Abstract: Many modern large-language models (LLMs) achieve high accuracy on complex reasoning benchmarks, but their multi-step inference processes remain largely opaque, making it difficult to trust ...