---
license: other
task_categories:
- text-classification
- text-generation
language:
- en
tags:
- peer-review
- scientific-papers
- iclr
- llm-evaluation
pretty_name: DeepReview-Bench
size_categories:
- 1K<n<10K
---

# DeepReview-Bench

A benchmark for evaluating automatic (LLM) paper reviewers against human peer
reviews. Each item is an ICLR submission paired with its human reviews; the
ground truth for a paper is aggregated from its 3–4 human reviewers (rating
1–10; Soundness / Presentation / Contribution each 1–4; Accept/Reject decision).

## Contents

`papers/<paper_id>/`:

| file | description |
|------|-------------|
| `paper.pdf` | original submission PDF |
| `paper.md` | cleaned Markdown conversion of the PDF (math `$...$` and tables preserved) |
| `review.json` | human reviews: per-reviewer rating, soundness, presentation, contribution, strengths, weaknesses, etc. |
| `metadata.json` | title, venue, decision, dates |
| `figures/` | figures extracted from the paper |
| `conversion_report.json` | (subset) automatic PDF→Markdown fidelity check |

Top-level:

- `splits/` — evaluation splits: `deepreview13k_test_dev.jsonl`,
  `..._dev_clean100.jsonl` (version-consistency-filtered dev set),
  `..._lite.jsonl`, `..._test.jsonl`, `all.jsonl`.
- `mismatch_flags.jsonl` — per-paper flag for **version mismatch**: whether the
  human reviews describe an earlier paper version than the shipped `paper.md`
  (added/revised content after rebuttal). Tags: `mismatch` / `clean` /
  `unaudited`.
- `markdown_manifest.jsonl`, `pdf_manifest.jsonl`, `markdown_clean_manifest.jsonl`
  — provenance manifests.

## Known caveats

- **Version mismatch.** A meaningful fraction of papers have human reviews that
  describe an earlier arXiv/OpenReview version than the `paper.md` shipped here
  (rebuttal-revised content). `splits/..._dev_clean100.jsonl` is a dev subset
  selected to be version-consistent, but its retained papers were audited
  heuristically and **a small number of residual mismatches remain** — treat
  `mismatch_flags.jsonl` as the audited signal and re-verify if you need a
  perfectly clean subset.
- **Copyright.** PDFs are the authors' original submissions and remain under
  their respective rights; redistributed here for research evaluation only.

## License

Set the appropriate license before relying on this card. The papers are
third-party academic works; the benchmark packaging is released for research use.

## Citation

<!-- TODO: add the DeepReview-Bench / source citation. -->