Tool‑assisted planning can succeed by leveraging external search and verification, leaving the model's unaided planning ability unclear. We isolate innate planning as the generation of a complete plan within a single response, without tools or execution feedback. A verifiable framework is introduced that decomposes success‑weighted plan efficiency as $$\mathrm{SPE}=R\times Q\times E$$, where $R$ denotes execution reliability, $Q$ captures coordination among legal plans, and $E$ measures action economy of completed plans. Trajectory replay and certified optimal solutions are used to verify each factor; the overall score also incorporates unsuccessful attempts. We evaluated 8,400 full‑plan outputs from seven model conditions on $4\times4$ sliding puzzles, controlling exact horizon length and obstacle count in the primary cohort. Findings show that even models that complete every trial become less economical as the horizon grows, and higher economy among completed plans can coexist with poor goal attainment. $\mathrm{SPE}$ separates the overall capability hierarchy, while its components distinguish coordination failures from inefficient completions. The framework makes long‑horizon limitations of unaided planning quantitatively verifiable.
Review