Learning How to Search for Plans with Exponentially Less Space

TEXT START: Heuristic search for a plan can store exponentially many states, even when its heuristic is almost perfect.

The Dissection

The paper compresses planning. It replaces memory-intensive state enumeration with a learned, domain-specific control policy using registers, modes, termination certification, and limited backtracking. The exponential burden is not eliminated; it is concentrated in choice depth.

That is still a meaningful automation advance. Planning becomes cheaper to run, easier to deploy under tight memory limits, and less dependent on brute-force search. The system converts a former computational bottleneck into a compact controller.

The Core Fallacy

The dangerous misreading is to treat search complexity as a permanent barrier to automation. This paper shows it is an engineering bottleneck that can be reorganized, compressed, and partially automated.

The result does not prove universal cognitive automation. Its policies are domain-specific, its execution time can remain exponential in choice depth, and its benchmarks do not establish dominance across real-world cognition. But those are limits on scope, not evidence that human planning labor is safe. The paper attacks one more excuse for requiring it.

Hidden Assumptions

• Domains can be represented cleanly enough for one learned specification per domain.
• Structural termination can be certified and remains valid outside the training tasks.
• Choice depth stays small under distribution shift and difficult edge cases.
• Benchmark success transfers to messy environments with incomplete information and changing objectives.
• The cost of generating, verifying, maintaining, and porting policies remains lower than the human labor they replace.
• The language-model counterexample loop scales without verification becoming the new bottleneck.
• Memory and search were the binding constraints, rather than sensing, actuation, reliability, or exception handling.

Social Function

Primary classification: partial truth. The technical result appears to solve a real planning constraint rather than offer motivational fiction.

Its broader function is transition management. It makes autonomous planning cheaper while leaving the labor consequences outside the frame. That omission is not neutral: capability research accumulates toward P1, while institutions continue pretending that each isolated efficiency gain has no systemic consequence.

This is not copium. It is infrastructure for the transition.

The Verdict

A clean technical advance with an ugly systemic implication: less memory, less brute-force computation, and fewer human planning bottlenecks. It does not establish the full Discontinuity Thesis, but it pushes directly in its direction. The post-WWII employment circuit is not defended by better planning algorithms; it is further deprived of a reason to keep humans inside the loop.