[Paper Review] 2048 Without New Tiles Is Still Hard

Ahmed Abdelkader, Aditya Acharya+1 more|arXiv (Cornell University)|Jan 15, 2015

Artificial Intelligence in Games2 references4 citations

TL;DR

This paper proves that the popular 2048 tile-sliding puzzle is NP-hard, even when no new tiles are generated during play. The authors reduce 3-SAT to a generalized 2048 game by designing custom board gadgets—such as collapser, variable, and clause gadgets—that simulate logical constraints. The key result is that determining whether a 2048 tile can be created from a given board configuration is computationally intractable in the NP sense.

ABSTRACT

We study the computational complexity of a variant of the popular 2048 game in which no new tiles are generated after each move. As usual, instances are defined on rectangular boards of arbitrary size. We consider the natural decision problems of achieving a given constant tile value, score or number of moves. We also consider approximating the maximum achievable value for these three objectives. We prove all these problems are NP-hard by a reduction from 3SAT. Furthermore, we consider potential extensions of these results to a similar variant of the Threes! game. To this end, we report on a peculiar motion pattern, that is not possible in 2048, which we found much harder to control by similar board designs.

Motivation & Objective

• To investigate the computational complexity of 2048 when no new tiles are generated.
• To determine whether creating a specific high-value tile (e.g., 2048) is decidable within a given board configuration.
• To extend the complexity analysis to Threes!, another slide-and-merge puzzle with similar mechanics.
• To formalize decision problems in generalized 2048 and prove their NP-hardness using reductions from known NP-complete problems.

Proposed method

• Constructed a generalized 2048 game with arbitrary board size and fixed initial tile configuration, excluding random tile spawns.
• Designed specialized board gadgets: collapser (for 2×2 tile shifts), variable (for truth assignment via left/right merges), and clause (for logical satisfaction via vertical alignment).
• Used a 2-4 checkerboard lattice to preserve parity and prevent unintended merges.
• Reduced 3-SAT to the 2048 decision problem by encoding variables and clauses into tile configurations that simulate logical constraints.
• Demonstrated that a successful sequence of moves leading to a 2048 tile corresponds exactly to a satisfying assignment of the 3-SAT instance.
• Adapted the construction for Threes! by adjusting tile values (3·2^k) and accounting for persistent gaps in tile dynamics.

Experimental results

Research questions

• RQ1 Is it NP-hard to determine whether a 2048 tile can be created from a given initial board configuration without new tile generation?
• RQ2 Can the 2048 game be used to simulate the computation of a 3-SAT instance through tile dynamics and merges?
• RQ3 How do the dynamics of Threes! compare in terms of computational complexity, especially with persistent gaps after merges?
• RQ4 Can the same reduction technique used for 2048 be adapted to prove NP-hardness for Threes! despite differences in tile movement and merging rules?

Key findings

• The decision problem 'Can a 2048 tile be created from a given board?' is NP-hard, proven via a many-one reduction from 3-SAT.
• The construction uses polynomially bounded gadgets (collapser, variable, clause) that simulate logical variables and clauses through controlled tile merges.
• The checkerboard lattice ensures that only intended merges occur, preserving the integrity of the logical encoding.
• For Threes!, the same reduction strategy applies, though gaps from merges require careful spatial separation to prevent interference.
• The authors conjecture that Threes! is also NP-hard, based on structural similarity and the feasibility of adapting the 2048 reduction.
• The paper corrects a prior claim of PSPACE-completeness for 2048, showing instead that the construction in that work only achieves NP-completeness due to irreversible merges and inconsistent move semantics.