A new user turn helps redirect failed reasoning

When a model’s reasoning goes wrong, must correction start over, or can it keep useful earlier steps and be redirected? Answering that matters to whether a model can recover from a failed attempt without retraining its main reasoner. UPTURN tests a specific way to choose a restart point and redirect the next response.

UPTURN learns where to restart a failed answer, keeps its earlier reasoning, and redirects the frozen model with an instruction in a new user turn.

Across four frozen reasoners and six math tests, researchers compared UPTURN with a baseline that generated 32 full answers per question. UPTURN generated 16 initial answers; a separate gate estimates whether each answer is correct, and responses below its fixed threshold were corrected once, for up to 32 generations.

Mean per-question accuracy, averaged over reasoners and tests, was 55.3% with UPTURN and 37.1% with Base.

The result supports the full correction procedure in these evaluations: retaining part of a response and redirecting its continuation can outperform generating more complete responses. The comparison does not isolate the effect of the learned cut, the instruction, or the new user turn individually.

The main comparison covers four frozen reasoners and six math tests, and its protocols allow up to 32 generations per question. Base receives a larger maximum generated-token allowance, so the result does not establish equal wall-clock time or isolate the effect of any single part of UPTURN. The study covers automatically checkable reasoning tasks and one minimal intervention.