---
license: apache-2.0
language:
- en
task_categories:
- question-answering
tags:
- mathematics
- benchmark
- evaluation
- benchfill
size_categories:
- n<1K
configs:
- config_name: AIME24
  data_files:
  - split: test
    path: AIME24.jsonl
- config_name: AIME25
  data_files:
  - split: test
    path: AIME25.jsonl
- config_name: AIME26
  data_files:
  - split: test
    path: AIME26.jsonl
- config_name: AMC23
  data_files:
  - split: test
    path: AMC23.jsonl
---

# BenchFill

Four augmented mathematics evaluation benchmarks from **Your Benchmark Is an Empirical Measure Over Difficulty** (NeurIPS 2026), by Yifan Sun, Naicheng Yu, Jiawen Gong, Jingyan Shen, and Huan Zhang.

BenchFill generates new items to fill gaps in a benchmark's difficulty distribution. This release contains the original benchmark items followed by the curated generated items.

- [Code and full artifact release](https://github.com/ASTRAL-Group/benchmark-difficulty-measure)
- [Project page](https://naicheng-yu.github.io/Benchfill-Website/)

## Why BenchFill?

**BenchFill helps detect training gains that existing benchmarks miss.** Filling gaps in benchmark difficulty makes localized improvements easier to observe. In the paper's SFT experiment, adding just one BenchFill question to each of three AIME benchmarks made training gains statistically detectable in all three cases; the original benchmarks and baseline augmentations detected none ([Table 12](https://naicheng-yu.github.io/Benchfill-Website/static/pdf/benchfill_paper.pdf#page=33)).

![How BenchFill works: training gains in a difficulty gap may be missed by the original benchmark; adding questions in that gap makes the gains easier to detect.](benchfill-overview.png)

## Subsets

All four configurations use the `test` split.

| Configuration | Original items | Generated items | Total | Original JSON | Hugging Face JSONL |
|---|---:|---:|---:|---|---|
| AIME24 | 30 | 20 | 50 | [AIME24.json](AIME24.json) | [AIME24.jsonl](AIME24.jsonl) |
| AIME25 | 30 | 22 | 52 | [AIME25.json](AIME25.json) | [AIME25.jsonl](AIME25.jsonl) |
| AIME26 | 30 | 23 | 53 | [AIME26.json](AIME26.json) | [AIME26.jsonl](AIME26.jsonl) |
| AMC23 | 40 | 40 | 80 | [AMC23.json](AMC23.json) | [AMC23.jsonl](AMC23.jsonl) |
| **Total** | **130** | **105** | **235** | | |

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("Yifan99/BenchFill", "AIME24", split="test")
# Other configurations: "AIME25", "AIME26", "AMC23"

original = dataset.filter(lambda row: "-origin-" in row["id"])
generated = dataset.filter(lambda row: "-fill-" in row["id"])
```

## Data format

Each record has three fields:

- `id`: question ID.
- `question`: mathematics problem text.
- `answer`: reference answer.

IDs use `-origin-` for original questions and `-fill-` for generated questions. Both JSON and JSONL files are provided; `load_dataset` uses JSONL.

## Source attribution and terms

The repository's existing Apache-2.0 metadata is retained. Original AIME and AMC competition materials are from the Mathematical Association of America (MAA) and retain their original source terms; the repository metadata does not relicense those materials.

See the upstream [THIRD_PARTY_NOTICES.md](https://github.com/ASTRAL-Group/benchmark-difficulty-measure/blob/53cf20cdb98dc9c2a4f780e582f00c559fb9e165/THIRD_PARTY_NOTICES.md), including its AMC23 and AIME24/25/26 attribution and [MAA educational-sharing notice](https://maa.org/wp-content/uploads/2024/08/2022-AMC-10A-solutions.pdf). Source notices and permissions apply to the corresponding source content.

## Citation

```bibtex
@misc{sun2026benchmarkempiricalmeasuredifficulty,
  title = {Your Benchmark Is an Empirical Measure Over Difficulty},
  author = {Yifan Sun and Naicheng Yu and Jiawen Gong and Jingyan Shen and Huan Zhang},
  year = {2026},
  eprint = {2610.29927},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url = {https://arxiv.org/abs/2610.29927}
}
```