ReCo Optimizes LLM Reasoning, Reduces Inference Costs.

Large Reasoning Models (LRMs) often employ lengthy chain-of-thought (CoT) reasoning, which, while effective for complex tasks, leads to "overthinking" and inflated inference costs due to large KV-caches. Existing KV-cache compression methods typically apply uniform policies and only consider what's removed from the cache, overlooking the varying tolerance to context loss along a reasoning trajectory and the potential for smaller caches to increase token generation. This research introduces ReCo (Reward-Coordinated Compression), a novel step-wise framework designed to optimize LRM efficiency. ReCo uses a lightweight process-reward estimator to score each completed reasoning step. This reward signal then drives three key components: first, reward-adaptive KV-cache compression, which aggressively shrinks the cache at high-reward steps and less so at low-reward ones; second, a reward-banded penalty on reflection tokens to curb redundant generation; and third, confidence-based early stopping when reasoning is deemed reliable. Evaluations across multiple reasoning models and benchmarks demonstrate that ReCo significantly reduces generated tokens by 37%-65% and improves end-to-end latency by 2.08x-2.35x compared to full CoT, all while maintaining high accuracy. This coordinated approach addresses both cache size and token generation, leading to more efficient LRM deployment.