We identify three fundamental bottlenecks in current DLMs: (1) Low introspective consistency. SDAR: 0.699 vs. I-DLM: 0.984. (2) Compute inefficiency. TiDAR: ~7.8x overhead vs. I-DLM: ~2.5x. (3) Infrastructure mismatch. SDAR slope=84 vs. I-DLM: 549.
作者系统性地识别了现有DLMs的三大瓶颈，并通过量化对比展示了I-DLM的显著优势。内省一致性从0.699提升到0.984，计算开销从7.8x降低到2.5x，基础设施效率从84提升到549，这些数据不仅验证了I-DLM的有效性，也为未来DLM研究指明了方向。