AI News

Large language models can plan, use tools, write code, and complete long-horizon workflows, but strong local actions do not guarantee effective project-level control. The paper uses “LLM Parkinsonism” as a narrowly defined, non-clinical metaphor for continuing to act after the original objective is satisfied, including low-value refinements, repeated checks, and repairs to complexity created by the agent itself. It argues that the failure is more directly associated with placing proposal generation, scope interpretation, progress assessment, and stopping authority in one self-conditioned loop than with autoregressive prediction alone. The authors introduce Global Executive Control (GEC) v0.2, an uncertainty-aware architecture that separates action generation from project-level governance. In a 24,000-episode matched-candidate benchmark with a 40,000-token ceiling, first-candidate control reached 67.42% hard-goal success, while candidate-set local control reached 96.53% and GEC reached 96.57%. Compared with candidate-set control, GEC maintained success while reducing mean token use from 19,782 to 12,574, a 36.4% decrease, and reducing mean tokens to completion from 16,136 to 13,114, an 18.7% decrease. The study also reports no measured pre-completion drift, substantially lower gross complexity, and favorable results after adding 500 synthetic governance tokens per cycle. The findings come from mechanistic simulations, so validation with live models remains necessary.