ASSCG: Just-Right Gating over Chattering for Fast–Slow LLM Planning in Autonomous Driving

Autonomous driving systems often use a fast-slow paradigm where a real-time planner handles immediate control while a compute-heavy LLM provides high-level guidance. However, existing methods for triggering the slow LLM are often inefficient, relying on fixed intervals or heuristic proxies that fail to account for the actual marginal utility of LLM reasoning. This paper asks: how can we optimize the timing and usage of slow-system guidance to maximize driving performance while minimizing computational costs?

The authors introduce the Adaptive Slow-System Control Gate (ASSCG), a lightweight controller that treats slow-system invocation as a sequential decision problem. At each frame, ASSCG predicts one of three actions: Query (refresh guidance), Cache (reuse existing guidance), or Drop (ignore guidance to prevent harmful interventions). The gate is implemented using an RWKV backbone—chosen for its efficient, constant-time inference over long horizons—and is trained using a combination of supervised fine-tuning and compute-aware reinforcement learning (GRPO). This allows the system to learn when to trust the LLM and when to rely solely on the fast planner.

ASSCG demonstrates significant improvements across two different fast-slow architectures. When integrated into AsyncDriver on the nuPlan Hard20 benchmark, it improved the planning score by 2.28 points while reducing end-to-end inference latency by approximately 60%. In a second architecture based on RecogDrive evaluated on NAVSIM, the system achieved a 0.6 increase in PDMS while simultaneously increasing the average vehicle speed by 25%. These results suggest that a learned, frame-level gate is more effective than traditional rule-based or complexity-based triggers.

This work addresses a critical bottleneck in deploying LLMs for real-time robotics. By moving away from rigid triggering rules toward a learned, context-aware gating mechanism, the authors show that it is possible to maintain the reasoning benefits of large models without incurring the prohibitive latency costs of per-frame execution. This approach provides a scalable template for integrating heavy reasoning modules into time-sensitive autonomous systems.