EquivPruner: Boosting Efficiency and Quality in LLM-Based Search via Action Pruning

Researchers introduce EquivPruner, a method that reduces token consumption in LLM reasoning searches by identifying and pruning semantically equivalent steps. Combined with MathEquiv, a new dataset for mathematical equivalence detection, the approach achieves 48.1% token reduction on GSM8K while maintaining or improving accuracy.

EquivPruner addresses a fundamental inefficiency in how LLMs conduct reasoning through search algorithms. Current approaches waste computational resources by exploring redundant pathways that, while semantically different in structure, represent equivalent logical steps. This redundancy translates directly to unnecessary token consumption and higher operational costs—a critical constraint for developers and organizations deploying LLM-based reasoning systems at scale.

The problem emerges from limitations in existing semantic similarity detection methods, which struggle in specialized domains like mathematical reasoning where surface-level differences mask logical equivalence. The introduction of MathEquiv, a purpose-built dataset for mathematical statement equivalence, enables training lightweight detectors that can reliably identify such equivalences without requiring massive compute resources themselves. This represents a shift toward domain-specific optimization rather than generic similarity metrics.

The practical implications are substantial for LLM economics. A 48.1% reduction in token consumption directly translates to lower API costs and faster inference times—material advantages for cost-sensitive applications and real-time systems. The simultaneous accuracy improvements suggest the pruning approach eliminates genuinely redundant exploration paths rather than cutting corners. This creates a win-win scenario where efficiency and quality align.

Looking forward, the framework's applicability across multiple models and tasks suggests broader potential beyond mathematics. Organizations optimizing LLM inference pipelines should monitor whether this pruning methodology extends effectively to other reasoning-intensive domains like code generation, logical deduction, and scientific problem-solving. The release of code and datasets enables community iteration and potential standardization of equivalence-based optimization techniques.

• →EquivPruner reduces token consumption by 48.1% on mathematical reasoning tasks while maintaining or improving accuracy.
• →MathEquiv is the first dataset specifically designed for training mathematical statement equivalence detection models.
• →The approach addresses inefficient exploration of semantically equivalent steps in LLM search algorithms.
• →Token efficiency improvements have direct cost implications for deploying reasoning-based LLM systems at scale.
• →The method generalizes across multiple models and tasks, suggesting broader applicability beyond mathematics.