Revisiting Chain-of-Thought Reasoning under Limited Supervision: Semi-supervised Chain-of-Thought Learning

2026-07-01 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

The Semi-CoT framework introduces Semi-supervised Chain-of-Thought Learning, a novel approach that reuses generated reasoning traces as semi-supervised learning signals for large language models. Unlike traditional Chain-of-Thought (CoT) methods that primarily use reasoning chains as inference-time prompts, Semi-CoT extends the self-training paradigm by constructing pseudo reasoning supervision from unlabeled questions. The framework operates by sampling multiple pseudo-CoTs for each unlabeled question, then estimating answer-level semantic entropy to select low-entropy reasoning chains as reliable pseudo-CoT demonstrations. Pilot experiments across AQuA, SVAMP, GSM8K, and MultiArith datasets demonstrated that the entropy gate effectively selects high-precision pseudo-CoTs, achieving pseudo-answer precision ranging from 91.36% to 100%. While Semi-CoT yielded small performance gains on SVAMP and GSM8K, it showed negative transfer on AQuA and reached a performance ceiling on MultiArith, indicating a need for improved demonstration selection or student training.

Key takeaway

For Machine Learning Engineers aiming to enhance large language model reasoning with limited supervision, Semi-CoT presents a viable strategy to generate pseudo reasoning signals from unlabeled data. While the entropy gate effectively identifies high-precision pseudo-CoTs, you should prioritize stronger demonstration selection mechanisms or more robust student training approaches. This will help mitigate negative transfer observed on some datasets and ensure consistent performance gains across diverse tasks, moving beyond inference-time prompting.

Key insights

Semi-CoT reuses generated reasoning traces from unlabeled data as semi-supervised learning signals for Chain-of-Thought.

Principles

• Generated reasoning traces can serve as pseudo-supervision.
• Semantic entropy can gate pseudo-CoT reliability.
• Unlabeled data offers reliable reasoning signals.

Method

Semi-CoT samples multiple pseudo-CoTs for unlabeled questions, estimates answer-level semantic entropy, and selects low-entropy chains as reliable pseudo-CoT demonstrations.

In practice

• Apply entropy gating to filter pseudo-labels.
• Explore semi-supervised CoT for LLM fine-tuning.
• Test pseudo-CoT generation on diverse datasets.

Topics

• Chain-of-Thought Reasoning
• Semi-supervised Learning
• Large Language Models
• Pseudo-labeling
• Semantic Entropy
• Machine Learning

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

Related on AIssential

• Scaling LLM Reasoning from Minimal Labels: A Semi-Supervised Framework with a Lightweight Verifier — A semi-supervised framework scales LLM reasoning by using a lightweight verifier to generate high-confidence pseudo-labels from minimal human input.
• What Makes Effective Supervision in Latent Chain-of-Thought: An Information-Theoretic Analysis — Latent CoT reasoning accuracy depends on information fidelity, suggesting mutual information maximization over geometric imitation.
• ThoughtFold: Folding Reasoning Chains via Introspective Preference Learning — ThoughtFold uses introspective preference learning to prune redundant steps in reasoning chains, significantly boosting LRM efficiency.
• On the Expressive Power and Limitations of Multi-Layer SSMs — Multi-layer SSMs have inherent compositional limitations, but online Chain-of-Thought reasoning can make them as powerful as streaming algorithms.
• Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability — Reasoning SFT can generalize cross-domain, but it is conditional on optimization, data quality, and model capability.
• Learning to Label: A Reinforced Self-Evolving Framework for Semi-supervised Referring Expression Segmentation — A reinforced self-evolving framework improves semi-supervised referring expression segmentation by learning to construct reliable pseudo-labels.

Open in AIssential →

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