L'IA peut-elle résoudre des énigmes logiques standardisées à un niveau supérieur au percentile 99 ?

Les jeux de logique LSAT, le raisonnement quantitatif GRE, des formats similaires — les LLMs modernes se situent confortablement dans le top décile.

Background

Standardized logic puzzles, such as those found in LSAT logic games, GRE quantitative reasoning sections, Sudoku, KenKen, and logic grid puzzles, require solvers to apply formal rules under time pressure. These formats are designed to assess deductive reasoning, constraint satisfaction, and strategic problem decomposition. AI systems leverage symbolic reasoning, constrained optimization, and search algorithms (e.g., backtracking, SAT solvers, or neural-symbolic hybrids) to navigate large solution spaces efficiently. Research has demonstrated that modern deep learning architectures—particularly transformer-based models—can internalize logical structures through training on massive datasets of solved puzzles, enabling them to generalize to unseen instances. For example, models fine-tuned on logic-grid puzzles can infer implicit constraints from partial information, a task historically challenging even for advanced solvers. Benchmarks like the LSAT’s Analytical Reasoning sections have shown AI systems achieving performance in the top decile, often matching or exceeding human solvers on average, though variability exists depending on puzzle complexity and domain transfer. Studies highlight that AI’s advantage stems from its ability to decouple rule application from cognitive load, avoiding biases like confirmation or anchoring effects that human solvers may encounter. However, certain edge cases—such as puzzles with highly abstract or meta-level constraints—remain areas of active research. Sources: Science Daily (Enriched May 9, 2026).