Output-Side Reasoning Augmentation

Improve hard tasks by eliciting intermediate tokens before the final answer — orthogonal to input-side tricks like instructions, few-shots, or soft prompts.

Many prompting methods improve the input: instructions, soft/learned prompts, few-shot exemplars, retrieval. Chain-of-thought prompting instead augments the output with intermediate natural-language steps that decompose multi-hop work before the final answer. Wei et al. place CoT as orthogonal to input-side instruction tuning: you can combine both.

Do not confuse this with post-hoc natural-language explanations that justify an already-chosen label — those target interpretability; CoT’s goal is sequential decomposition (see natural-language intermediate reasoning). Harnesses should version output-side cues (triggers, plan formats) separately from input packs in the prompt catalog, and still respect scale-dependent chain of thought: output-side reasoning is an emergent ability, not a free upgrade for tiny models. Pair with broad vs narrow prompting when choosing a single multi-task output cue vs per-task input templates. Discovering more broad prompts — not only CoT triggers — is the intended research direction beyond this baseline.

---

Sources

• Chain-of-Thought Prompting (Wei et al.), §7–§8; Appendix C