COMPASS: Efficient Structured Reasoning Via Action Policy Compression*

Abstract

As smart industry evolves toward human-centric collaboration, large language models are increasingly expected to function as cognitive agents that actively support complex reasoning and cooperative decision-making. However, although existing search-based reasoning methods can improve performance on complex reasoning tasks, their high computational cost and latency make them unsuitable for real-time collaborative decision-making. To address this limitation, we propose COMPASS (COMpressed Policy for Action-Selective Search), an efficient reasoning framework designed for complex knowledge reasoning and collaborative decision-making scenarios. COMPASS employs offline Monte Carlo Tree Search to obtain high-quality reasoning trajectories and compresses the embedded high-value decision patterns into an action selector, which then operates jointly with a language reasoning model to enable structured inference without explicit search at inference time. Experimental results on three benchmarks (GSM8K, GSM-Hard, and StrategyQA) across four open-source LLMs show that COMPASS consistently outperforms CoT-based and selfconsistency methods, while achieving performance comparable to search-based methods with token consumption reduced by approximately $\mathbf{9 5 \% - 9 7 \%}$. Further analysis indicates that COMPASS can learn stable high-level decision patterns, enabling efficient and goal-directed reasoning in complex multistep tasks.