Efficiency Collapse in Multi-Step LLM Execution: An Empirical Study of Cost, Redundancy, and Phase Dynamics

This work presents an empirical analysis of multi-step execution in large language models, focusing on how marginal contribution evolves relative to computational cost across sequential steps. Using redundancy-adjusted proxies for marginal information gain, we analyze execution trajectories across models, tasks, and prompt variations. We observe consistent patterns of early contribution concentration, redundancy accumulation, and diminishing marginal contribution under continued execution beyond early convergence points. The results highlight the absence of observable runtime signals for marginal contribution in current systems, where continuation decisions are not conditioned on execution state. This motivates the need for execution-aware mechanisms that incorporate trajectory and state into continuation decisions.