win7win/rap-benchmark-data

RAP: Research Attention Prediction in Language Agents

Paper · Evaluation code

RAP evaluates whether language agents can infer changes in research attention from literature search. Research attention is a direction's share of arXiv papers within its field. The benchmark combines two tasks with two temporal settings:

• Trend: Did a direction's share increase, or will it increase?
• Comparison: Which candidate direction gained, or will gain, the most share?
• State assesses a completed change; Forecast predicts a future change.

Benchmark

	Core	Extended
Research fields	54	269
Main evaluation questions	2,160	10,740
Trend questions per temporal setting	810	4,025
Comparison questions per temporal setting	270	1,345
Target periods	2024 H1–2026 H1	2024 H1–2026 H1

The environment has 2,187 benchmark directions, with 2–15 directions per field. Each field contributes up to three fixed Trend directions and one Comparison group of those directions. Questions repeat over five half-year periods. Forecast at T and State at T + six months address the same realised change. The Extended set includes the Core fields.

The complete inventory contains 11,814 questions, including auxiliary State questions for analyses at a shared cutoff. The dataset viewer shows this full inventory. Use scopes.json to select the intended evaluation scope.

Files

Path	Content
questions.jsonl	Questions, field IDs, cutoffs, time windows and selected directions
scopes.json	Fixed question IDs for Core, Extended and auxiliary scopes
codebooks/	English direction names and scopes for 269 fields
references/answers.jsonl	Evaluator-side acceptable answers and scoring eligibility
corpora/	Full field search corpora with titles and abstracts
corpus_manifest.json	Corpus paths, record counts and checksums
dataset.json	Dataset identity, protocol identity and data-file checksums
SHA256SUMS	Checksums for the published files

Direction IDs in the data refer to the field directory. The evaluator renders only selected definitions and assigns question-local D1–D3 identifiers for Comparison. Reference answers are read by the scorer, not the inference runner.

Reference answers

Each paper with confirmed direction memberships contributes unit mass, divided equally among its distinct confirmed directions. Direction shares are normalised over covered papers in the complete field directory. Candidate selection does not change that denominator.

Trend compares a direction's later share with its earlier share. Exact no-change events are unscored. Comparison selects the largest signed percentage-point change; when all candidates decline, the smallest decline wins. Tied reference winners are all accepted for accuracy. Brier scoring uses unique references.

Models return {"p_gain": p} for Trend or a probability for each displayed candidate for Comparison. These are probabilities of answer events, not paper shares. See the evaluation protocol for rendering, validation and scoring.

Search corpus

The benchmark was constructed from 356,357 arXiv papers first submitted between June 2022 and June 2026. This release provides the field corpora used for evaluation. Fields can overlap, so a paper can occur in more than one corpus.

Each corpus record contains id, date, title, abstract and categories. The search runner enforces the question's field and cutoff before constructing its BM25 index. Search returns full titles and abstracts from papers strictly before the cutoff. The corpora retain all dates needed by the rolling tasks; the per-question cutoff is applied by the evaluator.

Download and evaluate

Download the complete snapshot:

pip install -U huggingface_hub
hf download win7win/rap-benchmark-data --repo-type dataset \
  --revision attention-change-2026-11-01 --local-dir RAP_dataset

From the evaluation-code repository:

export RAP_DATA=/path/to/RAP_dataset
python -m rap_eval.dataset --data "$RAP_DATA" --corpora

File integrity can also be checked from the dataset root:

sha256sum -c SHA256SUMS

The release identifier in dataset.json is rap-state-forecast-change. The earlier fixed-eight-direction benchmark is preserved at revision legacy-eight-directions; it uses a different evaluation protocol.

License and citation

Original RAP-derived artifacts are released under CC BY 4.0. The descriptive arXiv metadata retain their source terms. See LICENSE and ARXIV_METADATA_NOTICE.md.

@misc{wu2026rap,
  title = {RAP: Research Attention Prediction in Language Agents},
  author = {Wu, Yingqian and Liang, Jingcong and Yue, Shengbin and Wang, Siyuan
            and Yin, Zhenfei and Yu, Junchi and Wei, Zhongyu},
  year = {2026},
  eprint = {2609.10092},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.10092}
}