Paper summary: Chain-of-Thought Reasoning Without Prompting

Problem

Getting LLMs to reason step by step usually relies on prompting: few-shot examples or “let’s think step by step”. This adds manual effort and makes it hard to tell what the model can do intrinsically, versus what the prompt induced.

Method

The authors change only the decoding. Greedy decoding takes the single most likely token at each step. CoT-decoding instead looks at the top-k alternative tokens at the first step, continues each branch, and inspects the resulting paths. Chain-of-thought paths often appear in these alternatives. The model’s confidence in its final answer is noticeably higher on paths that contain reasoning, so it can be used to select them.

Key results

According to the paper, CoT paths are frequently present in pre-trained models without any prompting, and selecting them by confidence improves reasoning results over greedy decoding across several benchmarks.

Limitations

There’s no dedicated limitations section in the extracted text. The method explores several decoding paths per question, so it costs more compute than one greedy pass.