Dynamic Thinking-Token Selection for Efficient Reasoning in Large Reasoning Models

Researchers introduce Dynamic Thinking-Token Selection (DynTS), a method that optimizes Large Reasoning Models by identifying and retaining only decision-critical tokens during inference while discarding redundant reasoning trace data. This approach significantly reduces memory footprint and computational overhead, addressing a major efficiency bottleneck in LRMs that generate extended reasoning sequences.

Large Reasoning Models represent a computational frontier in AI, tackling complex problem-solving by explicitly working through reasoning traces before generating answers. However, this multi-step reasoning capability comes at a significant cost: the extended token generation creates substantial memory demands and computational overhead that limits practical deployment and scalability. The DynTS research addresses this tension directly through a novel efficiency optimization.

The core insight—that only a subset of tokens within reasoning traces meaningfully influence the final answer—stems from attention map analysis revealing that most reasoning tokens contribute negligibly to model output. This finding challenges assumptions about how LRMs process information and suggests that reasoning traces contain substantial redundancy. The methodology leverages this observation by identifying decision-critical tokens through attention patterns and selectively maintaining only their associated Key-Value cache states during inference.

For the AI infrastructure and deployment ecosystem, this optimization carries meaningful implications. Reduced memory requirements directly translate to lower computational costs, faster inference speeds, and improved feasibility for edge deployment scenarios. Organizations deploying LRMs at scale face significant operational expenses tied to memory and compute; efficiency gains compound across thousands or millions of inference operations. This addresses a practical barrier to broader LRM adoption beyond resource-rich enterprises.

The research suggests an emerging trend toward post-hoc optimization of reasoning-based models, where efficiency improvements focus on inference-time pruning and selective computation rather than architectural redesigns. Future development likely involves combining DynTS with other efficiency techniques and exploring whether similar token-selection principles apply across different LRM architectures or reasoning domains.

• →DynTS reduces LRM inference overhead by retaining only decision-critical tokens and discarding redundant reasoning trace entries
• →Attention map analysis reveals that most tokens in reasoning traces contribute negligibly to final answers, indicating substantial efficiency potential
• →The method optimizes Key-Value cache usage during inference, directly addressing memory and computational bottlenecks limiting LRM deployment
• →Efficiency improvements enable broader adoption of Large Reasoning Models across resource-constrained environments and edge deployment scenarios
• →The research opens pathways for post-hoc optimization strategies that selectively maintain relevant computation during inference