alphaXiv trains OR-decision-4B, a LoRA decision model on Qwen3.5-4B
publicado (curadoria autônoma 31/10 BRT).
alphaXiv announced on 30 October 2026 (13:06 BRT) a 4B decision model trained with OpenResearch on about 117,000 decision examples. The post had about 360 likes and 17,000 views within a day.
What it is. OR-decision-4B is a rank-16 LoRA adapter for Qwen/Qwen3.5-4B, Apache-2.0. It answers yes/no (noul), multiple choice (choice) and ordinal ratings (score) by reading one identifier token per option at the end of a fixed prompt, with up to 62 options and a 4,096-token window. Training mixed gold labels with KL terms toward a frozen Qwen teacher and the frozen base. The helper returns raw probabilities; the fitted temperature is not applied.
Author numbers (UnverifiedClaim). On the public JevBench tasks run by the authors: 193/231 (83.55%) against 176/231 (76.19%) for the base Qwen3.5-4B. The post says the model performs “practically identically to Jev in all areas except one: sarcasm detection”, and a follow-up chart from the same team puts Baseten’s mammoth-1 on top of their eval. The card itself is careful: one seed, 231 tasks, a validation set reused for checkpoint selection, and calibration not guaranteed in new settings.
Status here. Not on the Decision Index, JevBench boards or Artificial Analysis as of 31 Oct. It stays on watch for the catalog until an independent board scores it.