When Adaptive Rewards Hurt: Causal Probing and the Switching-Stability Dilemma in LLM-Guided LEO Satellite Scheduling

Research reveals that adaptive reward mechanisms in AI-guided satellite scheduling systems actually hurt performance, with static reward weights achieving 342.1 Mbps versus dynamic weights at only 103.3 Mbps. The study found that fine-tuned LLMs performed poorly due to weight oscillation issues, while simpler MLP models achieved superior results of 357.9 Mbps.

• →Static reward weights significantly outperformed adaptive ones in LEO satellite scheduling due to the need for stable convergence in reinforcement learning.
• →Fine-tuned LLMs collapsed to poor performance (45.3 Mbps) due to weight oscillation rather than lack of domain knowledge.
• →Simple MLP models achieved the best results at 357.9 Mbps on known regimes and 325.2 Mbps on novel regimes.
• →Causal probing revealed counterintuitive findings that switching penalty increases yielded substantial performance gains.
• →The research suggests LLMs add value in natural language understanding but simpler methods suffice for technical optimization tasks.