minRLM Beats Sudoku With REPL Loop

The win here is not “LLMs learned Sudoku.” It’s that minRLM turns a brittle generation problem into a recoverable search process, which is exactly where LLMs are strongest when paired with tools.

• –Sudoku is a harsh test because one wrong digit contaminates the entire solution, so vanilla models tend to produce fluent-looking but invalid grids.
• –The generate → execute → fix → repeat loop gives the model state, verification, and rollback, which are the missing pieces in one-shot prompting.
• –This is a systems result more than a model breakthrough: the base LLM is still doing language work, while the REPL/backtracking layer does the actual search.
• –The benchmark message is broader than Sudoku. Any task with hard constraints, long dependency chains, or expensive recovery benefits from an interpreter, checker, or agent loop.
• –For builders, the takeaway is practical: don’t ask the model to be right once; wrap it in a workflow that can detect errors and continue.