RESEARCH · RESEARCH · #1106
Paper coins “LLM Parkinsonism” and introduces Global Executive Control (GEC) v0.2 to curb persistent, token-inefficient agent behavior
The arXiv preprint defines “LLM Parkinsonism” as persistent, low-value actions by language-model agents and argues this arises from conflating action proposal and project-level control. It introduces Global Executive Control (GEC) v0.2—an uncertainty-aware governance architecture that separates action generation from executive control—and reports on a 24,000-episode matched-candidate simulation where GEC matched candidate-set control on hard-goal success (96.57% vs 96.53%) while cutting mean token use from 19,782 to 12,574 (36.4%) and eliminating measured pre-completion drift; the authors note live-model validation remains necessary.
KEY POINTS
- The arXiv preprint defines “LLM Parkinsonism” as persistent, low-value actions by language-model agents and argues this arises from conflating action proposal and project-level control.
- It introduces Global Executive Control (GEC) v0.2—an uncertainty-aware governance architecture that separates action generation from executive control—and reports on a 24,000-episode matched-candidate simulation where GEC matched candidate-set control on hard-goal success (96.57% vs 96.53%) while cutting mean token use from 19,782 to 12,574 (36.4%) and eliminating measured pre-completion drift; the authors note live-model validation remains necessary.
- The work provides mechanistic evidence that explicit, uncertainty-aware governance separating action proposal from stopping and scope control can substantially reduce unnecessary computation and complexity in autonomous LLM agents.
WHY IT MATTERS
The work provides mechanistic evidence that explicit, uncertainty-aware governance separating action proposal from stopping and scope control can substantially reduce unnecessary computation and complexity in autonomous LLM agents.