PACE Framework Cuts Embodied Planning Delays via Interleaved Reasoning

A new framework called PACE, which stands for Planning with Adaptive Cognitive Effort, has been developed to address the long delays in reasoning that occur in large language models used for embodied systems, where planning can take several minutes. The problem comes from traditional methods that complete all reasoning before any actions can be taken, leading to inefficiencies. PACE introduces two main features: an Interleaved Think-Act architecture that allows for reasoning and action to happen at the same time, and a Dynamic Budget Allocator that modifies reasoning token budgets based on how much time is left for execution. In trials using the Robotouille benchmark with Qwen3-8B-AWQ, PACE achieved a 10% success rate, which is a 67% improvement over the ReAct+Think method. More details are available in a paper on arXiv (arXiv:2608.03034), which was announced as a cross-type submission.