A team at the Recurse Center has built an algorithmic solver for Semantle, the Wordle variant that challenges players to guess words based on semantic similarity rather than spelling. While humans typically need dozens of guesses to find the target word, this solver consistently finds answers in around three guesses by leveraging the geometric properties of word embeddings.

Semantle uses Google News word2vec embeddings, representing each word as a 300-dimensional vector and scoring guesses using cosine similarity. The solver works by maintaining a candidate list of all possible words and iteratively filtering based on similarity scores. Each guess constrains the target to a ring on the unit embedding sphere, dramatically reducing the candidate pool. After just one or two guesses, millions of candidates can be narrowed to a handful.

The approach exploits the sparsity of vocabulary within high-dimensional embedding space. While humans gradually hone in on answers through intuitive semantic connections, the solver takes a global approach, making seemingly random guesses that nonetheless converge quickly through geometric constraints. The team found that direct algebraic solutions were impractical, requiring hundreds of guesses, while their filtering method proved remarkably efficient.