Scaling LLM Reasoning from Minimal Labels: A Semi-Supervised Framework with a Lightweight Verifier

A new semi-supervised framework can train large language models to reason with far less human-labeled data by using a lightweight classifier to verify intermediate reasoning steps, according to a paper submitted in 2026 [1]. The method, detailed on arXiv, addresses a core bottleneck in developing large language models (LLMs): the need for vast quantities of correctly annotated answers to assess reasoning quality [1]. The researchers instead train a compact reasoning-correctness classifier on only a few labeled samples. This classifier judges whether the intermediate reasoning traces generated by an LLM are valid [1]. An entropy-based confidence threshold then filters out unreliable samples, and the remaining high-confidence reasoning traces are used to fine-tune the model [1]. The approach was tested on Verifiable Math Problems using an Orca-Math subset and on Question Answering on Image Scene Graphs (GQA) with Visual Programming [1]. The paper reports that the method achieves accuracy comparable to using 10-15 times more labeled data [1]. Ablation analyses confirmed that both the classifier and the entropy filtering step are essential for scalable and noise-resistant pseudo-labeling [1]. High-quality labeled training datasets for supervised and semi-supervised machine-learning algorithms are typically difficult and expensive to produce because of the large amount of time needed to label the data [4]. The new framework attempts to bypass this cost by replacing expensive answer-level supervision with lightweight reasoning verification [1]. The authors state this provides a practical path toward constructing large-scale reasoning resources and paves the way for future autonomous reasoning systems that learn from minimal human input [1]. The paper was submitted to arXiv on 15 June 2026 under the title "Scaling LLM Reasoning from Minimal Labels: A Semi-Supervised Framework with a Lightweight Verifier" [1]. The work falls within the broader effort to improve learning algorithms and reduce dependency on manually curated datasets, which remain an integral part of machine learning research [4].