Beyond Input Understanding: Diagnosing Multilingual Mathematical Reasoning with Directed Acyclic Trace Graphs
The gap between strong English math reasoning and weaker performance in other languages is not just about understanding problem statements. Controlling the model's reasoning language significantly impacts accuracy, suggesting language affects how reasoning executes. Researchers introduced DATG to map traces to mathematical structures, finding that non-English traces suffer from poor anchor coverage and dependency fidelity, especially in low-resource settings. Test-time controls improved performance.