Natural-Language Intermediate Reasoning

Extra tokens or post-answer rationales are not enough — useful CoT needs natural-language steps before the answer, not only equations or padding.

Ablations on arithmetic chain-of-thought prompting isolate what the intermediate text must do:

• Equation only helps short one-/two-step sets but barely moves GSM8K — hard word problems resist a single equation map.
• Variable compute only (emit dots matching equation length) ≈ standard prompting — burning tokens without semantic steps does not help.
• Reasoning after the answer ≈ baseline — a post-hoc rationale does not substitute for sequential reasoning before committing.

So harness defaults should elicit natural-language steps prior to the final answer, not merely longer completions or “explain yourself” after the fact. This supports when chain of thought helps on hard multi-step tasks and pairs with two-stage reasoning–answer extraction when you need a separate parseable answer. Prefer self-consistency decoding over padding when you want more compute on hard items.

---

Sources

• Chain-of-Thought Prompting (Wei et al.), §3.3