R-KV: Redundancy-aware KV Cache Compression for Training-Free Reasoning Models Acceleration

Reasoning models have demonstrated impressive performance in self-reflection and chain-of-thought reasoning. However, they often produce excessively long outputs, leading to prohibitively large key-value (KV) caches during inference. While chain-of-thought inference significantly improves performance on complex reasoning tasks, it can also lead to reasoning failures when deployed with existing KV cache compression approaches. To address this, we propose Redundancy-aware KV Cache Compression for Reasoning models (R-KV), a novel method specifically targeting redundant tokens in reasoning models. Our method preserves nearly 100% of the full KV cache performance using only 10% of the KV cache, substantially outperforming existing KV cache baselines, which reach only 60% of the performance. Remarkably, R-KV even achieves 105% of full KV cache performance with 16% of the KV cache. This KV-cache reduction also leads to a 90% memory saving and a 6.6X throughput over standard chain-of-thought reasoning inference. Experimental results show that R-KV consistently outperforms existing KV cache compression baselines across two mathematical reasoning datasets.

Entropy-based Exploration Conduction for Multi-step Reasoning

In large language model (LLM) reasoning, multi-step processes have proven effective for solving complex tasks. However, the depth of exploration can significantly affect the reasoning performance. Existing methods to automatically decide the depth often bring high costs and lack flexibility, and thus undermine the model's reasoning accuracy. To address these issues, we propose Entropy-based Exploration Depth Conduction (Entro-duction), a novel method that dynamically adjusts the exploration depth during multi-step reasoning by monitoring LLM's output entropy and variance entropy. We employ these two metrics to capture the model's current uncertainty and the fluctuation of uncertainty across consecutive reasoning steps. Based on the observed changes, the LLM selects whether to deepen, expand or stop exploration according to the probability. In this way, we balance the reasoning accuracy and exploration effectiveness. Experimental results across four benchmark datasets demonstrate the efficacy of Entro-duction. We further conduct experiments and analysis on the components of Entro-duction to discuss their contributions to reasoning performance.