SGP-CoT: A Self-Guided Chain-of-Thought Pruning Technique for Large Language Models to Independently Determine Their Reasoning Paths

The ACL 2026 main conference paper SGP-CoT proposes an unsupervised chain-of-thought pruning method that enables reasoning models to independently assess the importance of thinking steps. It significantly reduces computational overhead while maintaining reasoning quality, providing a new solution for large model reasoning optimization.

As large language models (LLMs) improve their performance on complex reasoning tasks, chain-of-thought (CoT) prompting has become the mainstream method to stimulate reasoning abilities. However, lengthy intermediate steps lead to high computational costs and long reasoning delays, limiting applications in resource-constrained environments. How to streamline reasoning paths while maintaining reasoning quality is a key challenge currently.

Core Idea: Your Reasoning Model Knows What Counts—without manual annotation or additional evaluation models, leveraging its own capabilities to judge the value of steps. Technical Mechanism: 1. Step Importance Evaluation: After generating a complete reasoning chain, guide the model to self-evaluate the importance of each step; 2. Dynamic Threshold Pruning: Adaptively adjust pruning intensity based on task difficulty; 3. Reconstruction of Optimized Reasoning Chain: Reorganize the retained steps into a coherent path.

1. Improved Computational Efficiency: Reduce token count, lower latency and resource consumption; 2. Fully Self-Supervised: No manual annotation required, can be seamlessly integrated into CoT-supported LLMs; 3. Enhanced Interpretability: Explicitly identify key steps, clearly show the decision-making process; 4. Flexible Adaptation: Dynamic thresholds adapt to different tasks, allowing adjustment of the latency-accuracy trade-off.

Application Scenarios: Real-time dialogue systems (improve user experience), mobile devices (local deployment feasible), multi-round complex reasoning (error analysis and debugging). Future Directions: Combine with speculative decoding and model quantization; expand to multimodal reasoning scenarios.

SGP-CoT is an important advancement in the field of chain-of-thought optimization. It proves that LLMs can independently identify and optimize reasoning processes, providing a new perspective for understanding and improving model thinking mechanisms. It has important reference value for researchers and engineers working on large model reasoning optimization.