reviewsearch-skill gives your coding agent a search engine over 205,988 real peer reviews and
their author rebuttals — every ICLR / ICML / NeurIPS / COLM paper with a public author response. Ask how
a kind of paper gets criticized, find the reviewers who raised a specific concern, and read how the
authors who answered it phrased their reply.

It's also live in your browser.

🔍 What it searches

The query is a concern, not a keyword — e.g. "reviewers asking for a statistical significance test"
or "papers criticized for weak baselines" — and each result is a real review that raised it, carrying
the author's full rebuttal as its payload.

python skills/reviewsearch/scripts/search_reviews.py \
  "reviews asking for a statistical significance test" --top-k 5 --accepted-only

Each result has venue, year, title, decision, summary, concern (the reviewer's weaknesses +
questions), rebuttal (the author's response — every result has one), and a relevance score. Filter
with --accepted-only, --year-min, --year-max (the corpus spans 2023–2026).

😎 Why it's different

We did not prompt a model to imagine what reviewers say. We collected every ICLR / ICML / NeurIPS /
COLM review that has an author response, and trained our own hybrid retriever on them: a fine-tuned
dense encoder and a
sparse encoder, fused with Reciprocal Rank
Fusion (RRF), and served on a free CPU tier.

So when you ask how a concern was answered, you get the case where a real author faced the same
criticism — and their whole response, not a plausible-sounding guess.

The search engine is live: try it on the website.

📦 Installation

Claude Code

/plugin marketplace add yjoonjang/reviewsearch-skill
/plugin install reviewsearch

Codex

Tell Codex:

Fetch and follow instructions from https://raw.githubusercontent.com/yjoonjang/reviewsearch-skill/main/.codex/INSTALL.md

Then just ask, for example: "find reviews asking for a significance test on a retrieval paper."

📚 Attribution

Review data from OpenReview (ICLR / ICML / NeurIPS / COLM), licensed
CC-BY-4.0.

📄 License

MIT — see LICENSE.