Dynamic Thinking-Token Selection for Efficient Reasoning in Large Reasoning Models
This addresses computational overhead for users of large reasoning models, though it is an incremental improvement on existing efficiency methods.
The paper tackles the efficiency bottleneck in Large Reasoning Models by identifying that only decision-critical tokens in reasoning traces significantly influence final answers, and proposes Dynamic Thinking-Token Selection to retain only these tokens' KV cache states during inference, achieving up to 40% memory reduction and 1.8× speedup with minimal accuracy loss.
Large Reasoning Models (LRMs) excel at solving complex problems by explicitly generating a reasoning trace before deriving the final answer. However, these extended generations incur substantial memory footprint and computational overhead, bottlenecking LRMs' efficiency. This work uses attention maps to analyze the influence of reasoning traces and uncover an interesting phenomenon: only some decision-critical tokens in a reasoning trace steer the model toward the final answer, while the remaining tokens contribute negligibly. Building on this observation, we propose Dynamic Thinking-Token Selection (DynTS). This method identifies decision-critical tokens and retains only their associated Key-Value (KV) cache states during inference, evicting the remaining redundant entries to optimize efficiency.