DISC: Dynamic Decomposition Improves LLM Inference Scaling

2025-02-23

AI Summary

Existing reasoning scaling methods rely on predefined, fixed-granularity step decomposition, limiting adaptability to varying problem complexity. This paper proposes a dynamic decomposition mechanism that adaptively partitions both the solution process and reasoning trace into variable-granularity steps during inference. Our core contributions are threefold: (1) the first confidence-driven real-time step decomposition; (2) priority-based sampling scheduling; and (3) multi-granularity reasoning path management—enabling difficulty-aware, dynamic allocation of computational resources and overcoming the limitations of static token-level or sentence-level step paradigms. Evaluated on APPS, MATH, and LiveCodeBench, our method achieves significant improvements in both reasoning efficiency and task accuracy, empirically validating the effectiveness and generalizability of adaptive step decomposition for scaling large language model reasoning.

Technology Category

• Planning, Routing, and Scheduling: Model-Based Reasoning
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Many inference scaling methods work by breaking a problem into smaller steps (or groups of tokens), then sampling and choosing the best next step. However, these steps and their sizes are usually predetermined based on human intuition or domain knowledge. This paper introduces dynamic decomposition, a method that automatically and adaptively splits solution and reasoning traces into steps during inference. This approach improves computational efficiency by focusing more resources on difficult steps, breaking them down further and prioritizing their sampling. Experiments on coding and math benchmarks (APPS, MATH, and LiveCodeBench) show that dynamic decomposition performs better than static methods, which rely on fixed steps like token-level, sentence-level, or single-step decompositions. These results suggest that dynamic decomposition can enhance many inference scaling techniques.

Problem

• Dynamic decomposition improves LLM inference scaling
• Automatically adapts problem splitting during inference
• Enhances computational efficiency in complex steps

Innovation

• Dynamic decomposition adaptively splits steps
• Improves efficiency by focusing on difficult steps
• Outperforms static methods in benchmarks