IS-CoT: Breaking the Long-form Generation Collapse via Interleaved Structural Thinking

arXiv·medium signal

Targets the persistent failure of LLMs to stay coherent and controllable over long-form generation, where reasoning-enhanced models succeed on logic-heavy tasks but still collapse on extended content. Proposes interleaving structural thinking with generation to maintain coherence. Practical for builders generating long documents, reports, or newsletters where mid-output drift is a known problem.