Paper-Reviewing-Exam

Harbor tasks for collecting peer reviews of scientific manuscripts. Each task hands an agent a complete paper and asks it to write a review; the review is archived for human experts to assess afterwards.

These tasks do not score reviews. The verifier checks only that review.md was written and is substantive. There is no rubric and no LLM judge, because deciding what counts as a good review is the judgement the human experts are for.

• 31 tasks from 27 papers
• Fields: not declared (specs/.yaml is optional)
• Each version of a paper is an independent task: a later version fixes the earlier one's problems, so a review of one is not a review of the other.

Running

Harbor reads this repository directly:

harbor run --repo https://huggingface.co/datasets/Jack-Jieke-Wu/Paper-Reviewing-Exam/tree/main/paper-review-exam -a opencode

Tasks declare network_mode = "no-network". Passing --allow-agent-host promotes the policy to an allowlist, so the same task serves closed-book and open-book runs:

harbor run --repo https://huggingface.co/datasets/Jack-Jieke-Wu/Paper-Reviewing-Exam/tree/main/paper-review-exam -a opencode \
  --allow-agent-host arxiv.org \
  --allow-agent-host api.semanticscholar.org

Whether a run had network access is a property of the invocation, not of the task. Record it alongside the review, or a later reader cannot tell whether an agent that checked no citations was unable to or merely did not.

Running the benchmark

To run one reproducible model condition over the six-task Issue #19 set and archive every trail, use pre-harbor benchmark from the paper-review-bench repository (see its docs/BENCHMARK.md). Trails land in the Paper-Reviewing-Exam-Trails dataset.

Layout

paper-review-exam/
├── dataset-manifest.jsonl      one row per task: provenance and source hash
└── <task-id>/
    ├── task.toml
    ├── instruction.md
    ├── environment/            Dockerfile + the manuscript
    ├── solution/solve.sh       placeholder-review oracle
    └── tests/                  submission-contract check

Contamination

The writing-time review trail and later revisions of each paper are excluded from every task environment and checked for at build time, because a review written by an agent that read them looks exactly like a very good review and nothing downstream can tell.

Two paths remain open and are not fixable inside a task:

1. Multi-version papers are public on arXiv. An agent reviewing v1 with arxiv.org reachable can fetch v2 and read what was fixed. Consider running version-set papers closed-book, or recording that the risk applies.
2. This repository also hosts the raw sources under papers/, including the writing-time trail for the solution-* manuscripts. huggingface.co is not in the suggested allowlist, so a task run cannot reach it, but do not add it.

Provenance

Built by pre-harbor from the LaTeX sources in papers/. dataset-manifest.jsonl records each task's source SHA-256, so a collected review can be traced to the exact bytes it was written from.

Benchmark canary GUIDs are embedded in every task. This data should not appear in training corpora.