LLM Parkinsonism: Uncertainty-Aware Global Executive Control Architecture for Autonomous LM Agents

What it claims

The headline result is narrower than the 96.57% figure suggests. In the 24,000-episode benchmark (40,000-token ceiling), most of the jump from 67.42% to roughly 96.5% hard-goal success comes from giving the agent access to several candidate actions. GEC itself adds nothing beyond that: a candidate-set local control reaches 96.53% without GEC.

What GEC adds over that control is cost and discipline:

• Mean tokens: down from 19,782 to 12,574 (36.4%).
• Restricted mean tokens to completion: down from 16,136 to 13,114 (18.7%) at the 40,000-token ceiling.
• Pre-completion drift: the measured drift is eliminated.
• Complexity: gross complexity drops sharply.

If you adopt only one idea, generating and choosing among candidate actions buys the success. Governance buys the efficiency.

How it works

• Goal Contract. The project is a contract G = (g, R_H, R_S, N, C, M):
• a primary goal
• hard requirements
• explicitly authorized soft objectives
• explicit non-goals
• constraints
• an amendment policy

An LLM proposal is not a requirement. Contract changes are revisioned amendment events authorized outside the executor. The reference implementation rejects anonymous or self-authorized amendments, which targets "self-generated requirement promotion" (for example, an optional "site-level alert aggregation" idea later being treated as a committed feature with tests).

• Independent scope adjudication. Each action gets a scope label: DIRECT, PREREQUISITE, VERIFICATION, RISK_MITIGATION, SOFT, NONE or FORBIDDEN. The generator's own label is not authoritative. The adjudicator could be a dependency graph, rule engine, separate model, human or ensemble; what matters is the separation of authority. NONE and FORBIDDEN actions count as unscoped unless the contract is amended.
• State-versioned evidence. Evidence for a criterion is stored as (status, verifier, state version, confidence, dependencies). Later changes can invalidate it, so a test that passed at version k does not certify completion at version k+m.
• Two-level stopping. Candidate actions are scored as V = E[ΔU] − λ_T·C_T − λ_K·C_K − λ_R·C_R − λ_V·C_V. That is expected verified utility gain minus weighted token, complexity, risk and verification costs.
• A low score (V ≤ τ) rejects that candidate only. It does not stop the project.
• Stopping is a state-level judgment over the whole candidate set.
• Unmet hard requirements with no feasible path end in BLOCKED, not GOOD ENOUGH.
• Stopping for economic reasons applies only after hard completion.

flowchart TD
  G[Generator proposes candidates] --> S{Independent scope adjudicator}
  S -->|NONE / FORBIDDEN| R[Reject unless contract amended]
  S -->|DIRECT / PREREQ / VERIFY / RISK / SOFT| V{Action score V > τ?}
  V -->|no| R
  V -->|yes| X[Execute]
  X --> E[Versioned evidence; invalidate stale]
  E --> H{Hard goals met with valid evidence?}
  H -->|no, path exists| G
  H -->|no, no path| B[BLOCKED]
  H -->|yes| C{Any soft continuation worth cost?}
  C -->|yes| G
  C -->|no| D[STOP]

Evidence

The paper reports the following metrics, all from the LPB v0.2 benchmark:

• Restricted mean tokens to completion (RMTTC_B). Episodes that fail to complete count as the full budget B rather than being dropped, which avoids survivorship bias. The paper treats conditional tokens-to-completion as secondary only.
• Termination Overrun Ratio (TOR). Measures the tail of work after the goal is complete.
• Pre-completion drift rate (GDR_pre). Kept separate from TOR so the two are not double-counted.
• Useful-token ratios. A direct ratio and a process ratio. The process ratio credits validated prerequisites and verification work.

The governance-overhead sensitivity analysis stayed favorable up to an extra 500 synthetic governance tokens per cycle. This is a rough budget for what a live adjudicator could cost before the savings disappear.

According to the paper, LPB v0.2 includes:

• matched candidate sets
• the candidate-set control
• budget and ablation studies
• noise robustness tests
• probes of beneficial soft work

The fetched text does not give a URL or licence for it.

Limits

• Simulations only. All results come from mechanistic simulations, and the authors say live-model validation is still needed. The 500-token overhead figure is synthetic.
• Action amplitude is not measured. The "action amplitude" signature behind the Parkinsonism metaphor is illustrative only and is not a validated endpoint.
• Constraints are not exercised. LPB v0.2 does not exercise a general-purpose constraint predicate engine, so the constraint set C is untested.
• Token efficiency is not independent evidence. When all compared trajectories succeed, the token-efficiency metric is just 1000/T_total and duplicates the token-cost result.
• Text is truncated. The fetched text stops in the metrics section, so the ablation and noise-robustness results are not visible here.