Adapting to a completely new environment and reasoning about which outcomes are possible under unseen rules is a fundamental component of general intelligence. This work investigates the ability in small discrete worlds where every grid configuration follows a shared base rule set plus a few extra constraints, yielding Sudoku‑like puzzles. A single depth‑recurrent model is trained on many rule combinations and evaluated on four novelty levels: a brand‑new puzzle, a new combination of known rules, a pairing of rules never seen together, and a rule instance absent from training. The model receives active rules either as a symbolic description or as two solved examples; a baseline model receives no rule information. Results show that for rule recombinations the description matches or exceeds the example condition (99% vs 95% success on new combinations). For rule instances whose components never co‑occurred during training, the description falls below the no‑information baseline while examples remain above it. The framework enables systematic testing of many generalization regimes, addressing limitations of benchmarks such as ARC‑AGI, and offers new ground for evaluating depth‑recurrent reasoning models that are usually tested only on saturated Sudoku benchmarks.
Review