UPTURN redirects LLM reasoning with a new user turn instead of fine-tuning
A 130K-parameter learned probe finds where a failed reasoning chain went wrong, and a new user turn — not an inline token — steers the model back, its authors report.
In plain words
Large language models often fail to spot and fix mistakes in their own reasoning. UPTURN is a correction method from researchers behind a new arXiv paper that adds no training of the model itself: a small learned component picks the point in a failed answer where things went wrong, and the model is asked to redo the rest after receiving an instruction as a fresh user message. The authors say this beats comparable self-correction methods on math tests.
What was announced
The paper's core idea is to intervene at a learned position rather than retrain a model to critique itself. The authors train a 130K-parameter MLP head on a frozen encoder, using only final-answer correctness as the signal, to decide where in a failed rollout to step in. Their tests showed that inline injections under a fixed generation budget did not reliably change a model's course, while presenting the same instruction as a new user turn did.
The resulting framework, UPTURN (User-Prompt TURN injection at a learned position), requires no LLM fine-tuning: it keeps a selected prefix of the reasoning, injects an instruction as a new user turn, and regenerates the remainder. Across four reasoners and six math benchmarks, the authors report that UPTURN improved average accuracy in every setting over a 32-rollout baseline, outperformed self-correction and reflection-token injection baselines under matched correction budgets, transferred to GPQA-Diamond and across reasoners without retraining, and complemented GRPO and self-correction methods.
What it does not show
- Evidence is limited to the paper's abstract; detailed results, method specifics and peer-review status cannot be verified from the supplied excerpt.
- Reported accuracy gains and baseline comparisons are the authors' own claims and have not been independently confirmed.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.