stanjsx/llm-review-panels

Dataset: Second Opinions or Echoes? Measuring Shared Errors Across LLM Peer Review Stages

Panels of LLM reviewers on 1,404 submissions to six machine learning venues, through independent review, rebuttal, discussion, and the area chair's decision: model-call records for the experiments reported in the paper, the coverage judge's labels, and the human annotation. The code that runs the panels and reproduces the paper's results reads these files from the repository root.

Venues and panels

Folder	Venues	Papers	Human reviews
iclr2021	ICLR 2021	296	before ChatGPT's release
iclr2022	ICLR 2022	225	before ChatGPT's release
neurips2022	NeurIPS 2022	183	before ChatGPT's release
2023-25	ICLR 2025, NeurIPS 2023, ICML 2025	700	after

The eight-family panel: GPT-5.6 Terra, Gemini 3.1 Pro, Claude Sonnet 5, DeepSeek V4 Flash, GLM-5.2, Qwen3.8, Grok 4.7, and Kimi K3. The four-family panel of the Families comparison: GPT-5.1, GPT-5.6 Luna, GPT-5.6 Terra, Gemini 3.1 Pro, Gemini 3.5 Flash, Gemini 3.5 Flash Lite, DeepSeek V4 Flash, GLM-5.2. Model keys in the records follow configs/models.yaml in the code.

Files

Path	Contents
data/manifests/<group>.json	One entry per paper: OpenReview id (paper_id), venue (venue_key), title, abstract, outcome (accept, reject, withdraw), human ratings (human_ratings, human_mean), and per review its OpenReview note id, anonymous reviewer id, and rating.
data/manifests/author_response_notes.json	For 180 of the 2023-25 papers, the OpenReview ids of the author notes given to the reviewers as the author responses (the authors' first round of responses).
results/first_reviews/<model>/<venue>/reviews.jsonl	First-review call logs (call_type: review), including failed attempts and retries (run_idx: three runs on ICLR 2022; on the other venues two for the eight-family panel, and one for the four models only in the four-family panel on ICLR 2021 and NeurIPS 2022): review items (strengths and weaknesses, each with its aspect in dimension), reasoning_text, emitted_score on the venue's scale, extra (summary, confidence, whether the model says it has seen the paper), raw_response, usage (tokens, cost). run_meta.json gives the venue's rating scale. Both panels on the 2021-22 venues; the eight-family panel on 2023-25.
results/<group>/scores.csv	2021-22: first-review scores on [0, 1] (score) with the human mean rating.
results/2023-25/four_family_scores.csv	The four-family panel on 2023-25: first-review scores (call_type review) and scores from the human reviews with ratings removed (score_from_human_reviews).
results/<group>/content/	human_points.jsonl: the points extracted from the human reviews (id, polarity, category, kind, text, and the quoted sources). coverage_labels.csv: for each first review (model, run) and point (id), whether the review raises it (recognition: specific, partial, absent) and its stance. human_coverage.csv: human committees' coverage of the same points with the same judge.
results/<group>/discussion/	records.jsonl: revisions of the first reviews (arm, table below), each with the revised review, emitted_score, first_review_score, and for the discussion conditions the other reviewers in the order shown (peers). records_rebuttal.jsonl: revisions after author responses. coverage_labels.csv: the coverage judge on the revisions; on 2023-25 against human_points.jsonl in this folder (332 papers). iclr2022/discussion_run2/, discussion_run3/: two repeated runs.
results/<group>/notes/	records.jsonl: revisions and fresh reviews after three notes (notes); design.json: the notes given per paper; adoption_labels.csv: how many of the notes a review raises (mentioned of n_notes), for revisions and their first reviews.
results/<group>/roles/	records.jsonl: first reviews with an assigned focus (role); assignment.json: the role per paper and model; coverage_labels.csv.
results/<group>/area_chair/	records.jsonl: the decisions (decision) and meta-reviews of nine LLM area chairs (ac_model_id) reading the human reviews or the panel's first reviews (arm), with the venue's decision (venue_decision). iclr2022/area_chair/records_run2.jsonl: a repeated run.
results/annotation/	Human annotations of 300 weakness and review pairs. items.csv: what the annotators saw; key.csv: each item's source, condition, and judge label; labels.csv: final recognition labels (label_A, label_B) and B's stance (stance_B). A's final labels follow a review of disagreements with the judge and B. initial_labels.csv: the independent recognition labels used for initial agreement and note uptake. Both label files join to the items by item.
results/recall_probe/<group>/	Whether a model recalls a paper's decision and rating from its abstract (arm: title_abstract, abstract_only, predict).
results/self_review/	The eight-family panel reviewing a draft of the paper itself (ICLR 2025 setup): first reviews, revisions after discussion, area-chair decisions, and the judge's labels for the weaknesses the reviews share.
MANIFEST.tsv	Data and result files with their sizes and SHA-256 hashes.

Conditions (field arm)

arm	Condition
plain_revision	The reviewer re-reads the paper and its own first review (control)
discussion	Plus the other seven first reviews with their scores
hidden_scores	Plus the other reviews without their scores
hidden_verdicts	Hidden scores, with sentences stating a verdict removed
independence	Discussion plus an instruction to check the other reviews against the paper rather than follow them
rebuttal	Plus the authors' real responses
reply_to_review	Plus a reply to the reviewer's own review, written by GPT-5.1 as the authors (ICLR 2022)
missed, already_raised, other_paper	Notes: weaknesses the human reviewers raised and the panel missed; weaknesses some panel review had raised; another paper's weaknesses
fresh_missed, fresh_other_paper	The same notes given before a fresh review instead of a revision
human_reviews, panel_reviews	Area chairs reading the human reviews or the panel's first reviews
first_review	The first review itself (coverage labels)

Calls (field call_type): review is a first review; score_from_human_reviews scores the paper from the human reviews with ratings removed.

Selecting first reviews

The call logs can contain more than one attempt per paper, model, and run. analysis/common.py:first_reviews selects call_type: review records with parse_ok: true and nonempty items for the requested run_idx, keeping the last qualifying record in file order for each paper and model. Its emitted_score is rescaled using the venue's rating scale. analysis/common.py:score_table uses the supplied scores.csv files for the 2021-22 venues, preserving the sample used for those score tables in the paper. Other analyses apply their own matching requirements.

Legacy per-item rubric calls (call_type: B) from an earlier workflow are excluded; they do not contribute to the paper's reported results.

Not included

Complete paper texts, complete human review documents, and the rendered prompts that contain them. Attributed review excerpts are included in sources[].quote to support the extracted points. The papers and reviews are public on OpenReview under the ids in data/manifests/.

License

The dataset is released under CC BY 4.0. The review excerpts come from OpenReview, where reviews and comments are released under CC BY 4.0; the manifests give each paper's OpenReview id and the note ids of its reviews. The code is released under the MIT License (llm-review-panels).