Princeton's PaperScorer fine-tunes Qwen to predict paper acceptance, hitting 77.5% accuracy

IanArawjo · x · 2026-10-27

Princeton researchers Sachin Konan, Jonathan Liu, and Zhuang Liu released PaperScorer, a NeurIPS '26 project that predicts paper acceptance from anonymized submissions, plus a web service: upload an anonymized PDF and receive a calibrated probability labeled Leans Reject, Leans Accept, or Likely Accept.

Method:

Data and results:

Notable validation cases: Mamba, rejected at ICLR 2024, gets a 71% score (2.3× the 31% ICLR base rate) — arguably it deserved acceptance; the early version of A Decade's Battle on Dataset Bias is correctly predicted as a reject, while its later accepted ICLR version is correctly predicted as an accept. Two of the showcased well-known good papers were actually rejected.

Limits: 5 papers per mailbox per day, first 15 pages only, results may be delayed by queue load.

Original post →

More from Models

Models channel →