DATA_CARD.md

RAP Benchmark Dataset Card

Task

RAP evaluates whether an agent can estimate the distribution of newly posted papers across eight operational research directions and forecast the distribution in the following six months. Each episode uses forecast_date=T as its boundary. The history windows describe the available past periods and the future window is [T, T+6 months).

The State task estimates the distribution observed in the most recent six-month window. The Forecast task predicts the distribution in the next six-month window.

Package contents

• inputs/ contains the model-facing Dev/Test episode inputs.
• targets/ contains evaluator-only recent and future counts, shares, and diagnostics.
• evaluator/ contains combined episode records for scoring and offline baselines.
• data/ contains the frozen evaluation scope and Search-manifest indexes.
• search_shards/ contains the metadata-only Search snapshot.

Each row in inputs/ contains the field, cutoff date, time-window metadata, and eight direction definitions. It does not contain history counts, recent shares, future counts, or future shares. Those evaluator-only values are stored in targets/ and evaluator/.

Splits and evaluation

Dev and Test are frozen field-level splits rather than random episode-level splits. Dev contains 335 episodes from 67 fields; Test contains 1,055 episodes from 211 fields. The evaluation-code repository provides the prompt protocol, scorer, contract checks, and offline statistical baselines.

Target shares are normalized across the eight directions, using primary-direction counts. Direction definitions and episode boundaries are provided in each input record. Use inputs/ for agent runs and targets/ or evaluator/ for scoring.

Search snapshot

search_shards/fixed/ enforces the window [T-6m, T), while search_shards/expanding/ enforces <T. Each shard contains exactly id, date, title, abstract, and categories. The shard manifests record SHA-256 digests, row counts, and date-bound checks.

The Search snapshot contains descriptive arXiv metadata only. See ARXIV_METADATA_NOTICE.md for provenance and source terms.

Integrity

From the dataset root, run:

sha256sum -c SHA256SUMS

License

This repository is marked other because it combines original RAP-derived artifacts with arXiv descriptive metadata. Read LICENSE and ARXIV_METADATA_NOTICE.md before redistribution or adaptation.

Citation

@misc{wu2026rap,
  title  = {RAP: Research Attention Prediction Reveals Target-Conditioned Evidence Acquisition Biases},
  author = {Wu, Yingqian and Liang, Jingcong and Wang, Siyuan and Yin, Zhenfei
            and Torr, Philip and Yu, Junchi and Wei, Zhongyu},
  year   = {2026},
  url    = {https://arxiv.org/abs/2609.10092}
}