# UPTURN preprint reports reasoning gains from user-turn corrections

_Published Monday, October 12, 2026 at 10:37 AM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers report in a preprint that UPTURN, a method for redirecting failed AI reasoning, improved average accuracy across four reasoning models and six math benchmarks over a 32-rollout baseline, without fine-tuning the language models.

The method trains a 130K-parameter neural network to select an intervention point using final-answer correctness. It retains the reasoning before that point, inserts an instruction as a new user turn, and regenerates the remainder. Under a fixed generation budget, the tested instructions inserted within existing reasoning did not reliably redirect it.

UPTURN also outperformed self-correction and reflection-token injection baselines under matched correction budgets and transferred to GPQA-Diamond without retraining.

## Sources

- [cs.CL updates on arXiv.org](https://arxiv.org/abs/2610.13672)

---
Canonical: https://techandbusiness.org/newswire/hLdwrZyAgTy3EjxjworpGu
Published: 2026-10-12T14:37:54.769Z
Story chronology: 2026-10-12T04:00:00.000Z
Retrieved: 2026-10-12T16:38:35.034Z
Publisher: Tech & Business (techandbusiness.org)