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

2026-06-15 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A new semi-supervised framework scales Large Language Model (LLM) reasoning from minimal supervision by transforming reasoning verification into a data creation mechanism. This approach trains a lightweight reasoning-correctness classifier using only a few labeled samples to determine the validity of intermediate reasoning traces generated by an LLM. An entropy-based confidence threshold then filters out unreliable samples, with the remaining high-confidence reasoning traces used to fine-tune the model. Experiments on Verifiable Math Problems (Orca-Math subset) and Question Answering on Image Scene Graphs (GQA) with Visual Programming demonstrate that this method achieves accuracy comparable to systems utilizing 10-15x more labeled data. Ablation analyses confirm the critical roles of both the classifier and entropy filtering in enabling scalable, noise-resistant pseudo-labeling. This framework offers a practical path for constructing large-scale reasoning resources, moving towards autonomous reasoning systems with minimal human input.

Key takeaway

For Machine Learning Engineers developing LLMs with scarce labeled data, consider implementing a semi-supervised framework that leverages lightweight reasoning verifiers. This approach allows you to achieve high reasoning accuracy comparable to 10-15x more data by turning verification into a data creation mechanism. You can significantly reduce expensive answer-level supervision, accelerating the development of robust reasoning capabilities and paving the way for more autonomous systems.

Key insights

A semi-supervised framework scales LLM reasoning by using a lightweight verifier to generate high-confidence pseudo-labels from minimal human input.

Principles

• Reasoning verification can create data.
• Entropy filtering enhances pseudo-labeling.
• Lightweight verifiers reduce supervision cost.

Method

Train a lightweight classifier on few labels to verify LLM reasoning traces. Filter high-confidence traces using an entropy threshold. Fine-tune the LLM with these pseudo-labels.

In practice

• Construct large-scale reasoning resources.
• Develop autonomous reasoning systems.
• Reduce data labeling costs for LLM training.

Topics

• Large Language Models
• Semi-Supervised Learning
• Reasoning Verification
• Pseudo-Labeling
• Data Efficiency
• Orca-Math

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language Models — Verifiable reasoning can be compressed into compact models, achieving frontier performance.
• Unified Data Selection for LLM Reasoning — High-Entropy Sum (HES) is a training-free metric for quantifying LLM reasoning quality via token entropy.
• Power-SMC: Low-Latency Sequence-Level Power Sampling for Training-Free LLM Reasoning — Power-SMC enables efficient sequence-level power sampling for LLM reasoning, significantly reducing latency compared to prior methods.
• Finding the Minimal Parameter Budget for Implicit Reasoning: A Data Complexity Driven Scaling Law for Language Models — Overparameterization in LMs can degrade reasoning performance on knowledge graphs due to excessive memorization, exhibiting a U-shaped scaling curve.
• Scaling Behaviors of LLM Reinforcement Learning Post-Training: An Empirical Study in Mathematical Reasoning — Larger LLMs demonstrate superior efficiency and performance in RL post-training for mathematical reasoning.
• Reasoning-Based Refinement of Unsupervised Text Clusters with LLMs — LLMs can validate and refine unsupervised text clusters by acting as semantic judges, improving coherence and interpretability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.