← News

ResearchSources: arXiv · 20 Nov 2026

Researchers train LLMs on QOC, a structured alternative to Chain-of-Thought

An arXiv paper reports that fine-tuning models to follow a Question–Option–Criteria decision process improved answer quality over base models, with gains concentrated in larger models.

In plain words

Many language models reason by writing out step-by-step thoughts, a technique called Chain-of-Thought. This paper instead trains models to break decisions into explicit questions, competing options, and evaluation criteria. The authors report this structured approach improved answer quality, mainly in larger models.

What was announced

The paper, posted on arXiv, asks whether the internal thinking behavior of a large language model can be deliberately changed through training, and whether such a change helps. The authors replace Chain-of-Thought prompting with Question–Option–Criteria (QOC), a decision process that decomposes a task into explicit questions, competing options, and evaluation criteria. They describe a three-generation development pipeline, moving from a basic QOC decision core to richer answer synthesis and then to statistically enriched QOC reasoning.

Evaluations covered both internal consistency and output quality. Across the reported experiments, QOC-based fine-tuning improved answer quality relative to non-fine-tuned base models, with particularly strong gains once models were large enough to run the underlying QOC computations reliably. The authors conclude that the internal thinking model of an LLM is trainable and suggest structured alternatives to Chain-of-Thought could lead to more reliable, interpretable, task-adapted decision support. These are the authors' own reported results and have not been independently confirmed.

What it does not show

  • The source is only the paper's abstract; no full results, benchmarks, model sizes, or task details are available to verify the claims.
  • No specific models, datasets, or numerical results are named in the supplied passages, so the size threshold for strong gains cannot be stated.

Sources

Source-support review completed 21 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 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.