ACIL: Auto Chain of Thoughts for In-Context Learning

2026-05-19 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, long

Summary

The Automatic Chain of Thought (Auto-CoT) framework enhances In-Context Learning (ICL) performance in Large Language Models (LLMs) by integrating automatically generated reasoning chains. This framework augments input-output pairs with structured explanations, prunes irrelevant or low-quality demonstrations, and systematically selects high-quality reasoning examples for the ICL prompt. Auto-CoT was validated across numerical function approximation tasks and textual tasks using datasets like LAMBADA and FinBERT. Experiments showed that Auto-CoT consistently reduced Mean Squared Error (MSE) in numerical tasks and achieved significantly lower loss in language modeling tasks, particularly for shorter context lengths (e.g., k=1), compared to baseline ICL methods. The approach leverages a GPT-2 model for reasoning chain generation and a variance-reduced policy gradient for optimal selection.

Key takeaway

For AI Engineers and Research Scientists working on improving LLM few-shot performance, Auto-CoT offers a systematic method to enhance In-Context Learning. By automatically generating and curating reasoning chains, you can significantly reduce prediction errors and improve model accuracy on both numerical and textual tasks. Consider implementing Auto-CoT's augmentation, pruning, and selection mechanisms to build more robust and adaptable LLM applications, especially when dealing with intricate reasoning requirements or limited context examples.

Key insights

Auto-CoT improves LLM in-context learning by automatically generating, pruning, and selecting reasoning chains.

Principles

• Explicit reasoning steps enhance ICL generalization.
• Diversity in demonstrations mitigates reasoning errors.
• Pruning low-quality chains improves inference quality.

Method

Auto-CoT augments training data with LLM-generated reasoning chains, prunes low-quality chains based on prediction error, and optimizes selection using a variance-reduced policy gradient.

In practice

• Use Auto-CoT for complex reasoning tasks in LLMs.
• Apply pruning to filter inconsistent reasoning chains.
• Optimize prompt selection for better inference.

Topics

• Auto Chain of Thought
• In-Context Learning
• Large Language Models
• Reasoning Chain Augmentation
• Demonstration Pruning

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

Related on AIssential

• CoT-Space: A Theoretical Framework for Internal Slow-Thinking via Reinforcement Learning — CoT-Space reframes LLM reasoning as continuous optimization in a semantic space, explaining optimal CoT length.
• Summarize Before You Speak with ARACH: A Training-Free Inference-Time Plug-In for Enhancing LLMs via Global Attention Reallocation — ARACH enhances LLMs by reallocating internal attention via a training-free context hub and logit offset at inference time.
• Where Do CoT Training Gains Land in LLM based Agents? — CoT training improves LLM agents' direct action prediction from prompts, with less impact on CoT's relative reasoning advantage.
• Context-CoT: Forcing LLMs to Actually Think (No ICL) — LLMs struggle to genuinely learn and reason from novel, complex knowledge provided solely in the prompt, often defaulting to pre-trained biases.
• A Language for Describing Agentic LLM Contexts — ACDL standardizes describing LLM agent context structure and evolution, improving communication and reproducibility.
• Researchers automated LLM reasoning strategy design and cut token usage by 69.5% — AutoTTS automates LLM test-time scaling strategy design, cutting token usage by 69.5% and improving accuracy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.