Reliability of inducing Long CoT structures via prompting in instruction-tuned models

Ascertain how reliably molecular-like Long Chain-of-Thought structures can be induced in instruction-tuned language models through prompting alone, as opposed to being obtained via distillation from strong reasoning models.

Background

The authors propose Mole-Syn, a structure-aware synthesis framework, to transfer behavior-transition structure without copying teacher outputs. Before introducing Mole-Syn, they state uncertainty about whether prompting instruction-tuned models can reliably induce the targeted Long CoT structural patterns without distillation.

This open question frames the need for synthesis methods that decouple structural transfer from surface forms and motivates evaluating instruction-driven generation of Long CoT behavior distributions.

References

LLMs may acquire advanced reasoning partly through exposure to explicit, structured Long CoT reasoning traces. However, it remains unclear how reliably such structures can be induced by prompting an instruction-tuned model, rather than obtained through distillation.

— The Molecular Structure of Thought: Mapping the Topology of Long Chain-of-Thought Reasoning  (2601.06002 - Chen et al., 9 Jan 2026) in Section “Synthetic Chemistry: Synthesis Long CoT Molecules from Scratch,” opening paragraph