Sudoku
Single
All projects
Work in progress: this project explores whether LLMs can solve Sudoku without external tools.
Abstract: the goal is to build and study a benchmark-driven setup for Sudoku reasoning, and test a core hypothesis: models still struggle on purely computational tasks even when no external tools or orchestration should be needed (for example, solving a simple Sudoku). As long as they lack this innate computational capability, I believe they do not truly understand computation and cannot solve Sudoku reliably over very long contexts. Current references include Can LLMs be Computers?, and Sudoku Bench.
for now, you can play Sudoku, if you're interested :)
Mode