# CoRP preprint reports model gains with one inference pass

_Published Monday, November 9, 2026 at 12:21 PM EST · AI, Science · Latest · Tier 2 — Notable_

Researchers report a method that folds rewarded model variations into a single language model, avoiding the multiple inference passes required by an ensemble. In a preprint, they say CoRP improved base models by 8.1 points on average across five models from 0.5B to 8B and five tasks covering math, code and creative writing.

The method combines reward-weighted aggregation, compatibility-aware reweighting and a held-out validation gate without passing gradients through the language model. Using one tenth of RandOpt's perturbation budget, it exceeded single-inference RandOpt by 6.5 points. It recovered more than half the gain of a 50-pass majority-vote ensemble while requiring one forward pass per test example.

## Sources

- [cs.CL updates on arXiv.org](https://arxiv.org/abs/2605.31494)

---
Canonical: https://techandbusiness.org/newswire/ojFHQyiHbjss-Guq98SMmS
Published: 2026-11-09T17:21:18.492Z
Story chronology: 2026-11-09T05:00:00.000Z
Retrieved: 2026-11-10T07:40:04.189Z
Publisher: Tech & Business (techandbusiness.org)