Technology
Hacker News

The Strongest Sokoban Solver

Source Entity

Hacker News

November 22, 2026
The Strongest Sokoban Solver

The newly developed Soksolv solver has outperformed existing benchmarks in the game Sokoban, achieving 24 world-first automated solutions. Developed entirely using AI-assisted coding, it demonstrates significant speed improvements and higher success rates over established solvers like Festival.

The Emergence of Soksolv: A New Benchmark in Automated Reasoning

The landscape of automated puzzle solving has shifted with the introduction of Soksolv, a high-performance solver designed to tackle the classic warehouse-keeper puzzle game, Sokoban. By achieving 24 world-first automated solutions across standard benchmark suites, Soksolv has effectively eclipsed the performance metrics of previously established solvers such as Festival, Sokolution, Takaken, YASS, and BoxSearch. This advancement represents a significant milestone in computational search algorithms, demonstrating both increased depth in problem-solving and a remarkable efficiency gain of approximately 3.2x in processing time compared to its predecessors.

Quantitative Superiority and Benchmark Performance

Testing across two major data sets—the 'Large Test Suite' comprising 3,272 levels and the 'Open Test Suite' spanning 10,043 levels—reveals the sheer scale of Soksolv’s capabilities. Data points indicate that at a 10-minute threshold, Soksolv maintains a lead of 10 and 18 levels ahead of the Festival solver, respectively. This performance is particularly noteworthy given that these comparisons were measured on identical hardware, ensuring that the results reflect algorithmic optimization rather than raw computational power disparity. The log-scale analysis of unsolved levels suggests that Soksolv is pushing the boundaries of what is computationally feasible within finite time constraints.

The Paradigm Shift: AI-Assisted Development

Perhaps the most compelling aspect of Soksolv is its origin story: the entire codebase was generated without a single line of handwritten code. Utilizing an extended session with Claude 3.5 Opus via Claude Code, the developer employed approximately 430 subagents to architect the solution. This marks a pivotal moment in software engineering where complex, performance-critical algorithms are being synthesized through high-level AI orchestration rather than traditional manual programming. This approach highlights the growing trend of AI as a force multiplier in algorithmic research.

Broader Implications for Algorithmic Search

Sokoban is a PSPACE-complete problem, meaning it serves as an excellent proxy for testing the efficiency of search heuristics and state-space exploration. The success of Soksolv suggests that the integration of large language models (LLMs) into the development loop can produce highly optimized logic that rivals or exceeds human-crafted code. As these solvers become more efficient, the techniques developed here could have implications for broader applications in logistics, pathfinding, and automated planning where complex constraints must be navigated within restricted environments.

Future Trends and Concluding Thoughts

As we look forward, the development of Soksolv signals a maturation in how developers approach niche, computation-heavy challenges. By leveraging AI to navigate the vast state-space of Sokoban, the developer has not only created a superior tool but has also provided a blueprint for future AI-driven software development. The ability to surpass legacy solvers in both speed and accuracy suggests that the next generation of automated solvers will likely rely heavily on these collaborative AI frameworks, setting a new standard for competitive puzzle-solving and computational research.

Verification Required?

Read the full report from the primary source

Go to Hacker News