs
diversity_baseline_raw	~15.8k	Baseline clustering and diversity-checker trees, as read directly by the analysis scripts
human_solutions_sample	7,244	Sampled human solutions used by the clustering-alignment check
technique_verifier_samples	100	The 100-item adaptivity subset the GPT-5.4 verifier is compared on

Usage

Restore the artifacts into data/ and outputs/ with the script provided in this repository, then run the evaluation as documented in the code release:

python scripts/data/restore_released_data.py --from-hub --repo-id INSAIT-Institute/ProofRank-outputs --overwrite
python scripts/results/eval_package.py --exclude-weak-baselines

Licensing

The problems are drawn from publicly available mathematics competitions and previously released benchmarks, and the human-written reference solutions are collected from public Art of Problem Solving posts and official competition materials. We abide by the sources' terms and conditions and redistribute this material under CC BY-NC-SA 4.0 - see LICENSE - for non-commercial research use and with proper credit to the original competitions.

The terms of the original sources continue to apply to the redistributed text, independently of that license. Namely, for official competition problems and previously released benchmarks, they typically in the public domain or released under comparable conditions, which likewise forbid use of the material for training.

Accordingly, this data must not be used to train any AI model. Users are responsible for complying with the original terms that apply to each entry.

Work produced by us - code, results, and evaluation scripts - is published under the more permissive MIT license.

Citation

@inproceedings{petrov2026proofrank,
  title={Not All Proofs Are Equal: Evaluating {LLM} Proof Quality Beyond Correctness},
  author={Petrov, Ivo and Dekoninck, Jasper and Dimitrov, Dimitar I. and Vechev, Martin},
  booktitle={The Fortieth Annual Conference on Neural Information Processing Systems Evaluations and Datasets Track},
  year={2026},
  url={https://openreview.net/forum?id=Zf9xKiTAom}
}