PaperScorer-Vision-ICLR-27-5k-sample

A random sample of 5,000 ICLR 2027 submissions, in the same row format as skonan/PaperScorer-Vision (openreview-iclr config), scored by PaperScorer-7B-Vision-OpenReview-ICLR before decisions were released. 4,974 papers were scored; 26 could not be (16 had fewer than 6 body pages, 10 were not anonymized) and are not included.

The predictions are model outputs, not ICLR decisions. label and the outcome fields in metadata are empty until decisions are out; this dataset is meant for checking the model's predictions against the real outcomes later.

from datasets import load_dataset
ds = load_dataset("skonan/PaperScorer-Vision-ICLR-27-5k-sample", "openreview-iclr", split="papers")

Schema

The first seven columns are exactly those of PaperScorer-Vision; four columns hold the model's prediction.

field	type	description
paper_id	string	OpenReview submission id
title	string	paper title
content	string	prompt-stripped body (title + abstract + <image> placeholders, one per page)
metadata	string	JSON with the PaperScorer-Vision keys: submission_id, answer, year (2027), ratings, decision, pct_rating, conference (iclr), authors ([], anonymous), citation, citation_normalized_by_year; outcome fields are null
label	string	null until decisions are released
references	list<list<string>>	[["iclr2027_sample5k", "test"]]
images	list<struct<bytes,path>>	one PNG per page the model saw, bytes inline
predicted_decision	string	Accept when logit >= 0, else Reject (exact ties count as Accept)
p_accept	float64	estimated chance of acceptance at ICLR's base rate (see below)
logit	float64	log P(Accept) − log P(Reject) at the model's decision token
rank	int64	1 = highest logit (ties broken by paper_id)

extras.jsonl lists, per paper_id, the OpenReview number, status and title at download time, body pages, page images, the PDF's SHA-256, the task prompt and the service deploy that scored it.

Calibration

The model was trained on a balanced (50/50) accept/reject mix, so its log-odds logit measure how accept-like a paper is under even odds. p_accept converts them to a chance of acceptance at ICLR's real rate:

p_accept = sigmoid(logit / 1.251 + log(0.314 / 0.686))

The temperature 1.251 was fit on the ICLR natural-rate validation split (31.4% accepted) with the base-rate term inside the fit. This only changes the reported probability, never the decision: predicted_decision is Accept exactly when p_accept >= 0.314. On the held-out ICLR 2025-26 natural-rate test set, this calibration has an expected calibration error of 0.023 for the acceptance probability. Across this sample the mean p_accept is 30.6%, in line with ICLR's ~30% acceptance rate. The decision rule targets balanced accept and reject recall rather than the base rate, so it labels 50.4% of papers Accept. Many Accept calls will be rejections (on the held-out ICLR natural-rate test, 47% of Accept calls were accepted), and p_accept is the better guide to a paper's actual chances.

Caveats

• The sample was drawn uniformly from all submissions listed on OpenReview at download time. It includes papers later desk-rejected (52) or withdrawn (25), with more withdrawals possible (openreview_status in extras.jsonl). Exclude them before comparing with decisions, since ICLR's acceptance rate is computed over reviewed papers.
• Page images are the PDF as submitted, rendered by the same pipeline as PaperScorer-Vision. The model saw at most the pages listed in images.

License & citation

License: see the PaperScorer collection. Submissions remain the property of their authors.

@inproceedings{konan2026paperscorer,
  title     = {PaperScorer: How Predictable Is Paper Acceptance?},
  author    = {Konan, Sachin and Liu, Jonathan and Liu, Zhuang},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}